A Multilevel Control Method for Monitoring Frequency of Edge Devices Based on Outlier Identification

Abstract

The present invention relates to the technical field of railway engineering construction operation and maintenance, and particularly relates to a multi-level control method for the monitoring frequency of edge devices based on outlier identification, which includes the following steps: constructing a knowledge graph for railway infrastructure monitoring; collecting railway infrastructure monitoring data, deploying a model on edge devices, and performing data preprocessing on sensor data; respectively using a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination; using the knowledge graph to find outlier discrimination thresholds for the current monitoring data to be discriminated, and performing control on the monitoring frequency of edge devices according to the threshold judgment results; through the above method, the problem of time delay generated by the transmission of monitoring data in the previous cloud computing method can be solved, outlier discrimination of multi-category data in the Internet of Things can be realized, and verification is carried out through a multi-level hierarchical discrimination method, effectively reducing the false alarm probability of the monitoring system and improving the monitoring efficiency and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of railway engineering construction operation and maintenance, and particularly to a multi-level control method for the monitoring frequency of edge devices based on outlier identification. Background Art

[0002] During the operation of railway engineering, ensuring the healthy service state of railway infrastructure is the key to the safe operation of the line. During railway operation and maintenance, multi-index monitoring is often carried out on infrastructure, and a plurality of monitoring devices form an Internet of Things to timely detect anomalies in structural monitoring data to ensure a good operating state of the line.

[0003] The existing monitoring outlier discrimination adopts a cloud computing mode, where the monitoring data obtained by the front-end monitoring devices is transmitted to the cloud computing center, and data calculation and analysis are performed through relevant algorithms deployed in the cloud, and then the results are transmitted to the user side to complete information feedback. In the case of a high monitoring frequency, there is a time delay in the transmission of a large amount of monitoring data during information transmission. The processing of data information by the cloud computing center is restricted by the network transmission speed, and railway operation and maintenance management staff cannot obtain effective information feedback in a timely manner. Starting from the Internet of Things monitoring technology, the reasons for the generation of outliers during the monitoring process include: external environmental impacts, poor transmission signals of monitoring devices, abnormal impacts of monitoring devices, and anomalies in the basic structure itself, etc.

[0004] The existing data analysis algorithms deployed in the cloud do not consider the situation of misjudgment of outliers caused by external factors, and cannot accurately eliminate "false outliers", resulting in the system blindly alarming when encountering outliers, and being unable to automatically adapt and adjust the acquisition frequency of monitoring devices to provide more available data information, increasing the workload of railway operation and maintenance personnel in identifying and preparing for abnormal situations, and being unable to provide an optimal solution for the allocation of human and material resources. In addition, there are various types of sensing devices in the railway monitoring Internet of Things, and the existing outlier discrimination algorithms are mostly for a single index. Due to the lack of adaptability to monitoring indicators, the algorithms perform poorly in the discrimination of diversified indicators.

[0005] In summary, it is very necessary to propose a multi-level control method for the monitoring frequency of edge devices based on outlier identification that can solve the problem of time delay in the transmission of monitoring data in the previous cloud computing method, effectively reduce the false alarm probability of the monitoring system, and improve the monitoring efficiency and reliability. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-level control method for the monitoring frequency of edge devices based on outlier identification, aiming to solve the problem of time delay in the transmission of monitoring data in the previous cloud computing method, effectively reduce the false alarm probability of the monitoring system, and improve the monitoring efficiency and reliability.

[0007] To achieve the above object, a multi-level control method for the monitoring frequency of edge devices based on outlier identification of the present invention includes the following steps:

[0008] Extract entities from the monitoring knowledge of railway infrastructure to construct a monitoring knowledge graph of railway infrastructure;

[0009] Collect monitoring data of railway infrastructure, deploy a model on edge devices, and perform data preprocessing on sensor data;

[0010] Use a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination respectively;

[0011] Use the knowledge graph to find the outlier discrimination threshold for the current monitoring data to be discriminated, perform threshold judgment, and control the monitoring frequency of edge devices according to the judgment result.

[0012] Among them, in the steps of using a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination respectively, the process of using the sequence outlier discrimination model for outlier discrimination is as follows:

[0013] Perform segmentation processing on sequence data;

[0014] Establish a sequence outlier discrimination model considering time factor weighting;

[0015] Discriminate sequence outliers.

[0016] Among them, in the steps of using a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination respectively, the process of using the outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination is as follows:

[0017] Construct multiple Latin hypercube sampling data sets;

[0018] Perform multi-level Fourier series expansion on the constructed data sets respectively, and use Monte Carlo simulation to form a confidence interval;

[0019] Use the confidence interval to discriminate outliers.

[0020] Among them, in the steps of using a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination respectively:

[0021] If the discrimination shows an anomaly, search for the outlier discrimination threshold based on the knowledge graph;

[0022] If the discrimination shows normal, return to the user side.

[0023] Among them, in the step of using the knowledge graph to find the outlier discrimination threshold for the current monitoring data to be discriminated, performing threshold judgment, and controlling the monitoring frequency of the edge device according to the judgment result:

[0024] If the threshold judgment shows an anomaly, perform anomaly identification.

[0025] If the threshold shows normal, increase the monitoring frequency of the current monitoring data.

[0026] Among them, in the step of performing anomaly identification when the threshold judgment shows an anomaly:

[0027] Mobilize the monitoring indicators with high relevant relationships according to the railway infrastructure monitoring knowledge graph for multi-model anomaly identification, and based on the identification results, construct an anomaly identification risk alarm model, and return an anomaly report and risk level to the user side.

[0028] Among them, in the step of adjusting the monitoring frequency of the current monitoring data when the threshold shows normal:

[0029] Adjust the monitoring frequency of the current monitoring data through the adaptive adjustment module, and repeat data preprocessing and anomaly discrimination for the current sensor data collected after adjusting the monitoring frequency in the next round.

[0030] A multi-level control method for the monitoring frequency of edge devices based on outlier identification according to the present invention extracts entities from the railway infrastructure monitoring knowledge, constructs a railway infrastructure monitoring knowledge graph; collects railway infrastructure monitoring data, deploys a model at the edge device, and performs data preprocessing on the sensor data; respectively uses a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for anomaly discrimination; uses the knowledge graph to find the outlier discrimination threshold for the current monitoring data to be discriminated, performs threshold judgment, and controls the monitoring frequency of the edge device according to the judgment result. By means of edge computing technology at the sensor end and based on the sensor monitoring data, local discrimination of monitoring outliers is realized; using the sequence outlier discrimination model considering time freshness and the outlier discrimination model based on Fourier and Monte Carlo as the primary discrimination methods, and the railway domain-specific knowledge graph as the secondary discrimination basis and the interface for mobilizing relevant indicators for discrimination, a comprehensive discrimination is formed to give an early warning for the screened and effective outlier discrimination; hierarchical identification is performed for the anomaly determination, false anomalies generated by the monitoring are effectively screened, and for different anomaly situations, the acquisition frequency of the edge monitoring device is adjusted specifically to improve the monitoring and early warning efficiency and ensure the reliability and accuracy of the anomaly alarm of the monitoring system. Description of the Drawings

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0032] Figure 1 It is a schematic flowchart of the multi-level control method for the monitoring frequency of edge devices based on outlier identification of the present invention.

[0033] Figure 2 It is a step flowchart of the multi-level control method for the monitoring frequency of edge devices based on outlier identification of the present invention.

[0034] Figure 3 It is a schematic diagram of the network model structure of the multi-task learning objective of the present invention.

[0035] Figure 4 It is a visualization graph of the knowledge graph for railway infrastructure monitoring of the present invention.

[0036] Figure 5 It is a schematic diagram of the structure of the sequence outlier discrimination model considering time factors of the present invention.

[0037] Figure 6 It is a schematic diagram of the structure of the outlier discrimination model based on Fourier and Monte Carlo of the present invention.

[0038] Figure 7 It is a flowchart of the adaptive adjustment of the acquisition frequency of the sensing device of the present invention. Detailed implementation manners

[0039] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present application.

[0040] The terms used in the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms of "a", "the", and "said" used in the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0041] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0042] Please refer to Figures 1 to 7 , wherein Figure 1 is a schematic flow chart of a multi-level control method for the monitoring frequency of edge devices based on outlier identification, Figure 2 is a step flow chart of a multi-level control method for the monitoring frequency of edge devices based on outlier identification, Figure 3 is a schematic diagram of the network model structure of multi-task learning objectives, Figure 4 is a visualization graph of the railway infrastructure monitoring knowledge graph, Figure 5 is a schematic diagram of the structure of a sequence outlier discrimination model considering time factors, Figure 6 is a schematic diagram of the structure of an outlier discrimination model based on Fourier and Monte Carlo, Figure 7 is a flow chart for the adaptive adjustment of the acquisition frequency of sensing devices.

[0043] The present invention provides a multi-level control method for the monitoring frequency of edge devices based on outlier identification, including the following steps:

[0044] S100: Extract entities from the railway infrastructure monitoring knowledge and construct a railway infrastructure monitoring knowledge graph;

[0045] In this embodiment, a network model with multi-task learning objectives is constructed based on deep learning to extract entities from the railway infrastructure monitoring knowledge and construct a railway infrastructure monitoring knowledge graph; the specific network structure is implemented as follows:

[0046] Input data - Word embedding layer - Shared encoding layer - [Task branch 1 (entity recognition) + Task branch 2 (attribute acquisition) + Task branch 3 (relationship acquisition)].

[0047] The network model with multi-task learning objectives first captures the semantic information of words in the word embedding layer, converting each word in the text into a fixed-length vector representation. Then, through the shared encoding layer, the self-attention mechanism is used to encode the context information of the text. Next, the information is passed to different task-specific high-level networks for specific task learning, including knowledge entity recognition, knowledge attribute acquisition, and knowledge relationship acquisition. In the knowledge entity recognition task, the sequence labeling method is used to assign a predefined entity label to each token in the text. In the knowledge attribute acquisition task, a fully connected layer classifier is adopted to achieve attribute recognition, and then a regression model is used to extract specific attribute values from the recognized entity context. In the knowledge relationship acquisition task, through the two-stream network, the two entity contexts are processed separately first, and then fused and merged and relationship classification are carried out to achieve the knowledge relationship acquisition task.

[0048] Design a multi-task joint loss function for the network model with multi-task learning objectives, and adopt dynamic weight adjustment to control the learning and training effect of the model. The specific loss function is defined as follows:

[0049] L total = λ1L ner + λ2L attr + λ3L re ;

[0050] L ner = -log P(y * |x; θ);

[0051]

[0052] In the formula: L ner is the knowledge entity recognition loss; L attr is the knowledge attribute acquisition loss; L re is the relationship acquisition loss; λ1, λ2, λ3 are loss balance adjustment coefficients; x is the input sequence, y * is the correct label sequence; θ is the model parameter; P is the probability of a specific label sequence given the input and the model's parameters; N is the number of knowledge entities; M is the number of knowledge attributes; y ij is a binary indicator indicating whether entity i has attribute j; p ij is the probability that the model predicts that entity i has attribute j; v i is the correct attribute value of entity i; is the model-predicted attribute value; R is the number of knowledge relationship categories.

[0053] S200: Collect railway infrastructure monitoring data, deploy the model on edge devices, and perform data preprocessing on sensor data;

[0054] In this embodiment, railway infrastructure monitoring data is collected, and models are deployed on edge devices to preprocess sensor data. Among them, the outlier discrimination algorithm model is deployed on edge devices, including format conversion, model compression, and deployment integration. When preprocessing the monitoring data, the key is to embed category labels into the data, which are used as query inputs in the constructed railway infrastructure monitoring knowledge graph to obtain the thresholds of indicators corresponding to the categories. The data preprocessing includes data missing value imputation and data noise reduction for the monitoring data c1, c2, c3... collected by the sensing devices.

[0055] The data preprocessing includes missing value imputation and noise reduction. In this embodiment, the missing value imputation uses the K-nearest neighbor imputation method, that is, the missing data imputation algorithm based on the K-nearest neighbor algorithm. According to the distance measure or correlation analysis, the K samples closest to the missing sample are selected, and the data of these K samples are weighted to estimate the missing data of the sample.

[0056] For the above K-nearest neighbor imputation method, assume two samples y a and y b The distance between them is D(y a ,y b ), and the heterogeneous Euclidean distance metric is used to define D(y a ,y b ) as follows:

[0057]

[0058] In the formula, Di(y ai ,y bi ) is the distance between the i-th variables of samples y a and y b ;

[0059] For the case where the i-th variable of samples y a and y b has missing values, the maximum distance 1 is returned, which is calculated through the D0 function and defined as:

[0060]

[0061] For the K-nearest neighbor imputation method, if the i-th variable of sample y a has a missing value, then the K samples closest to the sample are selected, and the i-th variables of these K samples have no missing values. The distances of these K samples from sample y a from near to far form a set:

[0062] (m j is the sample closest to sample y a );

[0063] By weighting each of the K nearest neighbor samples and summing the weights of each category, the category with the largest sum of weights is selected as the imputation value.

[0064] Wavelet transform is used for data denoising to remove the noise interference in the monitoring data and obtain useful data information. First, the Daubechies wavelet basis function is selected, then the decomposition level is determined for fast wavelet transform, and finally data information reconstruction is performed. The related formulas are as follows:

[0065]

[0066] D j [n] = ∑ k f[n]·ψ j,k [n];

[0067] ψ j,k [n] = ψ[n]·2 -j ·ψ(2 -j n - k);

[0068] In the formula: f[n] is the original monitoring data signal; k is the displacement parameter; C j [n] is the approximation coefficient; D j [n] is the detail coefficient; is the scaling function; ψ[n] is the discrete function of the mother wavelet; j is the decomposition level.

[0069] S300: The sequence outlier discrimination model considering time freshness and the outlier discrimination model based on Fourier and Monte Carlo are respectively used for outlier discrimination;

[0070] In this embodiment, the sequence outlier discrimination model considering time freshness and the outlier discrimination model based on Fourier and Monte Carlo are respectively used for outlier discrimination; where:

[0071] The sequence outlier discrimination considering time factors is to filter the false alarms of single mutation outliers caused by external environmental influences. The process of outlier discrimination using the sequence outlier discrimination model is as follows:

[0072] S311: Perform segmentation processing on the sequence data;

[0073] S312: Establish a sequence outlier discrimination model considering time factor weighting;

[0074] S313: Discriminate sequence outliers.

[0075] Furthermore, in S311: When performing segmentation processing on the sequence data, taking the current monitoring data to be discriminated as the benchmark, a time series A containing 21 data points is intercepted forward along the time dimension and divided into 3 time step windows T = 3;

[0076] Further, in S312: Establishing a sequential outlier discrimination model considering time factor weighting,

[0077] The structure of the sequential outlier discrimination model considering time factor is implemented as follows:

[0078] Construct a time influence degree function that satisfies the principle of "larger nearby and smaller far away" in the time dimension, and the function is non-negative and monotonically increasing. The constructed time influence degree function is as follows:

[0079]

[0080] In the formula: F(t) represents the weighted function considering time influence; t aq represents the start time point of the previous sequence; t a represents the start time point of the current sequence; e represents the base of the exponential function; S represents the error of the monitoring data.

[0081] The sequential outlier discrimination model considering time factor includes two parts: a decoder and an encoder. As Figure 5 shown, the specific structure is: Encoder: Input layer - LSTM layer - Time factor weighting layer - Reparameterization layer - Output layer; Decoder: Fully connected layer - LSTM layer - Activation layer - Output layer. The application principle of the time factor weight in the key steps implemented by the above network is as follows:

[0082] Time series data set The feature vector observed by the LSTM layer at time step t is a t . In the encoder, the specific implementation of the time factor weighting is as follows:

[0083] w t = F(a t ; θ F );

[0084]

[0085] where, w t is the weight at time step t; θ F is the parameter of the time influence degree function; F(·) is the operation of the time influence degree function; is the weighted feature vector; ⊙ represents element-wise multiplication; h t represents the output of the time factor weighting layer.

[0086] Further, in S313: Discriminating sequential outliers, the threshold of the sequential outlier discrimination model considering time factor is obtained through cross-validation to obtain the optimal threshold, and the reconstruction error is compared with the threshold to achieve outlier judgment. The judgment result is fed back to the user side and the edge device acquisition frequency adaptive adjustment module.

[0087] The outlier discrimination model based on Fourier and Monte Carlo is to construct a discrimination criterion formed according to the variation law of the monitoring data itself, and filter out false alarms caused by ignoring the natural variation over time but being restricted by artificial fixed values; the process of using the outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination is as follows:

[0088] S321: Construct multiple Latin hypercube sampling datasets;

[0089] S322: Perform multi-level Fourier series expansion on the constructed datasets respectively, and use Monte Carlo simulation to form a confidence interval;

[0090] S323: Use the confidence interval to discriminate outliers.

[0091] Furthermore, in S321: When constructing multiple Latin hypercube sampling datasets, to construct the outlier discrimination model based on Fourier and Monte Carlo, it is necessary to first construct a dataset of the current monitoring data through Latin hypercube sampling. For the obtained current monitoring data Divide the monitoring data into N uniformly spaced intervals, and randomly select points for each interval to form a new dataset When the number of selected points m = 1, the principle of dataset formation can be expressed as:

[0092]

[0093] Furthermore, in S322: When performing multi-level Fourier series expansion on the constructed datasets respectively and using Monte Carlo simulation to form a confidence interval, after constructing the dataset of the current monitoring through Latin hypercube sampling, for the new dataset Perform multi-level Fourier series expansion, based on the principle of Fourier series expansion:

[0094]

[0095] where a0, a n , b n are the coefficients during multi-level Fourier series expansion.

[0096] After Fourier series expansion, a new dataset is obtained Using the Monte Carlo simulation idea, execute S321 multiple times to obtain k groups of data, form a dataset group, and perform multi-level Fourier series expansion respectively to obtain the data matrix X M :

[0097]

[0098] Obtain the upper limit:

[0099]

[0100] Lower limit:

[0101]

[0102] Taking the 95% confidence interval of its upper and lower limits as the discrimination basis for the current monitoring data, the confidence interval can be expressed as:

[0103] [X s - 0.05(X s - X x ), X x + 0.05(X s - X x )].

[0104] Furthermore, in S323: Using the confidence interval to discriminate outliers, the confidence interval generated by the outlier discrimination model constructed based on the above Fourier series expansion and Monte Carlo simulation ideas is used to perform outlier discrimination on the current data.

[0105] The confidence interval generated by the outlier discrimination model constructed based on the above Fourier series expansion and Monte Carlo simulation ideas is used to perform outlier discrimination on the current data.

[0106] S400: Using the knowledge graph to find the outlier discrimination threshold for the current monitoring data to be discriminated, performing threshold judgment, and according to the judgment result, controlling the monitoring frequency of the edge device;

[0107] In this embodiment, the category label of the current discrimination data object is read, and the category label is used as the input to query the threshold corresponding to the category index in the railway infrastructure monitoring knowledge graph and used as the judgment threshold for the current discrimination data. Taking the post-construction settlement monitoring of the subgrade as an example, when it is detected that the sequence outlier discrimination model considering time freshness and the outlier discrimination model based on Fourier and Monte Carlo discriminate outliers, a search will be performed in the constructed knowledge graph. The threshold setting in the railway infrastructure monitoring knowledge graph is stored according to the specification requirements: the post-construction settlement of the ballastless track should not exceed 15 mm. If the specification threshold is exceeded, the post-construction settlement abnormal information of the ballastless track will be returned to the client.

[0108] Furthermore, if the threshold judgment shows an anomaly, outlier identification is performed. If the threshold shows normal, the monitoring frequency of the current monitoring data is adjusted and strengthened.

[0109] Furthermore, in the step of performing outlier identification, relevant monitoring indicators with high correlation are mobilized according to the railway infrastructure monitoring knowledge graph for multi-model outlier identification, and based on the identification result, an outlier identification risk alarm model is constructed to return an outlier report and a risk level to the user side.

[0110] Among them, according to the knowledge graph, other monitoring index data with a high correlation with the current monitoring data index are mobilized and input into the multi-model anomaly identification module for identification, and an anomaly identification risk alarm model is constructed. With the monitoring indexes with a correlation of more than 60% as assistance, assuming there are m relevant monitoring indexes that meet the requirements, the correlation coefficient B = {b1, b2, …, b m}, and the anomaly judgment results of each relevant index and the current monitoring data index are represented by a (0,1) distribution, and the risk value Y = {y1, y2, …, y m , y x} is obtained. Then the risk alarm model is constructed as follows:

[0111]

[0112] The corresponding risk alarm level is shown in Table 1:

[0113] Table 1 Risk Alarm Level Table

[0114] Risk level RiskSignal Yellow RiskSignal<0.5 Orange 0.5 ≤ RiskSignal < 0.7 Red 0.7 ≤ Risk Signal

[0115] Taking the post-construction settlement monitoring of the subgrade as an example, the monitoring indexes with a correlation exceeding 60% at the same moment are used as assistance. For example, the current anomaly identification value of the post-construction settlement monitoring of the subgrade is y x , the current anomaly identification values of the moisture content and horizontal displacement of the monitoring indexes that meet the requirements are y1 and y2 respectively, and their correlation coefficients are b1 and b2 (b1, b2 > 0.6). Then the risk alarm model is RiskSignal = 0.8y x +0.2×(b1y1 + b2y2) / 2. Alarm is made according to the value corresponding table of Risk Signal.

[0116] Furthermore, in the step of adjusting the monitoring frequency of the current monitoring data, the monitoring frequency of the current monitoring data is adjusted through the adaptive adjustment module, and the data preprocessing and anomaly discrimination are repeated for the current sensor data collected after the next round of monitoring frequency adjustment, that is, S200 - S300 are repeatedly executed. For the next round of collected data judged to be normal, the original collection frequency is restored through the adaptive adjustment module.

[0117] Among them, the sensor edge device receives the signal for adaptive adjustment of the collection frequency. If an abnormal result is received, the sensor edge device automatically increases the collection frequency to collect data, providing data guarantee for repeatedly executing S200 - S300; if the received result is normal, the sensor maintains the original collection frequency and works normally.

[0118] As Figure 7 shown, the steps of adaptive adjustment of the collection frequency of the sensing device include: receiving the signal, analyzing the signal property, adjusting the real-time collection frequency, and collecting the monitoring data.

[0119] Among them, for the received signal, the judgment result obtained by edge computing processing is used as the signal input to the adaptive adjustment module.

[0120] Among them, analyze the nature of the anomaly and analyze whether the input judgment result belongs to an anomaly or is normal.

[0121] Among them, adjust the real-time acquisition frequency. If it is an abnormal signal, the system will increase the data acquisition frequency to densely monitor the development of this abnormal situation to ensure that all key data can be captured. If it is a normal signal, maintain the current acquisition frequency or adjust it back to the normal level to avoid unnecessary data accumulation and resource waste.

[0122] Among them, collect monitoring data. The adaptive adjustment module controls the working mode of the sensor to achieve real-time adjustment of the acquisition frequency, and the operation feedback makes the adjusted acquisition frequency take effect immediately. The system continues to collect data according to the new acquisition frequency to provide data support for the abnormal value discrimination module. If the abnormal discrimination gives a yellow alarm, the acquisition frequency of the device is adjusted to 1.25 times; if it gives an orange warning, the acquisition frequency is adjusted to 1.5 times; if it gives a red warning, the acquisition frequency is adjusted to 2 times.

[0123] In the present invention, the monitoring data abnormal value discrimination model is deployed on the edge device to solve the time delay problem caused by transmitting a large amount of data back, and perform data judgment locally so that the staff can timely grasp the health status of the roadbed. Build a knowledge graph for railway infrastructure monitoring. The threshold of the abnormal value discrimination model is not for a single value, but is adjusted according to the category label of the input data to achieve multi-index discrimination, and the model has diversified capabilities. Build a sequence abnormal value discrimination method based on time freshness, determine the influence weight for the monitoring sequence data in the time dimension according to the principle of "larger for the near and smaller for the far", build an abnormal value discrimination model based on Fourier and Monte Carlo, and determine the confidence interval through the Monte Carlo simulation idea to eliminate the pseudo-abnormal phenomena caused by external factors and ensure the accuracy of the returned abnormal alarm information. According to the returned abnormal judgment information, control the acquisition frequency of the edge device in real time in a targeted manner, adaptively adjust the frequency, and provide richer and more effective monitoring information.

[0124] Those skilled in the art will readily think of other implementation schemes of the present application after considering the specification and the content disclosed herein. The present application aims to cover any variations, uses or adaptive changes of the present application, and these variations, uses or adaptive changes follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application.

[0125] It should be understood that the present application is not limited to the exact structure already described and shown in the drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A multi-level control method for the monitoring frequency of edge devices based on outlier identification, characterized in that It includes the following steps: Extract entities from railway infrastructure monitoring knowledge and construct a railway infrastructure monitoring knowledge graph; Collect railway infrastructure monitoring data, deploy a model on edge devices, and perform data preprocessing on sensor data; Use a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination respectively; The process of using the sequence outlier discrimination model for outlier discrimination is as follows: Perform segmentation processing on sequence data; Establish a sequence outlier discrimination model considering time factor weighting; Discriminate sequence outliers; The process of using the outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination is as follows: Construct multiple Latin hypercube sampling datasets; Perform multi-level Fourier series expansion on the constructed datasets respectively, and use Monte Carlo simulation to form a confidence interval; Use the confidence interval to discriminate outliers; Use the knowledge graph to find the outlier discrimination threshold for the current monitoring data to be discriminated, perform threshold judgment, and control the monitoring frequency of edge devices according to the judgment result; If the threshold judgment shows an anomaly, perform anomaly identification: mobilize monitoring indicators with high relevant relationships according to the railway infrastructure monitoring knowledge graph for multi-model anomaly identification, and construct an anomaly identification risk alarm model based on the identification result, and return an anomaly report and risk level to the user side; If the threshold shows normal, increase the monitoring frequency of the current monitoring data.

2. The multi-level control method for the monitoring frequency of edge devices based on outlier identification according to claim 1, wherein In the steps of using a sequence outlier discrimination model considering time freshness and an outlier discrimination model based on Fourier and Monte Carlo for outlier discrimination respectively: If the discrimination shows an anomaly, perform outlier discrimination threshold search based on the knowledge graph; If the discrimination shows normal, return to the user side.

3. The multi-level control method for the monitoring frequency of edge devices based on outlier identification according to claim 1, characterized in that In the step of adjusting the monitoring frequency of the current monitoring data when the threshold shows normal: Adjust the monitoring frequency of the current monitoring data through an adaptive adjustment module, and repeat data preprocessing and outlier discrimination on the current sensor data collected after adjusting the monitoring frequency in the next round.

Images