Well lid intelligent monitoring algorithm based on Beidou positioning and multi-element machine learning fusion

Abstract

The invention discloses a well lid intelligent monitoring algorithm based on Beidou positioning and multivariate machine learning fusion. The well lid intelligent monitoring algorithm comprises the steps of collecting well lid monitoring training data, preprocessing the data and the like. A Beidou positioning device, a gyroscope and an accelerometer are installed on the inner side of a well lid, the position, speed, positioning state, three-axis attitude, acceleration and other data of the well lid are collected in real time, multiple machine learning algorithms such as association rules, clustering and classification are intelligently fused, a multiple judgment mechanism is set, and the well lid can be accurately positioned. Therefore, four states of the well lid can be monitored and discriminated more accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of manhole cover monitoring, and specifically relates to an intelligent manhole cover monitoring algorithm based on the fusion of Beidou positioning and multivariate machine learning. Background Art

[0002] With the acceleration of urbanization, the number of manhole covers on the road has increased significantly, and their management and maintenance have become increasingly important. However, due to various reasons, such as long-term lack of maintenance, material aging, vehicle crushing, etc., manhole covers are prone to displacement, flipping, sinking or even loss. These conditions not only affect the traffic order of the city, but also pose a serious threat to the safety of pedestrians and vehicles. In order to cope with this threat, smart manhole covers came into being. Smart manhole covers use Beidou positioning + advanced Internet of Things technology, combining traditional manhole covers with sensors to monitor the status of manhole covers in real time, including key information such as position and posture. Once the algorithm detects an abnormality in the manhole cover, the sensor will immediately issue an alarm to notify relevant personnel to maintain or replace it in time. This intelligent monitoring method greatly improves the efficiency and accuracy of manhole cover maintenance and effectively prevents the occurrence of safety accidents.

[0003] At present, many manhole cover monitoring algorithms based on this application scenario use machine learning methods, such as neural network methods or other classification algorithms, to train a large amount of data collected in a static environment to obtain a classification model to determine the normal / abnormal state of real-time dynamic data. However, the classification model has the following shortcomings: 1. It depends on data quality. If the data quality is poor, the classification effect is not good; 2. It is not sensitive to abnormal data recognition and is easy to classify abnormal data into a certain category. Therefore, it is necessary to optimize a single classification algorithm to enhance the accuracy and stability of manhole cover status judgment. Summary of the invention

[0004] The problem to be solved by the present invention is to provide an intelligent monitoring algorithm for manhole covers based on the fusion of Beidou positioning and multivariate machine learning. By installing Beidou positioning equipment and two sensors, a gyroscope and an accelerometer, on the inside of the manhole cover, the position, speed, positioning status, three-axis attitude, acceleration and other data of the manhole cover are collected in real time, and multiple machine learning algorithms such as association rules, clustering, and classification are intelligently integrated, and multiple discrimination mechanisms are set to more accurately monitor and discriminate the four states of the manhole cover.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: an intelligent monitoring algorithm for manhole covers based on the fusion of Beidou positioning and multivariate machine learning, based on a Beidou positioning device, a gyroscope and an accelerometer arranged on the manhole cover, comprising the following steps: 11. Collect manhole cover monitoring training data and label the training data to obtain data set D. The elements in the training data include timestamp t, longitude lat, latitude lon, speed v, horizontal precision dilution HDOP, number of receivable satellites sat, carrier-to-noise ratio CNR of each satellite, three-axis angular velocity ωx, ωy and ωz, three-axis acceleration ax, ay and az, The label is a natural number ranging from 1 to 4, where 1 represents stillness, 2 represents slight vibration, 3 represents tilt, and 4 represents movement. 12. Perform data preprocessing on the data set D and convert the longitude lat and latitude lon in the data into plane rectangular coordinates x and y to obtain the data set DP; 13. Perform feature extraction on the data set DP to obtain the data set DF; 14. Discretize the data set A consisting of data with labels 1 and 2 and the data set B consisting of data with labels 3 and 4 in the data set DF, count the maximum and minimum values ​​of each element and divide them into p parts, then execute the Apriori association rule algorithm to mine frequent L-item sets and obtain two sets of rules PA and PB. The value range of L is [1, m-1], where m is the number of elements in a single data item in the data set DF. The value range of p is 1-100; 15. Perform K-Means clustering on the data in the data set DF to obtain the cluster center C of each type of data i ,in, ; 16. Count the data with the highest label quantity in each category and get the set S, and re-calculate the cluster center of each category of data according to the label label , , the rest of the data is set RS, If RS=∅, execute step 17; otherwise, count the number of types of label numbers in RS, and find the cluster center of each type of data according to the label. Then execute step 17. ; 17. For the data in set S, use the SVM algorithm to train a multi-classification model f, where the categories and labels in the model have the same meaning; 18. Collect real-time data of manhole cover monitoring. The first piece of data collected is not processed. The remaining data is preprocessed and features are extracted to obtain data. 19. If PA∩PB=∅, determine whether the data in data satisfies the rule PA or PB. If so, determine it as the state corresponding to the rule and execute step 24. Otherwise, execute step 20. 20. Calculate the distance Sdis between data and each cluster center in the set S respectively. If max(Sdis)≤dis_th, execute step 21, otherwise execute step 22. The value range of dis_th is 0-100. 21. Use the multi-classification model f to classify the data, output the category label of the data, and execute step 24; 22. If the set RS is not ∅, then calculate the distance Rdis between data and each cluster center in RS respectively. If max(Rdis)≤dis_th, execute step 23. Otherwise, determine data as abnormal data, output 0 and execute step 24. The value range of dis_th is 0-100. 23. Find the class label corresponding to min(Rdis), determine it as the state represented by the label, and output the label number label; 24. If the output label is 1 or 2, continue to step 18 without processing. If the output is 3 or 4, an alarm is triggered. If the output is 0, a "data abnormality" prompt is displayed.

[0006] The beneficial effects of the present invention are: 1. The discrimination mechanism formed by various intelligent algorithms such as association rules, clustering, classification, etc. is stable and reliable. 2. The hardware cost is low and the algorithm execution is efficient.

[0007] The present invention is described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0009] Referring to the accompanying drawings, the present invention provides an intelligent manhole cover monitoring algorithm based on the fusion of Beidou positioning and multivariate machine learning, based on a Beidou positioning device, a gyroscope and an accelerometer arranged at the center position of the inner side of the manhole cover.

[0010] The method of the present invention comprises the following steps.

[0011] 11. Collect manhole cover monitoring training data and label the training data to obtain data set D.

[0012] The elements in the training data include timestamp t, longitude lat, latitude lon, velocity v, horizontal precision dilution HDOP, number of receivable satellites sat, carrier-to-noise ratio CNR of each satellite, three-axis angular velocity ωx, ωy and ωz, and three-axis acceleration ax, ay and az.

[0013] The label is a natural number ranging from 1 to 4, where 1 represents stillness, 2 represents slight vibration, 3 represents tilt, and 4 represents movement.

[0014] The data in the data set D is in the form of {t, lat, lon, v, HDOP, sat, CNR, ωx, ωy, ωz, ax, ay, az, label}.

[0015] The state of a manhole cover generally includes four states, namely: static, the state without any external force factors, in which the position, attitude angle, acceleration, etc. do not change in an ideal state; slight vibration, generally the instantaneous or continuous pressure exerted on the manhole cover by people or vehicles passing by (for example, the wheels are always pressing on the manhole cover when parking), and the acceleration usually changes greatly; tilt, that is, the manhole cover has a slight or serious flip, and the attitude angle usually changes significantly. When the tilt is serious, the Beidou positioning device will be in a semi-blocked state, and the ability to receive satellite signals will be better than the normal state when it is not tilted; moving, it is generally believed that a theft has occurred. At this time, the position will change significantly and continuously, and the Beidou positioning device will be in a semi-blocked state or an unblocked state, the ability to receive satellite signals will be significantly stronger, and the attitude angle may change significantly (pick it up and steal it), or it may not change significantly (keep it in a horizontal posture and steal it).

[0016] Therefore, based on the above four states of the manhole cover, the data indicators that may be associated with each state are analyzed, and the data such as timestamp t, longitude lat, latitude lon, speed v, horizontal precision dilution HDOP, number of receivable satellites sat, carrier-to-noise ratio CNR of each satellite, three-axis angular velocity ωx, ωy and ωz, three-axis acceleration ax, ay and az are selected for collection. Among them, CNR is an array composed of sat data, and the other data are single scalars.

[0017] Before starting to collect training data, use high-precision positioning equipment or other means to obtain the real latitude and longitude of the manhole cover and record it in the background as a fixed standard position data. During the collection, the slight vibration state is simulated by two ways: people and vehicles walking over the manhole cover and staying on the manhole cover; the tilt state is simulated by turning the manhole cover to different degrees of 0°~180°; the moving state is simulated by lifting, holding up, moving quickly, and moving slowly.

[0018] 3000 data items are collected in each of the four states with a sampling frequency of 1 Hz, that is, the data set D contains 12000 data items in total, and the pattern of each data item is: {t, lat, lon, v, HDOP, sat, CNR, ωx, ωy, ωz, ax, ay, az, label}.

[0019] 12. Perform data preprocessing on the data set D and convert the longitude lat and latitude lon in the data into plane rectangular coordinates x and y to obtain the data set DP.

[0020] The data in the data set DP has the format of {t,x,y,v', HDOP',sat',CNR',ωx',ωy',ωz',ax',ay',az',label}.

[0021] Since the Beidou positioning equipment is installed under the manhole cover and is in an obstructed state, its positioning accuracy will be greatly affected; in addition, sensors such as gyroscopes and accelerometers also have their own device errors. Therefore, the collected data set D needs to be pre-processed by cleaning and fusion, such as using Kalman filtering or outlier detection methods to identify abnormal values ​​and perform smoothing or elimination.

[0022] In order to facilitate the subsequent feature extraction, the position data needs to be converted into coordinates that are easy to compare and calculate. The processed data set DP is obtained. The size of DP is the same as D, which is also 12,000. The style of each data is: {t,x,y,v',HDOP',sat',CNR',ωx',ωy',ωz',ax',ay',az',label}.

[0023] 13. Perform feature extraction on the data set DP to obtain the data set DF.

[0024] The feature extraction values ​​include position change (Δx, Δy), velocity change Δv, horizontal precision dilution HDOP', number of satellites sat_ch with carrier-to-noise ratio higher than the set threshold CNR_th, maximum acceleration amax, minimum acceleration amin, average acceleration a_ave, standard deviation of acceleration a_std, attitude angle (θ, ϕ, ψ), in: , , , , n is the total number of data in the data set D, amax=max(ax',ay',az'), amin=min(ax',ay',az'), , ; The attitude angle (θ, ϕ, ψ) is obtained using the quaternion method based on the three-axis angular velocity ωx, ωy, ωz and the timestamp t.

[0025] The style of each data in the data set DF is: {Δx,Δy,Δv,HDOP',sat_ch,amax,amin,a_ave,a_std,θ,ϕ,ψ,label}.

[0026] When the positioning is accurate, the position change (Δx, Δy) and velocity change Δv can be used to preliminarily determine whether the manhole cover has moved. The horizontal precision factor HDOP' is one of the key parameters for determining the positioning accuracy. It directly affects the accuracy and reliability of positioning. The smaller the HDOP', the higher the positioning accuracy. In addition, the positioning performance is also related to the number of received satellites sat and the carrier-to-noise ratio CNR of each satellite. The more sats and the higher the CNR, the better the performance. However, even in the blocked state, the positioning device may receive a large number of satellites, and the carrier-to-noise ratio will be low in most cases, which is not of reference significance. Therefore, sat_ch is used to screen satellites with a carrier-to-noise ratio higher than the set threshold CNR_th. The value range of CNR_th is generally [27,30], that is, satellites below 27 are generally considered to have poor signal quality. The various statistical indicators of acceleration amax, amin, a_ave, and a_std reflect the influence of the manhole cover on the outside world from different angles. The attitude angle (θ, ϕ, ψ) provides real-time attitude information of the manhole cover.

[0027] The data set DF after feature extraction has one less data in each of the four states due to the differentiation of variables such as position and speed, with a total of 11996 data items. The style of each data item is: {Δx, Δy, Δv, HDOP', sat_ch, amax, amin, a_ave, a_std, θ, ϕ, ψlabel}.

[0028] 14. Discretize the data set A consisting of data with labels 1 and 2 and the data set B consisting of data with labels 3 and 4 in the data set DF, count the maximum and minimum values ​​of each element and divide them into p parts, then execute the Apriori association rule algorithm to mine frequent L-item sets and obtain two sets of rules PA and PB. Use Aij and Bij to represent the jth part of the i-th element in A and B, i=1~m-1, j=1~p. The value range of p is 1-100.

[0029] The value range of L is [1, m-1], where m is the number of elements in a single data item in the data set DF.

[0030] State 1 and state 2 are both normal states and do not need to be processed; state 3 and state 4 are abnormal states and require alarm processing if they occur. The significance of the association rule here is to mine the association relationship contained in various data indicators under normal and abnormal states, and use this relationship as a criterion for determining the state of real-time data. The core of the Apriori association rule algorithm is to find a set of elements that appear at the same time by traversing a large amount of data, called an item set.

[0031] Each data in the set DF has 13 elements, that is, m=13. Excluding the last element label, there are 12 in total. Therefore, the number of frequent item sets L can be an integer in [1,12]. In this case, L is 3, that is, mining common frequent 3-item sets. Since the data usually processed by the Apriori association rule algorithm is discrete data, and the data involved in this method are all continuous numerical data, before executing the algorithm, it is necessary to discretize each element, that is, count the maximum and minimum values ​​of each element in the data set, and evenly divide the interval into p equal parts, in this case 10 equal parts, that is, p=10, and mark each equal part as a symbolic variable, such as: use Aij to represent the jth equal part of the i-th element, and all data falling into this interval are represented as Aij, i=1~m-1, j=1~p.

[0032] 15. Perform K-Means clustering on the data in the data set DF to obtain the cluster center C of each type of data i ,in, .

[0033] The traditional classification-based manhole cover monitoring method collects each group of data according to the label, trains the classification model, and then judges the real-time data. It does not take into account the natural similarity between the data, resulting in the real-time data being normal data, but more similar to the abnormal data, leading to false alarms, or the real-time data being abnormal data, but more similar to the normal data, leading to misjudgments.

[0034] Clustering is to classify data with high similarity into one category. It is an unsupervised learning algorithm. In this example, K-Means clustering is used, and the clustering parameter k is specified as 4. The data is re-divided into 4 categories, and the difference between the classification results and the previous state-labeled data is observed, and the difference data is processed in the next step. The metric used to calculate the similarity is Euclidean distance, and the last element label in the data is not included in the calculation. After clustering, the center Ci of each category is the average value of all data in the category. Similarly, the last element label is not included in the calculation. In this example, the number of data in the four categories is 2857, 3014, 3115, and 3010 respectively.

[0035] 16. After clustering, count the data with the highest label quantity in each category and get the set S, and re-calculate the cluster center of each category of data according to the label label. , , the rest of the data is set RS, If RS=∅, execute step 17; otherwise, count the number of types of label numbers in RS, and find the cluster center of each type of data according to the label. Then execute step 17. .

[0036] The four types of data after clustering are not necessarily the same as the four groups of data collected by label, which means that the data in the four states are not necessarily extremely distinguishable. In this step, the data with the highest proportion of label numbers in each data is counted and recorded as set S. The idea is to identify these data as the most distinguishable data in all training data, extract them and use them as training data for the next classification model, which can improve the purity of data and model and facilitate more accurate classification. However, because the cluster data has been re-allocated, the cluster center has changed. Therefore, it is necessary to update each cluster center. , Define RS as a resource pool. If RS=∅, it means that the four types of data are exactly the same as the four groups of data collected by label. Then the data purity is very high and no other processing is required. You can directly enter the following training model step. Otherwise, the remaining data needs to be put into RS as a backup set of discrimination methods. The following step 22 will explain its discrimination method. Since RS may not contain data of all four labels, it is necessary to first count the number of label numbers num, in this case num=4, and find the center of the data under each label. , .

[0037] 17. For the data in set S, use the SVM algorithm to train a multi-classification model f. The categories and labels in the model have the same meaning.

[0038] The traditional SVM algorithm is to train a binary classification model (such as ), in this example, there are four states involved: {stationary, slight vibration, tilt, and movement}. Then, {stationary} can be taken as one category and {slight vibration, tilt, and movement} as another category, and the algorithm can be executed to generate a model. Then, {slight vibration, tilt, and movement} can be divided into {slight vibration} and {tilt, movement}, and the algorithm can be executed to generate sub-models. In this way, lower-level sub-models can be generated to finally complete the four-state judgment model f.

[0039] 18. Collect real-time data of manhole cover monitoring. The first piece of data collected is not processed; the remaining data is preprocessed and features are extracted to obtain data.

[0040] Since this method needs to differentiate the horizontal and vertical coordinates, speed and other elements when extracting features, it is necessary to perform the above processing from the second item of real-time data, that is, to start judging from the second item of data.

[0041] 19. If PA∩PB=∅, determine whether the data in data satisfies rule PA or PB. If so, determine it as the state corresponding to the rule and execute step 24; otherwise, execute step 20.

[0042] PA∩PB=∅, which means that the two sets of rules are completely different and have obvious distinction. Then the data data either belongs to only one of the sets of rules or neither of them. The state can be judged. If the data rule PA, it means the state is normal, and the data PB means the state is abnormal. If neither of them belongs, it means that data has no relevant data features and needs further judgment. In this example, the frequent three-item set of set A is A11: Δx∈(0,1), A21: Δy∈(0,1), A31: Δv∈(0,1); if the first three elements Δx, Δy, and Δv in data meet the above range respectively, they meet the rule PA; the frequent three-item set of set B is B45: HDOP'∈(2,2.5), B93: a_std∈(2,3), B113: ϕ∈(36,54); if the three elements HDOP, a_std, and ϕ in data meet the above range respectively, they meet the rule PB; in this example, PA and PB are clearly distinguished and can be used as the first step judgment criterion for real-time data. However, the characteristics of data do not meet PA or PB, so further judgment is needed.

[0043] 20. Calculate the distance Sdis between data and each cluster center in the set S respectively. If max(Sdis)≤dis_th, execute step 21, otherwise execute step 22. The value range of dis_th is 0-100 If the PA and PB rules are not sufficiently differentiated and cannot be used, or if the data does not meet any of the PA or PB rules, a classification algorithm must be executed to determine its status. However, another problem with the classification algorithm is that any input data that conforms to the data format will be classified. When a problem occurs with the Beidou positioning device or one of the gyroscopes or accelerometers, the data will become abnormal and will not reflect the current status of the manhole cover, but the algorithm will give a judgment result, resulting in a misjudgment. In addition, misjudgment may also occur for some normal data whose status differentiation is not obvious.

[0044] This step is to determine whether the data is normal and has high state discrimination before entering the classification judgment. The distance Sdis between data and each cluster center in S is calculated, where Sdis is an array of 4 elements, and each component is the distance between data and one of the Ci. When calculating the distance, the reliability of the components Δx, Δy, and Δv in data and Ci needs to be considered. Because when the Beidou positioning device is in an obstructed state, the positioning performance is poor, and the participation of Δx, Δy, and Δv in the calculation may lead to a large error. It is reasonable to participate in the calculation only when the positioning performance is good. The positioning performance is better when the horizontal precision factor HDOP≤5 and the number of satellites with a carrier-to-noise ratio higher than 27 is sat_ch_th≥5. Therefore, the way to calculate the distance is as follows: when the horizontal precision factor HDOP in data and Ci is less than the threshold HDOP_th, and sat_ch is greater than the threshold sat_ch_th, Sdis(i) is the Euclidean distance between the vectors of data and Ci; otherwise, Sdis(i) is the Euclidean distance between the vector components of data and Ci after removing Δx, Δy, and Δv. The value range of HDOP_th is 1-2, and the value is 2 in this example; the value range of sat_ch_th is 5-10, and the value is 8 in this example. If all distances are within the set threshold range, that is, max(Sdis)≤dis_th, the current data can be considered normal data and can enter the classification model judgment, otherwise enter step 22 for secondary judgment. In this example, dis_th=20 is set, Sdis=(18.35,13.02,17.54,3.11), all of which meet the conditions and can enter the classification model judgment.

[0045] 21. Use the multi-classification model f to classify the data, output the category label of the data, and execute step 24.

[0046] After the data in this example enters the classification model f, the SVM algorithm is run and label=4 is obtained, which is preliminarily judged as "moving".

[0047] 22. If the set RS is not ∅, then calculate the distance Rdis between data and each cluster center in RS respectively. If max(Rdis)≤dis_th, execute step 23. Otherwise, judge data as abnormal data, output 0 and execute step 24. The value range of dis_th is 0-100.

[0048] If Sdis does not meet the set threshold dis_th in step 20, data is suspected to be abnormal data or normal data with unclear state distinction. At this time, it will no longer enter the classification model f for judgment, but will be compared with the cluster centers in RS. Calculate the similarity, expressed as the distance Rdis, where Rdis is an array of 4 elements, each component is data and one of the If all distances are within the set threshold range, that is, max(Sdis)≤dis_th, the current data can be considered normal data, and the final judgment is entered into step 23; if not, it can be considered abnormal data, and label is set to 0. In this example, another set of real-time data data1 in step 20 has Sdis=(25.14,13.87,17.38,33.65), which does not meet the condition, and enters this step for secondary judgment, where Rdis=(19.06,6.22,17.74,12.31), which meets the condition, and enters step 23 for final judgment.

[0049] 23. Find the class label corresponding to max(Rdis), determine it as the state represented by the label, and output the label number label.

[0050] data1 and In the calculation result Rdis, min(Rdis)=6.22, and the corresponding label is 2, which is preliminarily judged as "slight vibration".

[0051] 24. If the output label is 1 or 2, continue to step 18 without processing. If the output is 3 or 4, an alarm is triggered. If the output is 0, a "data abnormality" prompt is displayed.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention, which should be included in the scope of the technical solution for protection of the present invention.

Claims

1. An intelligent manhole cover monitoring algorithm based on the integration of Beidou positioning and multivariate machine learning, based on a Beidou positioning device, a gyroscope and an accelerometer arranged on the manhole cover, characterized in that: The steps include:

11. Collect manhole cover monitoring training data and label the training data to obtain data set D. The elements in the training data include timestamp t, longitude lat, latitude lon, speed v, horizontal precision dilution HDOP, number of receivable satellites sat, carrier-to-noise ratio CNR of each satellite, three-axis angular velocity ωx, ωy and ωz, three-axis acceleration ax, ay and az, The label is a natural number ranging from 1 to 4, where 1 represents stillness, 2 represents slight vibration, 3 represents tilt, and 4 represents movement.

12. Perform data preprocessing on the data set D and convert the longitude lat and latitude lon in the data into plane rectangular coordinates x and y to obtain the data set DP; 13. Perform feature extraction on the data set DP to obtain the data set DF; 14. Discretize the data set A consisting of data with labels 1 and 2 and the data set B consisting of data with labels 3 and 4 in the data set DF, count the maximum and minimum values ​​of each element and divide them into p parts, then execute the Apriori association rule algorithm to mine frequent L-item sets and obtain two sets of rules PA and PB. The value range of L is [1, m-1], where m is the number of elements in a single data item in the data set DF. The value range of p is 1-100; 15. Perform K-Means clustering on the data in the data set DF to obtain the cluster center C of each type of data i ,in, ; 16. Count the data with the highest label quantity in each category and get the set S, and re-calculate the cluster center of each category of data according to the label label , , the rest of the data is set RS, If RS=∅, execute step 17; otherwise, count the number of types of label numbers in RS, and find the cluster center of each type of data according to the label. Then execute step 17. ; 17. For the data in set S, use the SVM algorithm to train a multi-classification model f, where the categories and labels in the model have the same meaning; 18. Collect real-time data of manhole cover monitoring. The first piece of data collected is not processed. The remaining data is preprocessed and features are extracted to obtain data.

19. If PA∩PB=∅, determine whether the data in data satisfies the rule PA or PB. If so, determine it as the state corresponding to the rule and execute step 24. Otherwise, execute step 20.

20. Calculate the distance Sdis between data and each cluster center in the set S respectively. If max(Sdis)≤dis_th, execute step 21, otherwise execute step 22. The value range of dis_th is 0-100.

21. Use the multi-classification model f to classify the data, output the category label of the data, and execute step 24; 22. If the set RS is not ∅, then calculate the distance Rdis between data and each cluster center in RS respectively. If max(Rdis)≤dis_th, execute step 23. Otherwise, determine data as abnormal data, output 0 and execute step 24. The value range of dis_th is 0-100.

23. Find the class label corresponding to min(Rdis), determine it as the state represented by the label, and output the label number label; 24. If the output label is 1 or 2, do not process and continue to step 18. If the output is 3 or 4, an alarm is triggered. If the output is 0, a "data abnormality" prompt is displayed.

2. The manhole cover intelligent monitoring algorithm based on Beidou positioning and multivariate machine learning fusion according to claim 1 is characterized in that: The data in the data set D is in the form of {t, lat, lon, v, HDOP, sat, CNR, ωx, ωy, ωz, ax, ay, az, label}.

3. The manhole cover intelligent monitoring algorithm based on Beidou positioning and multivariate machine learning fusion according to claim 1 is characterized in that: The data in the data set DP has the format of {t,x,y,v', HDOP',sat',CNR',ωx',ωy',ωz',ax',ay',az',label}.

4. The manhole cover intelligent monitoring algorithm based on Beidou positioning and multivariate machine learning fusion according to claim 1 is characterized in that: In step 13, the feature extraction values ​​include position change (Δx, Δy), velocity change Δv, horizontal precision dilution HDOP', the number of satellites sat_ch with a carrier-to-noise ratio higher than a set threshold CNR_th, maximum acceleration amax, minimum acceleration amin, average acceleration a_ave, standard deviation of acceleration a_std, attitude angle (θ, ϕ, ψ), in: ， ， , , n is the total number of data in the data set D, amax=max(ax',ay',az'), amin=min(ax',ay',az'), ， ， The attitude angle (θ, ϕ, ψ) is obtained by the quaternion method based on the three-axis angular velocity ωx, ωy, ωz and the timestamp t. The value range of CNR_th is [27,30].

5. The manhole cover intelligent monitoring algorithm based on Beidou positioning and multivariate machine learning fusion according to claim 1 is characterized in that: The style of each data in the data set DF is: {Δx,Δy,Δv,HDOP',sat_ch,amax,amin,a_ave,a_std,θ,ϕ,ψ,label}.

6. The manhole cover intelligent monitoring algorithm based on Beidou positioning and multivariate machine learning fusion according to claim 1 is characterized in that: The distance calculation of Sdis in step 20 includes the following steps: i When the horizontal precision factor HDOP in is less than the threshold HDOP_th and sat_ch is greater than the threshold sat_ch_th, Sdis(i) is the difference between data and C i Otherwise, Sdis(i) is the distance between data and C. i The Euclidean distance of the vector components after removing Δx, Δy, and Δv, Among them, the value range of HDOP_th is 1-2, and the value range of sat_ch_th is 5-10.

Images