Dam stilling pond partition wall scour prevention stability monitoring system based on edge calculation
By introducing edge computing technology into the anti-shock stability monitoring system of the partition wall of dam Xunlitang, combined with sensors, cloud and edge node modules, the problem of data samples is solved and the real-time and accuracy of monitoring is improved.
Patent Information
- Application Number
- CN202411805789.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2025-03-14
AI Technical Summary
In the anti-shock stability monitoring of partition walls of dams, due to the different frequency of different risk events, the data samples are uneven, which affects the accuracy of classification analysis.
A dam vacuum pond partition wall anti-rush stability monitoring system is adopted based on edge computing, which includes sensor module, cloud module and edge node module. Data is collected through the sensor module, the cloud module conducts timing model training and sample annotation, and real-time data classification prediction and timing prediction are carried out through the edge node module, and early warning is issued in a timely manner.
Through edge computing, data transmission delay is reduced and real-time is improved; through sample supplementation and model training, the performance of classification models is improved; by intelligently selecting the sensor deployment location, manual decision bias is reduced and monitoring accuracy is improved.
Smart Images
Figure CN119622593A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data analysis, and in particular to a dam energy dissipation pond partition wall anti-scour stability monitoring system based on edge computing. Background Art
[0002] As a structure to protect the downstream riverbed, the safety of the dam's energy dissipation pond's own partition wall under the impact of high-speed water flow is the key to achieving the purpose of energy dissipation and anti-scouring. There are many examples of damage at home and abroad. The working and damage mechanism of the high dam energy dissipation pond protection structure is very complex, which is the result of the coupling effect of the high-speed water flow hydrodynamic characteristics and the structural dynamic characteristics of the protection structure.
[0003] When monitoring the anti-scouring stability of the dam energy dissipation pond partition wall, since the various risks faced do not have equal probability of occurrence, some common water scouring conditions may cause one type of risk to occur frequently, and a large number of corresponding data samples will accumulate; while the risks caused by some special geological factors or extreme climatic conditions occur less frequently, which makes the corresponding number of samples scarce, thus affecting the accuracy of classification during data analysis. Summary of the invention
[0004] The purpose of the present invention is to provide a dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing to solve the problems raised in the prior art.
[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a dam energy dissipation pond partition wall anti-scour stability monitoring system based on edge computing, the system comprising a sensor module, a cloud module and an edge node module; The sensor module is used to select the sensor layout location according to the historical report data and arrange the sensor at the corresponding location to collect data; the cloud module is used to train the time series model for the collected data, manually annotate the collected data, balance the number of samples of different classifications, and build a classification model for training; the edge node module is used to perform classification prediction and time series prediction on the newly collected data according to the trained time series model and classification model, and make corresponding warnings when the prediction situation belongs to the preset classification; The output end of the sensor module is connected to the output ends of the edge node module and the cloud module; the output end of the cloud module is connected to the input end of the edge node module; the output end of the edge node module is connected to the input end of the cloud module.
[0006] The sensor module includes a position selection unit, a sensor deployment unit and a data acquisition unit; The location selection unit is used to determine the deployment location of the sensor according to the historical damage record report and the risk assessment report; the sensor deployment unit is used to deploy the sensor according to the selected location; the data acquisition unit is used for the sensor to collect data at a fixed frequency; The output end of the position selection unit is connected to the input end of the sensor deployment unit; the output end of the sensor deployment unit is connected to the input end of the data acquisition unit; the output end of the data acquisition unit is connected to the input end of the cloud module and the edge node module.
[0007] The specific implementation method of the position selection is: define the sensor type: [S1, S2, ..., S n ]; where n is a positive integer, indicating the number of sensor types; S1~S n Indicates the 1st to nth sensors; Obtain the historical damage record report and risk assessment report of the dam stilling pond diaphragm wall; According to the historical damage record report, the historical damage location is obtained, which is expressed as: [L1, L2,…, L a ]; where a is a positive integer, indicating the number of historical damage locations obtained based on historical damage record reports; L1~L a Indicates the 1st to ath historical damage locations; According to the historical risk assessment report, for a sensor S at the same location i , in the period t1~t m The value of V Si (t1)~V Si (t m );Calculate V Si (t1)~V Si (t m ) standard deviation: σ(S i )=((Σ j=1 m (V Si (t1)-V Si (mean) 2 ) / m) 1 / 2 ; Where i is a positive integer, representing the sensor type sequence, i∈{1,2,…,n}, S i represents the i-th sensor; m is a positive integer, representing the number of moments, t1~t m represents the 1st to mth moments; j is a positive integer, representing the time series, j∈{1,2,…,m}; V Si (mean) represents V Si (t1)~V Si (t m ), V Si(mean)=(V Si (t1)+V Si (t2)+…+V Si (t m )) / m; For all sensors at this location, calculate the standard deviation σ(S1)~σ(S n ); calculate σ(S1)~σ(S n ) is expressed as the sensitivity coefficient U: U=(σ(S1)+σ(S2)+…+σ(S n )) / n; Calculate the sensitivity coefficients of all positions included in the historical risk assessment report and select the positions corresponding to the first u values with the largest values: [l1,l2,…,l u ]; where u is a positive integer set by the staff; Combine the historical damage location and the location with the larger sensitivity coefficient value as the sensor deployment location P: P={L1,L2,…,L a}∪{l1,l2,…,l u}.
[0008] The sensor deployment position P is expressed as: [P1, P2, ..., P v ]; where v is a positive integer, indicating the number of deployment locations, P1~P v Indicates the 1st to vth deployment positions; Install a set of sensors at each sensor deployment location; The sensor collects data at a fixed frequency, and the collection time is expressed as: [T1, T2, ..., T c ]; where c is a positive integer, indicating the number of moments, T1~T c Indicates the 1st to cth moments.
[0009] The cloud module includes a database unit, a model building unit, a sample annotation unit, a sample supplement unit and a model training unit; The database unit is used to store data collected by the sensor at the same frequency at different times and model prediction data; the model construction unit is used to build a time series model based on the data collected by the sensor at different times; the sample annotation unit is used to manually perform risk assessment on the data samples collected at different times and classify them into different categories according to the risk assessment status; the sample supplement unit is used to supplement the categories with a small number of samples to balance the number of samples of various categories; the model training unit is used to build a classification model based on the supplemented samples; The output end of the database unit is connected to the input end of the model construction unit and the sample annotation unit; the output end of the model construction unit is connected to the input end of the edge node module; the output end of the sample annotation unit is connected to the input end of the sample supplement unit; the output end of the sample supplement unit is connected to the input end of the model training unit; the output end of the model training unit is connected to the input end of the edge node module.
[0010] The database storage format is: {P1:[V S1 (T1),V S2 (T1),…,V Sn (T1)],[V S1 (T2),V S2 (T2),…,V Sn (T2)],…,[V S1 (T c ),V S2 (T c ),…,V Sn (T c )]},…,{P v :[V S1 (T1),V S2 (T1),…,V Sn (T1)],[V S1 (T2),V S2 (T2),…,V Sn (T2)],…,[V S1 (T c ),V S2 (T c ),…,V Sn (T c )]}; Each deployment point consists of c samples, representing T1~T c Data collected at all times; For deployment point P1, T1~T c The data collected by sensor S1 at this moment is V S1 (T1)~V S1 (T c );Build ARIMA time series model: Use ADF test to test the stability of time series data; Obtain the ARIMA model difference times d value by successive differences until the time series is stable; Calculate the autocorrelation function and partial autocorrelation function of the time series, and determine the autoregressive order p and moving average order q according to the tailing and tailing characteristics of the autocorrelation function and partial autocorrelation function; According to the autoregressive coefficients φ1~φ p and moving average coefficients θ1~θ qConstruct the maximum likelihood function and use the Newton-Raphson algorithm to find the parameter value that makes the likelihood function reach the maximum value; obtain the ARIMA time series model of the sensor S1 at the deployment point P1; For deployment point P1, an ARIMA time series model for each sensor is constructed and trained; an ARIMA time series model for each sensor at each deployment point is constructed and trained.
[0011] Summarize the risk types based on historical risk assessment reports: [K1, K2, …, K e ]; where e is a positive integer, indicating the number of risk types, K1~K e Indicates risk types 1 to e; Use manual labeling to classify v·c samples in the database according to risk types K1~K e Marking For deployment point P1, samples are classified according to different annotated risk types, the number of samples of each type is counted, and the type with the largest number of samples is used as the standard. Threshold A is set, and the difference between the number of samples of each type and the number of samples of the type with the largest number of samples is calculated; when the difference in the number of samples is less than or equal to threshold A, no processing is performed; when the difference in the number of samples is greater than threshold A, samples of this type are supplemented; The specific implementation method of the sample supplement is: when the type K α When the number of samples is the largest, the number of samples is expressed as Num(K α ); when type K β Sample size Num(K β )+A<Num(K α ), for type K β Samples are supplemented; type K is selected β The corresponding samples are rearranged in the order of the corresponding time of the samples, except for the first sample, and the time interval between each sample and the previous sample is calculated and marked as J; the sample time is reordered and expressed as: [t 1 ,t 2 ,…,t b ]; Among them, α and β are positive integers, representing the type sequence; K β , K α ∈{K1,K2,…,K e}; b is a positive integer, indicating the number of moments after reordering, and also indicating type K β The corresponding sample size, t 1 ~t b Indicates the 1st to bth moments; Construct the input matrix: [[V S1 (t 1),V S2 (t 1 ),…,V Sn (t 1 ),0],[V S1 (t 2 ),V S2 (t 2 ),…,V Sn (t 2 ),J(t 2 )],…,[V S1 (t b ),V S2 (t b ),…,V Sn (t b ),J(t b )]] T ; Among them, 0~J(t b ) indicates t 1 ~t b The time interval between the moment and the previous sample before reordering; V S1 (t b )~V Sn (t b ) indicates t b Time corresponding sensor S1~S n The value of The gated recurrent unit GRU time series model is constructed with the input matrix as input; the model is trained by minimizing the mean square error loss function to obtain type K β The GRU time series model is simplified to the function y=f(x); where x represents the input variable, expressed in vector form; y represents the output prediction result, and the format is consistent with x; f(x) represents the training process of the model; By assigning different values to x, the corresponding output prediction results are obtained as supplementary samples; supplementary samples are evaluated manually to determine whether they belong to type K. β Make a judgment; when the supplementary sample belongs to type K β When the supplemented sample does not belong to type K β When the supplementary sample is discarded; When Num(K β )+A=Num(K α ), stop the type K β The samples are supplemented; Construct an SVM classification model based on the supplemented samples of various types: use the sequential minimum optimization algorithm to find the optimal decision boundary to train the model based on different samples and their corresponding types; obtain the SVM classification model applied to the deployment point P1; construct and train the SVM classification model for each deployment point.
[0012] The edge node module includes a time series model unit, a classification model unit and an early warning unit; The time series model unit is used to predict the data at the next moment according to the time series model; the classification model unit is used to classify the data collected at the moment and the data at the next moment predicted by the time series model; the warning unit is used to issue a warning when the classification result belongs to one or more categories; The output end of the timing model unit is connected to the input end of the classification model unit; the output end of the classification model unit is connected to the input end of the early warning unit; the output end of the early warning unit is connected to the input end of the cloud module.
[0013] For each deployment point, the corresponding SVM classification model is used to classify the newly collected sample data, and corresponding warnings are issued based on the classification results; For the newly collected sample data, the corresponding ARIMA time series model is used to perform time series forecasting to obtain the predicted sample data; for the predicted sample data, the classification model is used again to classify it, and an early warning is issued based on the classification situation; The newly collected sample data and the corresponding predicted data are stored for updating the model; the classification of the newly collected sample data and the predicted data are stored for updating the model.
[0014] Compared with the prior art, the beneficial effects of the present invention are: by introducing edge computing, the present invention can quickly process data at the edge node, reduce data transmission delay, and improve real-time performance; the present invention can improve the performance of the classification model by supplementing categories with fewer samples; the present invention intelligently selects the deployment location of sensors through historical damage records, risk assessment reports and sensitivity coefficients; through data analysis, the system can more objectively determine important monitoring points, reduce the deviation of manual decision-making, and improve the accuracy of monitoring. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 It is a structural schematic diagram of the dam energy dissipation pond partition wall anti-scour stability monitoring system based on edge computing of the present invention. DETAILED DESCRIPTION
[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0017] Example: Figure 1 As shown, the present invention provides a technical solution, a dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing, the system includes a sensor module, a cloud module and an edge node module; The sensor module is used to select the sensor layout location according to the historical report data and arrange the sensor at the corresponding location to collect data; the cloud module is used to train the time series model for the collected data and manually annotate the collected data, balance the number of samples of different classifications and build a classification model for training; the edge node module is used to perform classification prediction and time series prediction on the newly collected data according to the trained time series model and classification model, and make corresponding warnings when the prediction situation belongs to the preset classification; The output end of the sensor module is connected to the output ends of the edge node module and the cloud module; the output end of the cloud module is connected to the input end of the edge node module; the output end of the edge node module is connected to the input end of the cloud module.
[0018] The sensor module includes a position selection unit, a sensor deployment unit and a data acquisition unit; The location selection unit is used to determine the deployment location of the sensor according to the historical damage record report and the risk assessment report; the sensor deployment unit is used to deploy the sensor according to the selected location; the data acquisition unit is used for the sensor to collect data at a fixed frequency; The output end of the position selection unit is connected to the input end of the sensor deployment unit; the output end of the sensor deployment unit is connected to the input end of the data acquisition unit; the output end of the data acquisition unit is connected to the input end of the cloud module and the edge node module.
[0019] The specific implementation method of the position selection is: define the sensor type: [S1, S2, ..., S n ]; where n is a positive integer, indicating the number of sensor types; S1~S n Indicates the 1st to nth sensors; Obtain the historical damage record report and risk assessment report of the dam stilling pond diaphragm wall; According to the historical damage record report, the historical damage location is obtained, which is expressed as: [L1, L2,…, L a ]; where a is a positive integer, indicating the number of historical damage locations obtained based on historical damage record reports; L1~L a Indicates the 1st to ath historical damage locations; According to the historical risk assessment report, for a sensor S at the same location i , in the period t1~t m The value of V Si(t1)~V Si (t m );Calculate V Si (t1)~V Si (t m ) standard deviation: σ(S i )=((Σ j=1 m (V Si (t1)-V Si (mean) 2 ) / m) 1 / 2 ; Where i is a positive integer, representing the sensor type sequence, i∈{1,2,…,n}, S i represents the i-th sensor; m is a positive integer, representing the number of moments, t1~t m represents the 1st to mth moments; j is a positive integer, representing the time series, j∈{1,2,…,m}; V Si (mean) represents V Si (t1)~V Si (t m ), V Si (mean)=(V Si (t1)+V Si (t2)+…+V Si (t m )) / m; For all sensors at this location, calculate the standard deviation σ(S1)~σ(S n ); calculate σ(S1)~σ(S n ) is expressed as the sensitivity coefficient U: U=(σ(S1)+σ(S2)+…+σ(S n )) / n; Calculate the sensitivity coefficients of all positions included in the historical risk assessment report and select the positions corresponding to the first u values with the largest values: [l1,l2,…,l u ]; where u is a positive integer set by the staff; Combine the historical damage location and the location with the larger sensitivity coefficient value as the sensor deployment location P: P={L1,L2,…,L a}∪{l1,l2,…,l u}.
[0020] The sensor deployment position P is expressed as: [P1, P2, ..., P v ]; where v is a positive integer, indicating the number of deployment locations, P1~P v Indicates the 1st to vth deployment positions; Install a set of sensors at each sensor deployment location; The sensor collects data at a fixed frequency, and the collection time is expressed as: [T1, T2, ..., T c ]; where c is a positive integer, indicating the number of moments, T1~T c Indicates the 1st to cth moments.
[0021] The cloud module includes a database unit, a model building unit, a sample annotation unit, a sample supplement unit and a model training unit; The database unit is used to store data collected by the sensor at the same frequency at different times and model prediction data; the model construction unit is used to build a time series model based on the data collected by the sensor at different times; the sample annotation unit is used to manually perform risk assessment on the data samples collected at different times and classify them into different categories according to the risk assessment status; the sample supplement unit is used to supplement the categories with a small number of samples to balance the number of samples of various categories; the model training unit is used to build a classification model based on the supplemented samples; The output end of the database unit is connected to the input end of the model construction unit and the sample annotation unit; the output end of the model construction unit is connected to the input end of the edge node module; the output end of the sample annotation unit is connected to the input end of the sample supplement unit; the output end of the sample supplement unit is connected to the input end of the model training unit; the output end of the model training unit is connected to the input end of the edge node module.
[0022] The database storage format is: {P1:[V S1 (T1),V S2 (T1),…,V Sn (T1)],[V S1 (T2),V S2 (T2),…,V Sn (T2)],…,[V S1 (T c ),V S2 (T c ),…,V Sn (T c )]},…,{P v :[V S1 (T1),V S2 (T1),…,V Sn (T1)],[V S1 (T2),V S2 (T2),…,V Sn (T2)],…,[V S1 (T c ),V S2 (Tc ),…,V Sn (T c )]}; Each deployment point consists of c samples, representing T1~T c Data collected at all times; For deployment point P1, T1~T c The data collected by sensor S1 at this moment is V S1 (T1)~V S1 (T c );Build ARIMA time series model: Use ADF test to test the stability of time series data; Obtain the ARIMA model difference times d value by successive differences until the time series is stable; Calculate the autocorrelation function and partial autocorrelation function of the time series, and determine the autoregressive order p and moving average order q according to the tailing and tailing characteristics of the autocorrelation function and partial autocorrelation function; According to the autoregressive coefficients φ1~φ p and moving average coefficients θ1~θ q Construct the maximum likelihood function and use the Newton-Raphson algorithm to find the parameter value that makes the likelihood function reach the maximum value; obtain the ARIMA time series model of the sensor S1 at the deployment point P1; For deployment point P1, an ARIMA time series model for each sensor is constructed and trained; an ARIMA time series model for each sensor at each deployment point is constructed and trained.
[0023] Summarize the risk types based on historical risk assessment reports: [K1, K2, …, K e ]; where e is a positive integer, indicating the number of risk types, K1~K e Indicates risk types 1 to e; Use manual labeling to classify v·c samples in the database according to risk types K1~K e Marking For deployment point P1, samples are classified according to different annotated risk types, the number of samples of each type is counted, and the type with the largest number of samples is used as the standard. Threshold A is set, and the difference between the number of samples of each type and the number of samples of the type with the largest number of samples is calculated; when the difference in the number of samples is less than or equal to threshold A, no processing is performed; when the difference in the number of samples is greater than threshold A, samples of this type are supplemented; The specific implementation method of the sample supplement is: when the type K α When the number of samples is the largest, the number of samples is expressed as Num(K α ); when type K β Sample size Num(K β )+A<Num(K α ), for type K βSamples are supplemented; type K is selected β The corresponding samples are rearranged in the order of the corresponding time of the samples, except for the first sample, and the time interval between each sample and the previous sample is calculated and marked as J; the sample time is reordered and expressed as: [t 1 ,t 2 ,…,t b ]; Among them, α and β are positive integers, representing the type sequence; K β , K α ∈{K1,K2,…,K e}; b is a positive integer, indicating the number of moments after reordering, and also indicating type K β The corresponding sample size, t 1 ~t b Indicates the 1st to bth moments; Construct the input matrix: [[V S1 (t 1 ),V S2 (t 1 ),…,V Sn (t 1 ),0],[V S1 (t 2 ),V S2 (t 2 ),…,V Sn (t 2 ),J(t 2 )],…,[V S1 (t b ),V S2 (t b ),…,V Sn (t b ),J(t b )]] T ; Among them, 0~J(t b ) indicates t 1 ~t b The time interval between the moment and the previous sample before reordering; V S1 (t b )~V Sn (t b ) indicates t b Time corresponding sensor S1~S n The value of The gated recurrent unit GRU time series model is constructed with the input matrix as input; the model is trained by minimizing the mean square error loss function to obtain type K βThe GRU time series model is simplified to the function y=f(x); where x represents the input variable, expressed in vector form; y represents the output prediction result, and the format is consistent with x; f(x) represents the training process of the model; By assigning different values to x, the corresponding output prediction results are obtained as supplementary samples; supplementary samples are evaluated manually to determine whether they belong to type K. β Make a judgment; when the supplementary sample belongs to type K β When the supplemented sample does not belong to type K β When the supplementary sample is discarded; When Num(K β )+A=Num(K α ), stop the type K β The samples are supplemented; Construct an SVM classification model based on the supplemented samples of various types: use the sequential minimum optimization algorithm to find the optimal decision boundary to train the model based on different samples and their corresponding types; obtain the SVM classification model applied to the deployment point P1; construct and train the SVM classification model for each deployment point.
[0024] The edge node module includes a time series model unit, a classification model unit and an early warning unit; The time series model unit is used to predict the data at the next moment according to the time series model; the classification model unit is used to classify the data collected at the moment and the data at the next moment predicted by the time series model; the warning unit is used to issue a warning when the classification result belongs to one or more categories; The output end of the timing model unit is connected to the input end of the classification model unit; the output end of the classification model unit is connected to the input end of the early warning unit; the output end of the early warning unit is connected to the input end of the cloud module.
[0025] For each deployment point, the corresponding SVM classification model is used to classify the newly collected sample data, and corresponding warnings are issued based on the classification results; For the newly collected sample data, the corresponding ARIMA time series model is used to perform time series forecasting to obtain the predicted sample data; for the predicted sample data, the classification model is used again to classify it, and an early warning is issued based on the classification situation; The newly collected sample data and the corresponding predicted data are stored for updating the model; the classification of the newly collected sample data and the predicted data are stored for updating the model.
[0026] In this embodiment: The sensor type is [pressure sensor, displacement sensor, vibration sensor]; Obtain the historical damage record report of the dam stilling pond diaphragm in the past five years, covering multiple flood seasons and other working conditions, as well as the annual risk assessment report issued by a professional organization. From the historical damage record report, the historical damage locations [L1, L2] were sorted out, which were located at the bottom of the diaphragm near the outlet L1 and the middle of the diaphragm where water flow frequently impacted the area L2.
[0027] For each sensor at the same position, the values taken at the past 10 moments t1-t10, after calculating the standard deviation σ of the pressure sensor S1, are [10,12,9,11,13,10,12,9,11,13], and its mean V(mean)=(10+12+9+11+13+10+12+9+11+13) / 10=11. According to the standard deviation calculation formula, σ=1.41. Similarly, the standard deviation of the displacement sensor S2 is σ(S2)=1.2, the standard deviation of the vibration sensor S3 is σ(S3)=1.6, and the sensitivity coefficient U=(1.41+1.2+1.6) / 3=1.4. According to this method, the sensitivity coefficients of all locations included in the historical risk assessment reports were calculated, and the first three corresponding locations with the largest values were selected, [l1 (near the junction between the dam foundation and the partition wall), l2 (the bend at the top of the partition wall impacted by the water jump), l3 (the turbulent water flow impact area on the side of the partition wall)].
[0028] Finalize the sensor deployment location: P={L1,L2}∪{l1,l2,l3}=[P1(L1),P2(L2),P3(l1),P4(l2),P5(l3)]; A group of sensors is installed at each sensor deployment location. Each group contains one pressure sensor, one displacement sensor, and one vibration sensor. The sensors collect data at a fixed frequency of once every 30 minutes. The collection time is expressed as: [T1 (0:00), T2 (0:30), T3 (1:00), …, Tc (23:30)]. It is set to collect data 48 times a day.
[0029] The sensor collected data is stored in the set format. For example, for the deployment point P1, the storage format is: {P1: [[VS1(T1), VS2(T1), VS3(T1)], [VS1(T2), VS2(T2), VS3(T2)], …, [VS1(T48), VS2(T48), VS3(T48)]]}, and so on to store the data of other deployment points P2-P5.
[0030] Taking the deployment point P1 sensor S1 as an example, an ARIMA time series model is constructed, a maximum likelihood function is constructed, and the parameters are solved by the Newton-Raphson algorithm to obtain the ARIMA time series model of the deployment point P1 sensor S1. Similarly, the ARIMA model of each sensor at each deployment point is constructed.
[0031] According to the historical risk assessment report, the risk types are summarized as follows: [K1 (minor scour risk), K2 (moderate displacement risk), K3 (severe vibration risk)]. 5*48=240 samples in the database are manually labeled according to the risk type. After the deployment point P1 is labeled, there are 80 K1 type samples, 60 K2 type samples, and 40 K3 type samples. The K1 type with the largest number of samples is used as the standard, and the threshold A=20 is set. For the K3 type, the number of samples is 40+20<80, and additional samples are required. The samples corresponding to K3 are selected and reordered, the interval J is calculated, and the input matrix is constructed. The gated recurrent unit GRU time series model is constructed with it as the input. The GRU time series model of K3 is trained, and after multiple attempts, different values of x are assigned to generate additional samples. After manual evaluation, appropriate samples are added until the number of K3 samples reaches 60.
[0032] An SVM classification model is constructed based on the supplemented samples of various types. Based on different samples and their corresponding types, a sequential minimum optimization algorithm is used to find the optimal decision boundary training model, and the SVM classification model applied to each deployment point P1-P5 is obtained.
[0033] At the edge node, for the newly collected sample data, the data at the next moment is predicted based on the corresponding ARIMA time series model.
[0034] Use the SVM classification model corresponding to P1 to classify the newly collected sample data and the T49 time data predicted by the time series model to determine the risk type to which it belongs.
[0035] If the classification result belongs to the severe vibration risk K3 type, the early warning unit will immediately issue an early warning and send the early warning information and related data back to the cloud module for subsequent model update and analysis. At the same time, the newly collected sample data and the classification of the predicted data are stored for subsequent model update and optimization.
[0036] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.
Claims
1. The dam energy dissipation pond partition wall anti-scour stability monitoring system based on edge computing is characterized by: The system includes a sensor module, a cloud module and an edge node module; The sensor module is used to select the sensor layout location according to the historical report data and arrange the sensor at the corresponding location to collect data; the cloud module is used to train the time series model for the collected data, manually annotate the collected data, balance the number of samples of different classifications, and build a classification model for training; the edge node module is used to perform classification prediction and time series prediction on the newly collected data according to the trained time series model and classification model, and make corresponding warnings when the prediction situation belongs to the preset classification; The output end of the sensor module is connected to the output ends of the edge node module and the cloud module; the output end of the cloud module is connected to the input end of the edge node module; the output end of the edge node module is connected to the input end of the cloud module.
2. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 1 is characterized by: The sensor module includes a position selection unit, a sensor deployment unit and a data acquisition unit; The location selection unit is used to determine the deployment location of the sensor according to the historical damage record report and the risk assessment report; the sensor deployment unit is used to deploy the sensor according to the selected location; the data acquisition unit is used for the sensor to collect data at a fixed frequency; The output end of the position selection unit is connected to the input end of the sensor deployment unit; the output end of the sensor deployment unit is connected to the input end of the data acquisition unit; the output end of the data acquisition unit is connected to the input end of the cloud module and the edge node module.
3. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 2 is characterized by: The specific implementation method of the position selection is: define the sensor type: [S1, S2, ..., S n ]; where n is a positive integer, indicating the number of sensor types; S1~S n Indicates the 1st to nth sensors; Obtain the historical damage record report and risk assessment report of the dam stilling pond diaphragm wall; The historical damage location is obtained according to the historical damage record report, which is expressed as: [L1, L2,…, L a ]; where a is a positive integer, indicating the number of historical damage locations obtained based on historical damage record reports; L1~L a Indicates the 1st to ath historical damage locations; According to the historical risk assessment report, for a sensor S at the same location i , in the period t1~t m The value of V Si (t1)~V Si (t m );Calculate V Si (t1)~V Si (t m ) standard deviation: σ(S i )=((Σ j=1 m (V Si (t1)-V Si (mean) 2 ) / m) 1 / 2 ; Where i is a positive integer, representing the sensor type sequence, i∈{1,2,…,n}, S i represents the i-th sensor; m is a positive integer, representing the number of moments, t1~t m represents the 1st to mth moments; j is a positive integer, representing the time series, j∈{1,2,…,m}; V Si (mean) represents V Si (t1)~V Si (t m ), V Si (mean)=(V Si (t1)+V Si (t2)+…+V Si (t m )) / m; For all sensors at this location, calculate the standard deviation σ(S1)~σ(S n ); calculate σ(S1)~σ(S n ) is expressed as the sensitivity coefficient U: U=(σ(S1)+σ(S2)+…+σ(S n )) / n; Calculate the sensitivity coefficients of all positions included in the historical risk assessment report and select the positions corresponding to the first u values with the largest values: [l1,l2,…,l u ]; where u is a positive integer set by the staff; Combine the historical damage location and the location with the larger sensitivity coefficient value as the sensor deployment location P: P={L1,L2,…,L a }∪{l1,l2,…,l u }.
4. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 3 is characterized by: The sensor deployment position P is expressed as: [P1, P2, ..., P v ]; where v is a positive integer, indicating the number of deployment locations, P1~P v Indicates the 1st to vth deployment positions; Install a set of sensors at each sensor deployment location; The sensor collects data at a fixed frequency, and the collection time is expressed as: [T1, T2, ..., T c ]; where c is a positive integer, indicating the number of moments, T1~T c Indicates the 1st to cth moments.
5. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 4 is characterized by: The cloud module includes a database unit, a model building unit, a sample annotation unit, a sample supplement unit and a model training unit; The database unit is used to store data collected by the sensor at the same frequency at different times and model prediction data; the model construction unit is used to build a time series model based on the data collected by the sensor at different times; the sample annotation unit is used to manually perform risk assessment on the data samples collected at different times and classify them into different categories according to the risk assessment status; the sample supplement unit is used to supplement the categories with a small number of samples to balance the number of samples of various categories; the model training unit is used to build a classification model based on the supplemented samples; The output end of the database unit is connected to the input end of the model construction unit and the sample annotation unit; the output end of the model construction unit is connected to the input end of the edge node module; the output end of the sample annotation unit is connected to the input end of the sample supplement unit; the output end of the sample supplement unit is connected to the input end of the model training unit; the output end of the model training unit is connected to the input end of the edge node module.
6. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 5 is characterized by: The database storage format is: {P1:[V S1 (T1),V S2 (T1),…,V Sn (T1)],[V S1 (T2),V S2 (T2),…,V Sn (T2)],…,[V S1 (T c ),V S2 (T c ),…,V Sn (T c )]},…,{P v :[V S1 (T1),V S2 (T1),…,V Sn (T1)],[V S1 (T2),V S2 (T2),…,V Sn (T2)],…,[V S1 (T c ),V S2 (T c ),…,V Sn (T c )]}; Each deployment point consists of c samples, representing T1~T c Data collected at all times; For deployment point P1, T1~T c The data collected by sensor S1 at this moment is V S1 (T1)~V S1 (T c );Build ARIMA time series model: Use ADF test to test the stability of time series data; Obtain the ARIMA model difference times d value by successive differences until the time series is stable; Calculate the autocorrelation function and partial autocorrelation function of the time series, and determine the autoregressive order p and moving average order q according to the tailing and tailing characteristics of the autocorrelation function and partial autocorrelation function; According to the autoregressive coefficients φ1~φ p and moving average coefficients θ1~θ q Construct the maximum likelihood function and use the Newton-Raphson algorithm to find the parameter value that makes the likelihood function reach the maximum value; Get the ARIMA time series model of the sensor S1 at the deployment point P1; For deployment point P1, an ARIMA time series model for each sensor is constructed and trained; an ARIMA time series model for each sensor at each deployment point is constructed and trained.
7. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 6 is characterized by: Summarize the risk types based on historical risk assessment reports: [K1, K2, …, K e ]; where e is a positive integer, indicating the number of risk types, K1~K e Indicates risk types 1 to e; Use manual labeling to classify v·c samples in the database according to risk types K1~K e Marking For deployment point P1, samples are classified according to different annotated risk types, the number of samples of each type is counted, and the type with the largest number of samples is used as the standard. Threshold A is set, and the difference between the number of samples of each type and the number of samples of the type with the largest number of samples is calculated; when the difference in the number of samples is less than or equal to threshold A, no processing is performed; when the difference in the number of samples is greater than threshold A, samples of this type are supplemented; The specific implementation method of the sample supplement is: when the type K α When the number of samples is the largest, the number of samples is expressed as Num(K α ); when type K β Sample size Num(K β )+A<Num(K α ), for type K β Samples are supplemented; type K is selected β The corresponding samples are rearranged in the order of the corresponding time of the samples, except for the first sample, and the time interval between each sample and the previous sample is calculated and marked as J; the sample time is reordered and expressed as: [t 1 ,t 2 ,…,t b ]; Among them, α and β are positive integers, representing the type sequence; K β , K α ∈{K1,K2,…,K e }; b is a positive integer, indicating the number of moments after reordering, and also indicating type K β The corresponding sample size, t 1 ~t b Indicates the 1st to bth moments; Construct the input matrix: [[V S1 (t 1 ),V S2 (t 1 ),…,V Sn (t 1 ),0],[V S1 (t 2 ),V S2 (t 2 ),…,V Sn (t 2 ),J(t 2 )],…,[V S1 (t b ),V S2 (t b ),…,V Sn (t b ),J(t b )]] T ; Among them, 0~J(t b ) indicates t 1 ~t b The time interval between the moment and the previous sample before reordering; V S1 (t b )~V Sn (t b ) indicates t b Time corresponding sensor S1~S n The value of The gated recurrent unit GRU time series model is constructed with the input matrix as input; the model is trained by minimizing the mean square error loss function to obtain type K β The GRU time series model is simplified to the function y=f(x); where x represents the input variable, expressed in vector form; y represents the output prediction result, and the format is consistent with x; f(x) represents the training process of the model; By assigning different values to x, the corresponding output prediction results are obtained as supplementary samples; supplementary samples are evaluated manually to determine whether they belong to type K. β Make a judgment; when the supplementary sample belongs to type K β When the supplemented sample does not belong to type K β When the supplementary sample is discarded; When Num(K β )+A=Num(K α ), stop the type K β The samples are supplemented; Construct an SVM classification model based on the supplemented samples of various types: use the sequential minimum optimization algorithm to find the optimal decision boundary to train the model based on different samples and their corresponding types; obtain the SVM classification model applied to the deployment point P1; construct and train the SVM classification model for each deployment point.
8. The dam energy dissipation pond partition wall anti-scour stability monitoring system based on edge computing according to claim 7 is characterized by: The edge node module includes a time series model unit, a classification model unit and an early warning unit; The time series model unit is used to predict the data at the next moment according to the time series model; The classification model unit is used to classify the data collected at this moment and the data at the next moment predicted by the time series model; the warning unit is used to issue a warning when the classification result belongs to one or more categories; The output end of the timing model unit is connected to the input end of the classification model unit; the output end of the classification model unit is connected to the input end of the early warning unit; the output end of the early warning unit is connected to the input end of the cloud module.
9. The dam energy dissipation pond partition wall anti-scouring stability monitoring system based on edge computing according to claim 8 is characterized by: For each deployment point, the corresponding SVM classification model is used to classify the newly collected sample data, and corresponding warnings are issued based on the classification results; For the newly collected sample data, the corresponding ARIMA time series model is used to perform time series forecasting to obtain the predicted sample data; for the predicted sample data, the classification model is used again to classify it, and an early warning is issued based on the classification situation; The newly collected sample data and the corresponding predicted data are stored for updating the model; the classification of the newly collected sample data and the predicted data are stored for updating the model.