A groundwater monitoring system based on convolutional neural network

Abstract

The application relates to the technical field of groundwater monitoring investigation and evaluation systems, in particular to a groundwater monitoring system based on a convolutional neural network, which comprises a data unit, a processing unit connected with the data unit in a signal mode, a screening unit and a monitoring unit connected with the screening unit in a signal mode, is loaded with a convolutional neural network model, and is used for dividing a second sequence matrix into a training set and a test set; the convolutional neural network model can output a monitoring set by inputting the test set after being trained by the training set. The first sequence matrix is taken as basic data, a KNN algorithm based on distance weighting is used to screen the spatial correlation of the first sequence matrix, and a second sequence matrix is constructed according to the screening result; the second sequence matrix is divided into the training set and the test set, the training set is used for model training of the neural convolution network, the test set is used for model verification and input of a monitoring target; and finally, target monitoring data is obtained by weighting and fusing output values of the neural convolution network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of groundwater monitoring investigation and evaluation system, and particularly relates to a groundwater monitoring system based on a convolutional neural network. BACKGROUND

[0002] As an important support for groundwater environmental monitoring work, the health of groundwater environmental monitoring wells directly affects the accuracy of monitoring results. Due to the lack of professional knowledge and the fact that the management department has not yet developed a systematic maintenance and management method, some monitoring wells have problems such as inadequate post-maintenance management and loss of monitoring function, and even become pollution channels. For monitoring wells that are still in operation, it is also necessary to evaluate whether they can meet the requirements of groundwater monitoring work, and to strengthen inspection and maintenance and unified management according to the evaluation results.

[0003] Due to the influence of environmental factors and collection strategies, the collected water quality monitoring data often deviates from the actual situation, which affects the accuracy of water quality detection.

[0004] In addition, during the monitoring of groundwater, the search for abnormal data is mainly carried out manually. However, continuous manual analysis of abnormal data is prone to errors and cannot ensure the stability of the monitoring data. Manual search for abnormal data takes a long time and cannot guarantee the timeliness of the data. SUMMARY

[0005] The present application aims to provide a groundwater monitoring system based on a convolutional neural network, which is used to reduce the influence of abnormal values of original data on the monitoring process, and on this basis, to improve the monitoring accuracy by training using a sequence matrix that has spatial correlation with the monitoring target.

[0006] The present application is realized by the following technical scheme:

[0007] The application discloses a groundwater monitoring system based on a convolutional neural network, comprising a monitoring well, and a data unit for collating groundwater level space-time monitoring data and compiling a data set and collecting groundwater data in the monitoring well; a processing unit connected with the data unit and used for pre-processing the data set to obtain a first sequence matrix; a screening unit connected with the processing unit and used for screening the first sequence matrix for spatial correlation and constructing a second sequence matrix according to the screening result; and a monitoring unit connected with the screening unit and loaded with a convolutional neural network model, and used for dividing the second sequence matrix into a training set and a test set, wherein the convolutional neural network model can output a monitoring set by inputting the test set after training, and the monitoring unit compares the received groundwater data with the monitoring set, and outputs the groundwater data if a threshold condition is met, or repeats the groundwater data collection if the threshold condition is not met.

[0008] It should be noted that most of the traditional groundwater level dynamic monitoring methods only monitor and study the time series data, the method is relatively single, and the spatial factor is rarely or even not considered. The groundwater level monitoring value is a typical geographical space-time sequence data, and its dynamic change is not only a natural geographical phenomenon, but also a result of the joint action of different regional human economic activities. The dynamic change of groundwater and the monitoring simulation should be considered from the time and space multiple influences, not only affected by the historical data of the target monitoring well, but also affected by the data of other monitoring wells related to the target monitoring well.

[0009] Based on the above problems, a groundwater monitoring system based on a convolutional neural network is provided, which collates groundwater level space-time monitoring data and compiles a data set, eliminates abnormal values and reduces the adverse effects of abnormal values on the monitoring reasoning process, that is, the first sequence matrix of the groundwater level data with reduced noise is obtained by wavelet analysis of the data set; and the first sequence matrix is taken as the basic data, the KNN algorithm based on distance weighting is used to screen the first sequence matrix for spatial correlation, and the second sequence matrix is constructed according to the screening result; then the second sequence matrix is divided into a training set and a test set, the training set is used for model training of the neural convolution network, the test set is used for model verification and input of the monitoring target; finally, the output values of the neural convolution network are weighted and fused to obtain the target monitoring data.

[0010] Further, the processing unit comprises: a signal decomposition module configured to decompose original signals in the data set to obtain scale coefficients; a threshold processing module connected with the signal decomposition module and configured to screen and retain the decomposed scale coefficients; and a signal reconstruction module connected with the threshold processing module and configured to reconstruct the discrete signals after threshold processing into continuous signals, i.e., to obtain smooth real signals after denoising, and the set of real signals is a first sequence matrix. It should be noted that, due to large changes in the time series, individual data may exist in the periodic changes. After checking the original data, it is found that the causes of abnormal values include incorrect entry of actual monitoring data of the monitoring well and influence of overexploitation of adjacent monitoring wells, and thus it is difficult to accurately monitor the groundwater level in the study area. Therefore, in subsequent construction of the groundwater level monitoring model, the original monitoring well data needs to be preprocessed, such as denoising, to improve the accuracy of the model.

[0011] Based on the above content, according to the principle that a time domain signal is converted into frequency spectrum energy distributed in different frequency bands in a frequency domain through wavelet transform, high-frequency noise interference is suppressed by setting a reasonable frequency spectrum energy threshold, and high-frequency information and high-frequency noise are effectively distinguished.

[0012] Further, the screening unit selects a sequence set having spatial correlation with the monitoring target by using a KNN algorithm based on distance weighting and constructs a second sequence matrix. According to the spatial and temporal characteristics of the groundwater level data and the characteristics of the deep learning algorithm, the groundwater level monitoring sequence has obvious spatial and temporal correlation characteristics, the neural network model has good self-adaptation and self-learning ability, the KNN algorithm is introduced for spatial correlation screening, the spatial and temporal data set and the neural network model algorithm are reconstructed, the hybrid monitoring model is constructed, and the model is suitable for monitoring of the spatial and temporal sequence of the groundwater level.

[0013] Further, the screening unit is provided with an Euler distance threshold, and the selection condition is less than the Euler distance threshold, and the second sequence matrix is arranged in ascending order based on the size of the Euler distance. It should be noted that the present application uses the Euclidean distance to measure the similarity degree between the groundwater level sequences of the monitoring wells.

[0014] Further, the monitoring unit comprises an expansion module, the expansion module is configured to extract features of the second sequence matrix based on the convolutional neural network model, to obtain dense relationships after extraction, and to obtain expanded data reasoning rules based on the dense relationships, wherein the expanded data reasoning rules comprise a supplement rule, an and rule, a place inverse rule, a transmission rule and an augmentation rule.

[0015] Based on the knowledge graph of groundwater monitoring, this study investigates a sparse representation method for the reasoning relationships of groundwater monitoring objects. Convolutional neural networks are used for feature extraction, resulting in dense relationships. Based on these dense relationships, expanded data reasoning rules are obtained and conflict detection is performed. After conflict detection, new reasoning rules are assigned practical interpretations, resulting in a set of reasoning rules. The expanded knowledge reasoning rules include complementary rules, union rules, reflexive rules, transitive rules, and augmentation rules, thereby expanding, supplementing, and enriching the knowledge reasoning rule set.

[0016] Furthermore, the monitoring unit also includes a training module signal-connected to the extension module. This training module inputs the training set into the convolutional neural network model and completes the training process based on the data inference rules. By continuously adjusting key model parameters and using cross-validation to verify the model's predictive performance, an optimal convolutional neural network model is constructed.

[0017] Furthermore, after training, the convolutional neural network model inputs a test set and outputs a monitoring set. The monitoring unit also includes a fusion module connected to the training module. The fusion module can fuse the monitoring set according to the weight values ​​in the second sequence matrix to obtain target monitoring data.

[0018] Furthermore, the convolutional neural network model is a GRU neural network model, and the hidden layers of the GRU neural network model are enhanced with Dense layers, and the output layer is enhanced with a softmax layer. The GRU network is a variant of the LSTM neural network, a type of neural network with temporal prediction properties, suitable for predicting temporal data. The network parameters of the GRU neural network are easy to adjust, and its network structure is simpler than that of LSTM. Similarly utilizing the temporal nature of data information, by improving the GRU network, the optimal combination of data inference rules under different environmental conditions can be predicted. The improvement to the GRU network here is the addition of Dense layers to the hidden layers and a softmax layer to the final output layer, which improves the transformation capability of the feature space, converting the prediction of data inference into the prediction of data volatility.

[0019] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0020] 1. This invention decomposes a dataset using wavelet analysis to obtain a first sequence matrix of groundwater level data with reduced noise. Using this first sequence matrix as the base data, a distance-weighted KNN algorithm is employed to filter the spatial correlation of the first sequence matrix, and a second sequence matrix is ​​constructed based on the filtering results. The second sequence matrix is ​​then divided into a training set and a test set. The training set is used for training the neural convolutional network model, and the test set is used for model validation and as input to the monitoring target. Finally, the output values ​​of the neural convolutional network are weighted and fused to obtain the target monitoring data.

[0021] 2. This invention constructs an optimal convolutional neural network model by continuously adjusting key parameters of the model and using cross-validation to verify the model's predictive performance.

[0022] 3. This invention adds a Dense layer to the hidden layer of the convolutional neural network model and a softmax layer to the final output layer, thereby improving the transformation capability of the feature space and transforming the prediction of data inference into the prediction of data volatility. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0024] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0025] Figure 2 This is a schematic diagram of the processing unit of the present invention.

[0026] Figure 3 This is a data comparison chart for Example 3;

[0027] Figure 4 This is a schematic diagram of the structure of Example 4.

[0028] The attached diagram shows the markings and corresponding component names:

[0029] 1-Enclosed unit, 11-Enclosed seat, 12-Enclosed cover, 13-Enclosed component, 14-Bracket, 15-Slide plate, 16-Slide hole, 17-Slide cylinder, 18-Claw, 19-Limiting cylinder, 20-First connecting rod, 21-Second connecting rod, 22-Monitoring cylinder. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are for illustrative purposes only and are not intended to limit the invention. It should be noted that this invention is already in the actual research and development stage.

[0031] Example 1:

[0032] As attached Figure 1 and Figure 2 As shown, a groundwater monitoring system based on a convolutional neural network includes a monitoring well. The system comprises: a data unit for organizing spatiotemporal monitoring data of groundwater levels and compiling a dataset, and for collecting groundwater data within the monitoring well; a processing unit connected to the data unit for preprocessing the dataset to obtain a first sequence matrix; a filtering unit connected to the processing unit for performing spatial correlation filtering on the first sequence matrix as base data, and constructing a second sequence matrix based on the filtering results; and a monitoring unit connected to the filtering unit, equipped with a convolutional neural network model, for dividing the second sequence matrix into a training set and a test set. The convolutional neural network model, after being trained on the training set, can output a monitoring set after inputting the test set. The monitoring unit compares the received groundwater data with the monitoring set. If a threshold condition is met, the groundwater data is output; otherwise, the groundwater data collection is repeated.

[0033] The groundwater data includes water level data, water temperature data, water quality data, etc. The threshold condition can be that the deviation between the groundwater data and the monitoring set data is within 20%. If the threshold condition is not met after 5 repeated collections, the average value of the groundwater data will be output.

[0034] It should be noted that most traditional groundwater level dynamic monitoring methods only monitor and study its time-series data, making the methods relatively simplistic and rarely, if ever, considering spatial factors. Groundwater level monitoring values, however, are typical geographic spatiotemporal series data; their dynamic changes are both a natural geographical phenomenon and a result of the combined effects of human economic activities in different regions. Groundwater dynamic changes and monitoring simulations should consider multiple temporal and spatial influences, not only being affected by the historical data of the target monitoring well but also by the influence of data from other monitoring wells related to the target well.

[0035] It should also be noted that, based on the above problems, a groundwater monitoring system based on convolutional neural networks is proposed. By organizing the spatiotemporal monitoring data of groundwater levels and compiling a dataset, outliers are removed, reducing their adverse impact on the monitoring inference process. Specifically, wavelet analysis is performed on the dataset to decompose it into a first sequence matrix of groundwater level data with reduced noise. Using the first sequence matrix as the basic data, a distance-weighted KNN algorithm is used to filter the spatial correlation of the first sequence matrix, and a second sequence matrix is ​​constructed based on the filtering results. The second sequence matrix is ​​then divided into a training set and a test set. The training set is used for model training of the neural convolutional network, and the test set is used for model validation and as input to the monitoring target. Finally, the output values ​​of the neural convolutional network are weighted and fused to obtain the target monitoring data.

[0036] Wavelet analysis enables localized analysis across time (space) and frequency. High-frequency signals exhibit high time resolution and low frequency resolution, while low-frequency signals possess high frequency resolution and low time resolution. Based on this characteristic, the signal is decomposed into a superposition of wavelet functions. Multi-scale refinement analysis of the signal (function) is achieved through scaling, translation, and other operations. Adaptive time-frequency analysis is realized by appropriately adjusting the time and frequency windows, ultimately achieving time subdivision at high frequencies and frequency subdivision at low frequencies. In this application, discrete wavelet transform is used to discretize the scale parameters in continuous wavelets, which is simple to implement and requires less time.

[0037] It should be noted that the processing unit includes: a signal decomposition module, used to decompose the original signal in the dataset to obtain coefficients at each scale; a threshold processing module, connected to the signal decomposition module, used to filter and retain the decomposed coefficients at each scale; and a signal reconstruction module, connected to the threshold processing module, used to reconstruct the threshold-processed discrete signal into a continuous signal, i.e., to obtain a denoised and smooth real signal, the set of which is the first sequence matrix. It should be noted that due to the large fluctuations in the time series, anomalies in periodic variations of individual data cannot be ruled out. Verification of the original data revealed that the outliers were caused by errors in the actual monitoring data entry of the monitoring wells and the influence of over-exploitation by adjacent monitoring wells, making accurate monitoring of the groundwater level in the study area difficult. Therefore, in the subsequent construction of the groundwater level monitoring model, data preprocessing such as denoising of the original monitoring well data is necessary to improve the model's accuracy.

[0038] Based on the above, and according to the principle that time-domain signals are transformed into spectral energy distributed in different frequency bands in the frequency domain through wavelet transform, high-frequency noise interference can be suppressed by setting a reasonable spectral energy threshold, thus effectively distinguishing high-frequency information from high-frequency noise.

[0039] It should be noted that the filtering unit uses a distance-weighted KNN algorithm to select a set of sequences that are spatially correlated with the monitoring target and constructs a second sequence matrix. Based on the spatiotemporal characteristics of groundwater level data and the characteristics of deep learning algorithms, it is known that groundwater level monitoring sequences have obvious spatiotemporal correlation characteristics. Neural network models have good adaptive and self-learning capabilities. By introducing the KNN algorithm for spatial correlation filtering, reconstructing the spatiotemporal dataset and neural network model algorithm, a hybrid monitoring model is constructed, suitable for monitoring spatiotemporal sequences of groundwater levels.

[0040] It should be noted that the filtering unit has an Eulerian distance threshold, the selection condition is that the distance is less than the Eulerian distance threshold, and the second sequence matrix is ​​arranged in ascending order based on the magnitude of the Eulerian distance. It should also be noted that this application uses Eulerian distance to measure the similarity between groundwater level sequences of each monitoring well.

[0041] It should be noted that the monitoring unit includes an extension module. The extension module performs feature extraction on the second sequence matrix based on the convolutional neural network model. After extraction, dense relationships are obtained, and then extended data inference rules are obtained based on the dense relationships. The extended data inference rules include: complement rules, union rules, inverse rules, transitive rules, and augmentation rules.

[0042] Based on the knowledge graph of groundwater monitoring, this study investigates a sparse representation method for the reasoning relationships of groundwater monitoring objects. Convolutional neural networks are used for feature extraction, resulting in dense relationships. Based on these dense relationships, expanded data reasoning rules are obtained and conflict detection is performed. After conflict detection, new reasoning rules are assigned practical interpretations, resulting in a set of reasoning rules. The expanded knowledge reasoning rules include complementary rules, union rules, reflexive rules, transitive rules, and augmentation rules, thereby expanding, supplementing, and enriching the knowledge reasoning rule set.

[0043] It should be noted that the monitoring unit also includes a training module connected to the extension module. This training module inputs the training set into the convolutional neural network model and completes the training process based on the data inference rules. By continuously adjusting key model parameters and using cross-validation to verify the model's predictive performance, an optimal convolutional neural network model is constructed.

[0044] It should be noted that after training, the convolutional neural network model outputs a monitoring set after inputting a test set. The monitoring unit also includes a fusion module connected to the training module. The fusion module can fuse the monitoring set according to the weight values ​​in the second sequence matrix to obtain the target monitoring data.

[0045] It should be noted that the convolutional neural network model described is a GRU neural network model, with added Dense layers in the hidden layers and a softmax layer in the output layer. The GRU network is a variant of the LSTM neural network, a type of neural network with temporal prediction properties, suitable for predicting time-series data. The network parameters of the GRU neural network are easy to adjust, and its network structure is simpler than that of LSTM. Similarly utilizing the temporal nature of data information, by improving the GRU network, the optimal combination of data inference rules under different environmental conditions can be predicted. The improvement to the GRU network here involves adding Dense layers to the hidden layers and a softmax layer to the final output layer, thus improving the feature space transformation capability and converting the prediction of data inference into the prediction of data volatility. For the hidden layers, feedback connections are added, and the output value of each hidden layer is determined by the current input value and the output value of the previous time step. The GRU neural network model controls the input, forgetting, and output of information through the combined action of three gate structures in the memory block, allowing information to be retained for a long time during network transmission. It is suitable for building time series models, improving the stability and reliability of the models, and can be used to predict groundwater level time series.

[0046] Regarding the specific monitoring process:

[0047] 1. Divide the denoised dataset into sample sequences and target sequences;

[0048] 2. Calculate the Euclidean distance d between each monitoring well and the target monitoring well. For any monitoring well at any location and the target monitoring well, evaluate the true distance between the two monitoring wells by calculating the sum of the squares of the differences in the kilometer grid coordinates between the two monitoring wells.

[0049] 3. Sort the monitoring wells in ascending order based on the calculated Euclidean distance d between them;

[0050] 4. Select the denoised groundwater level sequences of the top K distance-corresponding monitoring wells and construct a sequence matrix of groundwater level monitoring wells for D consecutive years after spatial filtering.

[0051] 5. Select the data of the K monitoring wells after spatial filtering for the previous D-1 years as the training set, input them into the GRU neural network model for learning and training, output the predicted water level value of these K monitoring wells in the D year, and fuse the predicted values ​​of each monitoring well according to different weight values ​​to obtain the groundwater level prediction data of the target monitoring well in the D year.

[0052] In this process, the prediction error of the target monitoring well is calculated for different K values. The prediction result with the smallest prediction error is selected as the prediction data before the model is optimized. At this time, the K value is the K monitoring wells with the greatest spatial correlation with the monitoring wells corresponding to the final spatial screening result.

[0053] Example 2:

[0054] This embodiment only describes the parts that differ from Embodiment 1, specifically:

[0055] For wavelet thresholding denoising, based on the principle that a time-domain signal is transformed into spectral energy distributed across different frequency bands in the frequency domain through wavelet transform, a reasonable spectral energy threshold is set to suppress high-frequency noise interference, effectively distinguishing high-frequency information from high-frequency noise. Assuming the original groundwater level dataset is X, the wavelet thresholding denoising algorithm is as follows:

[0056] 1. Wavelet decomposition: Perform wavelet transform on the original dataset X to obtain its wavelet coefficients. X ~;

[0057] 2. Threshold processing: To separate the signals, a threshold is set. α and with X Comparing the absolute values ​​of ~, it will be less than α The wavelet coefficients are assigned a value of 0, greater than α The wavelet coefficients are preserved;

[0058] 3. Wavelet reconstruction: The spectral energy signal separated by thresholding is subjected to inverse WT transform, and the noise component of the original data is obtained after signal reconstruction.

[0059] 4. Noise-reduced signal: After calculating the noise component, subtract the noise component from the original data to obtain the noise-reduced true signal.

[0060] Example 3:

[0061] This embodiment only describes the differences from Embodiment 2. Specifically, it uses a groundwater monitoring project in a certain area of ​​Sichuan as an example. This embodiment selects monthly water level monitoring records from 33 monitoring wells from January 2004 to December 2018 for 15 consecutive years as the data source, totaling 180 periods of water level data, stored in a 33×180 matrix. The data from the first 14 years (2004-2017) is selected as the training set, and the data from the 15th year (2018) is selected as the test set. That is, the first 168 periods of data in the sample dataset are used as the training set for model learning, and the last 12 periods of data are used as the test set for the model.

[0062] The GRU neural network model in this embodiment is built on the MXNet deep learning framework, and the backend program is implemented using the GluonTS toolkit. The specific project environment configuration is as follows: Processor: Intel(R) Core(TM) i5-8300H CPU @ 2.30GHz, Operating System: Windows 10 Enterprise Edition 64-bit, Programming Language: Python, Development Platform: PyCharm 2019.1.2. The main parameters are set as follows: loss function is MSE, optimizer is Adam, activation function is tanh, learning rate is 0.0001, and Dropout is 0.2, which means that the neuron dropout rate during model training is 20%.

[0063] Based on the principle of selecting contrasting models, this embodiment selects three contrasting models: a single prediction contrasting model LSTM, a commonly used nonlinear prediction model, and a typical linear regression model. The experimental results are attached. Figure 3 As shown. (From the appendix) Figure 3 It can be seen that the technical solution of this application has better monitoring effect than LSTM, SVR and ARIMA models, and can more accurately reflect the change trend of groundwater level in monitoring well over time, and has good feasibility in practical applications.

[0064] Example 4:

[0065] This embodiment only describes the parts that differ from Embodiment 1, specifically as shown in the attached document. Figure 4 As shown, for the data unit, during the acquisition process, the presence of pressure head causes significant deviations in the data acquisition equipment of the groundwater monitoring well. This is because the presence of pressurized water in the gravity flow well leads to instability of the data acquisition equipment and prominent issues related to support and installation. At the same time, the gas or vacuum state at the top of the monitoring well has a significant impact on the monitoring data, directly reducing the accuracy of the monitoring data.

[0066] Based on the above problems, targeted improvements were made to the data unit. Specifically, a sealing unit 1 was installed on the monitoring well. The sealing unit 1 includes a sealing seat 11, a sealing cover 12, and a sealing member 13. The sealing cover 12 is located at the wellhead of the sealed well. The sealing seat 11 is located above the sealing cover 12, and the bottom of the support 14 is inserted into the ground. Slide plates 15 are evenly distributed in the middle of the lower end face of the sealing seat 11. Several sliding holes 16 corresponding to the slide plates 15 are opened in the middle of the sealing cover 12. A sliding cylinder 17 is also provided on the lower end face of the sealing cover 12. The sealing member 13 is hinged to the bottom of the slide plates 15 and the sliding cylinder 17. The sealing member 13 includes a claw 18. When the sealing seat 11 moves downward, the claw 18 can abut against the inner wall of the monitoring well by moving the slide plates 15 in the sliding holes 16.

[0067] Specifically, the lower end face of the sealing cover 12 is provided with a limiting cylinder 19, and the limiting cylinder 19 is provided with a plurality of limiting holes. The sealing member 13 further includes: a first connecting rod 20 hinged to the sliding plate 15 and a second connecting rod 21 hinged to the bottom of the sliding cylinder 17. The free end of the first connecting rod 20 is hinged to the middle of the second connecting rod 21. The free end of the second connecting rod 21 moves through the limiting hole and is connected to the claw 18. The interior of the sliding cylinder 17 is provided with a monitoring cylinder 22 for accommodating the data unit.

[0068] It should be noted that the monitoring cylinder 22 has several monitoring nodes arranged from top to bottom. Each monitoring node can be opened, closed, and sealed through a sealing component. Each monitoring node is equipped with a sensor for monitoring groundwater data, and each monitoring node has an opening connected to the groundwater via a valve. When monitoring different water layers, only the corresponding monitoring node needs to be opened for groundwater monitoring.

[0069] It should also be noted that, based on the above structure, under the weight of the sealing seat 11, there is always a downward pressure on the sliding plate 15, which causes the sliding plate 15 to tend to move downward and exert a continuous force on the first connecting rod 20. This force is manifested as the pressure of the chuck 18 on the inner wall of the monitoring well, thereby enabling the sealing operation of the monitoring wellhead and avoiding the instability of the data during the monitoring process. In addition, in order to ensure the accuracy of the monitoring data, monitoring variables are introduced as the monitoring set of the measured data to ensure the stability of the measured data.

[0070] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A groundwater monitoring system based on convolutional neural networks, comprising a monitoring well, characterized in that: The monitoring well is equipped with a sealing unit, which includes a sealing seat, a sealing cover, and a sealing component. The sealing cover is located at the wellhead of the monitoring well, and the sealing seat is located above the sealing cover with the bottom of the support inserted into the ground. The lower end face of the sealing seat has evenly spaced sliding plates. The middle of the sealing cover has several sliding holes corresponding to the sliding plates, and the lower end face of the sealing cover is also provided with a sliding cylinder. The sealing component is hinged to the bottom of both the sliding plates and the sliding cylinder. The sealing component includes a claw. When the sealing seat moves downward, the movement of the sliding plates within the sliding holes enables the claw to abut against the inner wall of the monitoring well. The lower end face of the sealing cover is provided with a limiting cylinder, and the limiting cylinder is provided with a plurality of limiting holes. The sealing member also includes: a first connecting rod hinged to the sliding plate and a second connecting rod hinged to the bottom of the sliding cylinder. The free end of the first connecting rod is hinged to the middle of the second connecting rod. The free end of the second connecting rod moves through the limiting hole and is connected to the claw. The interior of the sliding cylinder is provided with a monitoring cylinder for accommodating the data unit. The monitoring cylinder is provided with several monitoring nodes from top to bottom. Each monitoring node is opened and closed and sealed by a sealing component. Each monitoring node is provided with a sensor for monitoring groundwater data and has an opening connected to the groundwater through a valve. The system also includes: The data unit is used to organize the spatiotemporal monitoring data of groundwater level and compile the dataset, and is also used to collect groundwater data in the monitoring wells; A processing unit, signal-connected to the data unit, is used to preprocess the dataset to obtain a first sequence matrix; A filtering unit, signal-connected to the processing unit, performs spatial correlation filtering on the first sequence matrix as the basic data, and constructs a second sequence matrix based on the filtering results. The monitoring unit is signal-connected to the screening unit, equipped with a convolutional neural network model, and used to divide the second sequence matrix into a training set and a test set. After being trained on the training set, the convolutional neural network model can output a monitoring set after inputting the test set. The monitoring unit compares the received groundwater data with the monitoring set. If the threshold condition is met, the groundwater data is output. If the threshold condition is not met, the groundwater data collection is repeated.

2. The groundwater monitoring system based on a convolutional neural network according to claim 1, characterized in that: The processing unit includes: The signal decomposition module is used to decompose the original signal in the dataset to obtain the scale coefficients. The threshold processing module is connected to the signal decomposition module and is used to filter and retain the scale coefficients after decomposition. The signal reconstruction module is connected to the threshold processing module and is used to reconstruct the discrete signal after threshold processing into a continuous signal, that is, to obtain a smooth real signal after denoising. The set of real signals is the first sequence matrix.

3. The groundwater monitoring system based on a convolutional neural network according to claim 1, characterized in that: The screening unit uses a distance-weighted KNN algorithm to select a set of sequences that are spatially correlated with the monitoring target and constructs a second sequence matrix.

4. A groundwater monitoring system based on a convolutional neural network according to claim 3, characterized in that: The filtering unit is set with an Euler distance threshold, the selection condition is that the distance is less than the Euler distance threshold, and the second sequence matrix is ​​arranged in ascending order based on the size of the Euler distance.

5. A groundwater monitoring system based on a convolutional neural network according to claim 1, characterized in that: The monitoring unit includes an extension module, which performs feature extraction on the second sequence matrix based on the convolutional neural network model. After extraction, dense relationships are obtained, and then extended data inference rules are obtained based on the dense relationships. The extended data inference rules include: complement rules, union rules, inverse rules, transitive rules, and augmentation rules.

6. A groundwater monitoring system based on a convolutional neural network according to claim 5, characterized in that: The monitoring unit also includes a training module that is signal-connected to the extension module. The training module inputs the training set into the convolutional neural network model and completes the training process based on the data inference rules.

7. A groundwater monitoring system based on a convolutional neural network according to claim 6, characterized in that: After training, the convolutional neural network model inputs a test set and outputs a monitoring set. The monitoring unit also includes a fusion module connected to the training module. The fusion module can fuse the monitoring set according to the weight values ​​in the second sequence matrix to obtain target monitoring data.

8. A groundwater monitoring system based on a convolutional neural network according to claim 1, characterized in that: The convolutional neural network model is a GRU neural network model, and the hidden layers of the GRU neural network model are enhanced with Dense layers, and the output layer is enhanced with a softmax layer.

Images