Groundwater environment monitoring method and system based on big data

Abstract

The application relates to the field of intelligent monitoring, and specifically discloses a groundwater environment monitoring method and system based on big data, which first decomposes complex water quality time series data into three components of trend, season and residual by STL decomposition technology, so that the inherent periodicity of the groundwater system and random disturbance are effectively separated, on the basis of which, the trend and season components are innovatively used as a dynamic macro background for interpreting the true meaning of the residual component, and then combined with the normal disturbance mode of the LSTM self-encoder deep learning in a specific background, and finally a warning judgment is made by the SVM model. In this way, the periodic change of the groundwater environment can be effectively adapted, the normal water quality fluctuation caused by natural factors can be accurately identified, the false positive rate can be significantly reduced, the monitoring sensitivity to real pollution events can be improved, and finally the precise dynamic evaluation and reliable warning of the groundwater environment state can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent monitoring, and more specifically, to a groundwater environment monitoring method and system based on big data. BACKGROUND

[0002] Due to the characteristics of slow, hidden and difficult to repair of groundwater pollution, if it cannot be discovered and controlled in time, it will cause serious consequences to the sustainable use of water resources and human health. Therefore, building an intelligent monitoring system that can monitor the dynamic of groundwater environment in real time and accurately, and can timely warn potential pollution risks, has strategic significance for protecting water resources safety and effectively managing and protecting water environment.

[0003] To meet this monitoring demand, a variety of groundwater environment monitoring technologies based on time series data analysis have been developed. These technologies usually analyze the water quality data collected by sensors to identify abnormal fluctuations to determine whether a pollution event has occurred. However, the existing monitoring methods have technical limitations in dealing with complex groundwater environment data. The core problem is that they mostly define normal water quality state as a relatively stable or slowly changing baseline, without fully considering the inherent strong nonlinearity and periodicity of the groundwater system. In fact, groundwater quality will show significant periodic fluctuations with seasonal changes, rainfall recharge and other natural factors, and these fluctuations are part of the normal operation of the system. Due to the lack of effective modeling of this temporal context and periodic pattern, traditional models cannot establish a dynamic baseline that can adapt to different hydrological backgrounds. This makes it difficult for the model to distinguish between mutations caused by real pollution events and normal large fluctuations caused by natural factors (such as systematic differences between wet and dry seasons), resulting in a large number of false positives, which seriously restricts the accuracy and reliability of the monitoring system.

[0004] Therefore, an optimized groundwater environment monitoring method based on big data is expected. SUMMARY

[0005] In order to solve the above technical problems, the present application is proposed. Embodiments of the present application provide a groundwater environment monitoring method and system based on big data, which first decomposes complex water quality time series data into trend, season and residual components by STL decomposition technology, so as to effectively separate the inherent periodicity of the groundwater system and random disturbance, and on this basis, innovatively uses the trend and season components as a dynamic macro background to interpret the true meaning of the residual component, and combines the normal disturbance pattern of the LSTM autoencoder deep learning in a specific background, and finally makes a warning judgment by the SVM model. In this way, the periodic changes of the groundwater environment can be effectively adapted, the normal water quality fluctuations caused by natural factors can be accurately identified, the false positive rate can be significantly reduced, the monitoring sensitivity to real pollution events can be improved, and finally the accurate dynamic evaluation and reliable warning of the groundwater environment state can be realized.

[0006] According to one aspect of the present application, a groundwater environment monitoring method based on big data is provided, which comprises:

[0007] obtaining water quality time series data from a specific monitoring well;

[0008] inputting the water quality time series data into an STL decomposition module to obtain trend components, season components and residual components;

[0009] extracting a water quality residual feature vector sequence from the residual components with the trend components and the season components as background information;

[0010] performing abnormal score calculation on the water quality residual feature vector sequence based on an LSTM autoencoder to obtain an abnormal score value sequence;

[0011] inputting the abnormal score value sequence into a warning analysis engine based on an SVM model to obtain a warning analysis result.

[0012] According to another aspect of the present application, a groundwater environment monitoring system based on big data is provided, which comprises:

[0013] a data acquisition module for obtaining water quality time series data from a specific monitoring well;

[0014] an STL decomposition module for inputting the water quality time series data into the STL decomposition module to obtain trend components, season components and residual components;

[0015] a water quality residual feature extraction module for extracting a water quality residual feature vector sequence from the residual components with the trend components and the season components as background information;

[0016] an abnormal score calculation module for performing abnormal score calculation on the water quality residual feature vector sequence based on an LSTM autoencoder to obtain an abnormal score value sequence;

[0017] The early warning analysis module is configured to input the sequence of abnormal score values into an early warning analysis engine based on an SVM model to obtain an early warning analysis result.

[0018] Compared with the prior art, the underground water environment monitoring method and system based on big data provided by the present application firstly decomposes complex water quality time series data into three components of trend, season and residual by using STL decomposition technology, so as to effectively separate the inherent periodicity of the underground water system and random disturbance, and on this basis, innovatively takes the trend and season components as a dynamic macro background for interpreting the real meaning of the residual component, and further combines the normal disturbance mode of the LSTM self-encoder deep learning in a specific background, and finally makes a warning judgment by using the SVM model. In this way, the periodic change of the underground water environment can be effectively adapted, the normal water quality fluctuation caused by natural factors can be accurately identified, the false positive rate can be significantly reduced, the monitoring sensitivity to real pollution events can be improved, and finally the precise dynamic evaluation and reliable warning of the underground water environment state can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The drawings provided in the present application are used to provide a further understanding of embodiments of the present application and constitute a part of the specification, which serve to explain the present application together with the embodiments of the present application, but do not constitute a limitation of the present application. In the drawings, the same reference numerals generally designate the same components or steps.

[0020] Figure 1 A flowchart of the underground water environment monitoring method based on big data according to the embodiments of the present application is shown in FIG. 1.

[0021] Figure 2 A data flow schematic diagram of the underground water environment monitoring method based on big data according to the embodiments of the present application is shown in FIG. 2.

[0022] Figure 3 A block diagram of the underground water environment monitoring system based on big data according to the embodiments of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0023] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0024] As shown in the present application and claims, unless the context clearly indicates otherwise, the words "a," "an," "the," and / or "this" do not necessarily refer to the singularly but can include the plural, or vice versa. Generally, the terms "comprising," "including," and "containing" are intended to be open-ended and do not exclude additional steps, elements, or components.

[0025] While the present application makes various references to certain modules in the system according to embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are merely illustrative, and different aspects of the system and method can use different modules.

[0026] Flowcharts are used in the present application to illustrate operations performed by the system according to embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in sequence. Rather, various steps can be processed in reverse order or simultaneously, as desired. Other operations can also be added to or removed from these processes, or one or more steps can be removed from these processes.

[0027] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It is obvious that the described embodiments are only a part of the embodiments of the present application, and not all embodiments of the present application, and it should be understood that the present application is not limited by the example embodiments described herein.

[0028] In the technical solution of the present application, a groundwater environment monitoring method based on big data is proposed. Figure 1 A flowchart of the groundwater environment monitoring method based on big data according to embodiments of the present application. Figure 2 A data flow diagram of the groundwater environment monitoring method based on big data according to embodiments of the present application. As shown in Figure 1 and Figure 2 As shown in the groundwater environment monitoring method based on big data according to embodiments of the present application, the method comprises the steps of: S1, obtaining water quality time series data from a specific monitoring well; S2, inputting the water quality time series data into an STL decomposition module to obtain a trend component, a seasonal component, and a residual component; S3, extracting a water quality residual feature vector sequence from the residual component with the trend component and the seasonal component as background information; S4, performing abnormal score calculation based on an LSTM autoencoder on the water quality residual feature vector sequence to obtain an abnormal score value sequence; and S5, inputting the abnormal score value sequence into a pre-warning analysis engine based on an SVM model to obtain a pre-warning analysis result.

[0029] In particular, the S1 obtains water quality time series data from specific monitoring wells. By systematically and continuously collecting objective physicochemical indicators of the groundwater environment, it provides real, effective, and high-density raw data input for subsequent data decomposition, feature extraction, anomaly detection, and early warning analysis, providing data support for subsequent models to accurately analyze the dynamic evolution of the groundwater environment, thereby enabling timely detection and reliable early warning of potential pollution events.

[0030] wherein the specific monitoring well refers to a dedicated observation well placed at a representative geographic location or key hydraulic connection node based on hydrogeological exploration results. These monitoring wells are equipped with professional water quality sensing equipment as long-term, fixed-point groundwater quality data collection front-ends. The water quality time series data refers to a collection of a series of water quality parameter measurements continuously collected from a single or multiple specific monitoring wells at a predetermined and fixed time interval (e.g., every hour, every minute). Each data point in this data collection contains an accurate timestamp and multiple water quality parameter values such as pH, conductivity, turbidity, dissolved oxygen, and specific pollutant concentration, forming a structured data stream that reflects the dynamic changes in water quality over time.

[0031] In specific implementation, first, one or more multi-parameter water quality monitoring probes or sensor arrays are deployed in the specific monitoring well, which can measure the pre-set water quality indicators in real time. Subsequently, an embedded data collection unit automatically reads the measurement values of all sensors at a determined sampling frequency (e.g., every 30 minutes). The collected raw data is subjected to preliminary formatting and encapsulated into data frames containing the unique identifier of the monitoring well, the collection timestamp, and the key-value pairs of various water quality parameters. Then, using a wireless data transmission module (Data Transmission Unit), such as through GPRS, 4G / 5G, or NB-IoT Internet of Things communication technology, the encapsulated data frames are remotely sent to the central data server in a safe and reliable manner. After receiving the data, the central server performs analysis and verification and stores the valid data in a database dedicated to processing time series data (such as the time series database InfluxDB or HBase), thereby forming a structured water quality time series data set that can be called by subsequent analysis modules.

[0032] In particular, the S2, the water quality time series data is input into the STL decomposition module to obtain the trend component, the seasonal component and the residual component. It should be understood that the original water quality time series data is a mixed signal formed by the joint action of multiple complex factors, which contains not only slow changes (trend) driven by long-term evolution or persistent influence of the groundwater system, but also periodic fluctuations (seasonality) caused by natural laws such as seasonal change and rainfall cycle, and also random disturbances (residual) caused by sudden and accidental events. If the mixed signal is not distinguished for abnormal detection, normal seasonal fluctuations will be misjudged as pollution events, resulting in high false alarm rate of the system. Therefore, in the technical solution of the present application, the water quality time series data is input into the STL decomposition module to structurally deconstruct the mixed signal and accurately separate each component, so as to strip the real random abnormal signal that needs to be focused on for the subsequent analysis module, and lay a solid data foundation for realizing high-precision abnormal event recognition and early warning.

[0033] Among them, the STL decomposition module is a functional unit specially used for time series decomposition, and its name is the abbreviation of "Seasonal and Trend decomposition using Loess", which means using local weighted regression scatter smoothing method for seasonal and trend decomposition. It is a powerful and robust time series decomposition algorithm that can effectively process various types of time series data. The trend component output refers to the overall trend and internal law of the water quality data in a long time span, which is continuously rising, falling or remaining stable; the seasonal component is the regular fluctuation pattern that appears repeatedly in a fixed time period (for example, one year or one day); the residual component is the part left after the trend component and the seasonal component are removed from the original time series, which represents random noise, measurement error and potential abnormal signals caused by sudden events that cannot be explained by the trend and seasonal patterns.

[0034] In specific implementation, the STL decomposition module receives the water quality time series data obtained in the previous step as input and processes the sequence according to a specific mathematical model; specifically, the STL decomposition module processes the water quality time series data to obtain the trend component, the seasonal component and the residual component according to the following formula:

[0035] ,

[0036] Among them, is the trend component, is the seasonal component, The residual component is decomposed. In this process, the STL decomposition module first smoothes the water quality time series data to preliminarily estimate the trend component, and then subtracts the trend from the water quality time series to obtain a detrended sequence; then the detrended sequence is smoothed to estimate the seasonal component; then the seasonal component is subtracted from the water quality time series to obtain a seasonally adjusted sequence, and the sequence is smoothed to update the trend component. This process is iterated until the estimates of the trend component and the seasonal component converge stably, and finally the water quality time series data is decomposed into three independent components.

[0037] In particular, the S3 extracts a water quality residual feature vector sequence from the residual component with the trend component and the seasonal component as background information. It should be understood that the conventional mechanism often only retains the residual component for analysis after decomposing the original time series data, and completely discards the trend component and the seasonal component which contain the system macro state. The fundamental defect of this processing method is that it staticizes and isolates the meaning of the residual. In the real groundwater environment, a residual disturbance with the same value has a high dependence on the system background in which it is located. For example, in the flood season, a larger residual value may belong to the normal category because the flow is abundant and the hydrological conditions are volatile; however, in the dry season, a residual of the same size is most likely to be a significant abnormal signal because the hydrological activity is flat. Therefore, in the technical solution of the present application, the trend and seasonal information representing the macro state of the system are used as the dynamic background for interpreting the residual signal, that is, the data sequence formed by combining the trend component and the seasonal component together represents the predictable and regular macro state of the groundwater system at any time. Further, through a deep interaction mechanism, a residual feature is generated which can contain context information and has stronger expression ability, so that the subsequent anomaly detection model can make more accurate and more realistic judgments.

[0038] In implementation, first, the trend component and the seasonal component are spliced to obtain a background component; then, the background component and the residual component are interacted to obtain a water quality residual feature vector sequence. Here, it should be understood that the groundwater system itself is jointly affected by natural periods and long-term trends, and although the residual component has been stripped of the main trend and seasonal factors, it still may implicitly contain structural information related to the background state. If the original residual is directly used for anomaly detection, it may fail to fully consider the internal relationship between the residual and the current hydrological background (trend and season), resulting in misjudgment of normal fluctuations or missed reports of real anomalies. Therefore, in the technical solution of the present application, through manifold transmission modulation, the residual feature and the background feature are deeply fused at the structural level, so that the generated water quality residual feature vector sequence not only contains the original abnormal signal of the residual, but also encodes the deviation degree and nature of the abnormal signal relative to the current system expected state (defined by the trend and seasonal components). In this way, the system can better distinguish between normal water quality fluctuations caused by rainfall recharge and abnormal mutations caused by pollutant leakage, thereby reducing the false positive rate at the root and improving the accuracy and reliability of the early warning results, providing core technical support for building a truly intelligent groundwater environment monitoring system.

[0039] Specifically, first, local time sequence feature extraction is performed on the background component and the residual component to obtain a sequence of background component local time sequence feature encoding vectors and a sequence of residual component local time sequence feature encoding vectors. It should be understood that each component obtained by STL decomposition is a macroscopic time sequence structure, and direct interaction may ignore its internal fine-grained and dynamically changing local patterns. The trend of the groundwater system is not a straight line that remains unchanged, and its change rate may fluctuate due to long-term climate or human activities; the seasonal component is also not a perfect sine wave, and its amplitude and shape may have slight differences in different periods. The residual component is a complex sequence that contains rich short-term fluctuation information. Traditional global feature representation cannot capture the details of the evolution of these components over time, and interacting these components as smooth wholes will lose important local context information for anomaly detection, making it difficult to accurately quantify the deviation of the residual anomaly from the current background state at a specific time point. Therefore, in the technical solution of the present application, local time sequence feature extraction can be used to convert the macroscopic background and residual sequences into a series of feature encoding vector sequences rich in local context information. In this way, the subsequent manifold alignment and anomaly detection process greatly enhances the perception of complex hydrological dynamics, thereby providing a key technical guarantee for precise early warning. In a specific example of the present application, one-dimensional convolutional neural network-based local time sequence feature extraction can be used to obtain a sequence of background component local time sequence feature encoding vectors and a sequence of residual component local time sequence feature encoding vectors.

[0040] Specifically, the background component and the residual component are subjected to local temporal feature extraction to obtain a sequence of background component local temporal feature encoding vectors and a sequence of residual component local temporal feature encoding vectors according to the following formula:

[0041] ,

[0042] ,

[0043] wherein, is the background component, is the residual component, denotes one-dimensional convolution encoding, and are the i-th background component local temporal feature encoding vector and the i-th residual component local temporal feature encoding vector in the sequence of background component local temporal feature encoding vectors and the sequence of residual component local temporal feature encoding vectors respectively, is the sequence of background component local temporal feature encoding vectors, is the sequence of residual component local temporal feature encoding vectors.

[0044] Next, a manifold transmission modulation matrix of the sequence of background component local temporal feature encoding vectors with respect to the sequence of residual component local temporal feature encoding vectors is calculated. It should be understood that the background component carries the long-term evolution law and periodic pattern of the groundwater system, while the essence of the residual component is an abnormal signal in which the system deviates from the expected normal behavior under a specific background state. The traditional method of directly calculating the similarity or distance between the two often only scratches the surface and cannot capture the deep and structured dependency relationship between them. Therefore, in the technical solution of the present application, a globally optimal and structured alignment scheme is constructed for the two feature manifolds of background and residual through the optimal transport theory to obtain the manifold transmission modulation matrix. Through the matrix, the system can penetrate the manifold structure information of the background feature into the explanation of the residual anomaly in a globally optimal manner, which lays a solid foundation for the final accurate judgment of the nature of the anomaly and greatly improves the ability of the system to distinguish natural fluctuations from pollution events, and essentially reduces the false alarm rate of monitoring and early warning.

[0045] Specifically, in the technical solution of the present application, the specific process of calculating the manifold transmission modulation matrix of the sequence of background component local temporal feature encoding vectors with respect to the sequence of residual component local temporal feature encoding vectors includes:

[0046] First, a weighted cross-covariance matrix between the sequence of background component local temporal feature encoding vectors and the sequence of residual component local temporal feature encoding vectors is calculated, which is represented by the formula as follows:

[0047] ​ ,

[0048] ,

[0049] wherein, is a time weight, is a weighted cross-covariance matrix, is a transpose vector of the residual component local temporal feature encoding vector, is a background component local temporal feature encoding vector;

[0050] Further, singular value decomposition and modulation matrix extraction are performed on the weighted cross-covariance matrix to obtain a manifold transmission modulation matrix, which is expressed by the formula:

[0051] ,

[0052] ,

[0053] wherein, is the manifold transmission modulation matrix.

[0054] Further, based on the manifold transmission modulation matrix, feature alignment is performed on the sequence of background component local temporal feature encoding vectors to obtain a sequence of background component remodeled local temporal feature encoding vectors. It can be understood that although the background component describes the long-term trend and periodic law of the system, its original feature representation is not adaptively adjusted for the current specific residual anomaly pattern. Under different hydrological conditions, the same background state has different explanatory power for residual anomalies. For example, in the stable period of the dry season, the background feature indicates that the system is in a low activity state, at which time a small residual anomaly may be of great significance; while in the active period of the rainy season, the same residual anomaly may belong to the normal fluctuation range. If the background feature is not calibrated specifically, direct fusion will not accurately reflect this dynamic and context-dependent correlation, resulting in inaccurate anomaly judgment. Therefore, in the technical solution of the present application, the optimal transmission mapping relationship obtained by calculation is used to intelligently reshape the background feature sequence in a situational manner, so as to distort the manifold structure of the background feature to a form that best matches the current residual anomaly feature manifold in a globally optimal way, so that the remodeled background feature can maximize the absorption and reflection of the abnormal pattern information contained in the current residual sequence. Based on this, the system can more sensitively identify residual anomalies that still have significant structural deviation from the remodeled background feature in the subsequent fusion stage, thereby greatly improving the ability of the monitoring system to distinguish true and false anomalies in complex hydrological environments, and achieving precise early warning.

[0055] Specifically, based on the manifold transmission modulation matrix, the sequence of background component local timing feature coding vectors is feature-aligned to obtain the sequence of background component reshaped local timing feature coding vectors according to the following formula:

[0056] ,

[0057] wherein, is the transpose of the sequence of background component local timing feature coding vectors, is the sequence of background component reshaped local timing feature coding vectors.

[0058] Subsequently, each group of background component reshaped local timing feature coding vectors and residual component local timing feature coding vectors in the sequence of background component reshaped local timing feature coding vectors and the sequence of residual component local timing feature coding vectors is feature-fused to obtain the water quality residual feature vector sequence. It should be understood that the reshaped background feature has already contained the context information for the current residual anomaly pattern, but it and the residual feature are still two independent signal streams, and the system is still difficult to capture the subtle and dynamic difference pattern between the expected background state and the observed abnormal signal, and these differences are the key to distinguish real pollution events from natural fluctuations. Therefore, in the technical solution of the present application, through the fine feature fusion operation, those residual anomalies that do not conform to the expectation under a specific background environment are highlighted, thereby providing the most discriminative information for the final anomaly score. This enables the subsequent LSTM autoencoder to learn a more accurate normal pattern representation, thereby being more sensitive to anomalies that are significantly inconsistent with the background environment, greatly improving the accuracy of the system in identifying real pollution events under complex hydrological conditions, while effectively reducing false positives caused by natural fluctuations, and finally achieving the core goal of an intelligent monitoring system.

[0059] Specifically, each group of background component reshaped local timing feature coding vectors and residual component local timing feature coding vectors in the sequence of background component reshaped local timing feature coding vectors and the sequence of residual component local timing feature coding vectors is feature-fused to obtain the water quality residual feature vector sequence according to the following formula:

[0060] ,

[0061] wherein, is the residual component local timing feature coding vector, is the background component reshaped local timing feature coding vector at the corresponding position, represents and the corresponding water quality residual feature vector.

[0062] In particular, the S4 performs an abnormal score calculation based on the LSTM autoencoder for the water quality residual feature vector sequence to obtain an abnormal score value sequence. It should be understood that the water quality residual feature vector sequence generated in the previous step contains context information, but is still high-dimensional complex data, which is difficult to directly use for early warning judgment. In the technical solution of the present application, a deep learning model with strong modeling capability for time series data is introduced to convert the high-dimensional and complex feature expression into an intuitive and quantifiable scalar score, which directly reflects the degree of deviation of the current water quality state from its normal behavior pattern, providing a direct, explicit and reliable basis for subsequent early warning decision-making.

[0063] The LSTM autoencoder is an unsupervised neural network model, which is composed of two main parts: an encoder based on LSTM units and a decoder also based on LSTM units. The encoder is responsible for reading the input feature vector sequence and compressing it into a fixed-length internal representation (or latent vector) containing the core information of the entire sequence; the decoder attempts to reconstruct the original input sequence as accurately as possible from this internal representation. The core idea of the model is to train it on a large amount of normal data so that it is proficient in compression and reconstruction of normal patterns. The abnormal score measures the difference between the reconstructed sequence by the LSTM autoencoder and the water quality residual feature vector sequence. In an example, if the water quality residual feature vector sequence conforms to the normal pattern learned by the model during the training phase, the reconstruction error will be small; on the contrary, if the water quality residual feature vector sequence presents an abnormal pattern that the model has never seen before, the model will not be able to effectively reconstruct, resulting in a huge reconstruction error.

[0064] In specific implementation, first, the water quality residual feature vector sequence is input into the trained LSTM autoencoder to obtain a reconstructed water quality residual feature vector sequence. That is, in the real-time monitoring stage, the water quality residual feature vector sequence generated in the previous step is input into the trained LSTM autoencoder model as a complete input. The encoder part reads each feature vector in the sequence by time step and updates its internal state step by step, finally generating a latent vector that can summarize the information of the entire input sequence. Then, the decoder part receives this latent vector and starts to generate a new sequence by time step, i.e., the reconstructed water quality residual feature vector sequence; if the water quality residual feature vector sequence is normal, the reconstructed water quality residual feature vector sequence should be highly similar in numerical value to the original input sequence.

[0065] Further, the reconstruction error between each corresponding water quality residual feature vector and reconstructed water quality residual feature vector in the water quality residual feature vector sequence and the reconstructed water quality residual feature vector sequence is calculated to obtain the anomaly score value sequence. In this process, the system compares the original sequence and the reconstructed sequence one by one at each time point. The mean square error between each corresponding water quality residual feature vector and reconstructed water quality residual feature vector is calculated as the reconstruction error. Specifically, assuming that at time point , the original water quality residual feature vector is , the reconstructed vector is , and the dimension of the vector is , the calculation formula of the anomaly score at this time point can be expressed as:

[0066] ,

[0067] wherein, is the anomaly score value corresponding to and .

[0068] In particular, S5, the anomaly score value sequence is input into the early warning analysis engine based on the SVM model to obtain the early warning analysis result. It should be understood that simply relying on a static anomaly score threshold (for example, a score greater than a certain constant is an alarm) is often too rigid and cannot adapt to the complex dynamic characteristics of abnormal events in the development process, which is prone to false alarms for short-lived and harmless noise spikes, or slow response to slowly developing but persistent potential threats. Therefore, in the technical solution of the present application, a professional classification model trained by supervised learning is introduced to learn and distinguish the subtle differences between score fluctuations in normal state and score patterns caused by real abnormal events, so as to provide a more reliable, accurate and business logic final early warning judgment, ensuring the timeliness and accuracy of the early warning. It is worth mentioning that SVM is a widely recognized and superior supervised learning classifier in the field of machine learning, which is good at finding an optimal decision hyperplane in high-dimensional space that can separate different class data points with the largest margin. In the present application, the SVM model is configured as a binary or multi-classification model for determining the current water quality state. The obtained early warning analysis result is a discrete and qualitative state label, rather than a continuous value. For example, the result can be "normal", "level one early warning", "level two early warning", etc. explicit levels, directly corresponding to different emergency response levels.

[0069] In particular implementation, this step contains two core stages: offline training stage of the model and online inference stage. In the offline training stage, a high-quality, labeled training dataset is prepared for the SVM model. The construction process of this dataset is as follows: first, a long enough historical abnormal score value sequence containing multiple known normal operating conditions and at least one or more types of real abnormal events (such as historical pollution accidents) is collected. Then, the sequence is labeled, and an explicit label such as "normal" or "abnormal" is given to each time point or time period. Next, in order to enable the SVM model to capture the dynamic trend of the abnormal score rather than just the instantaneous value, the system will use the sliding window technique to perform feature engineering on the labeled abnormal score value sequence. For example, for each time point t, take the abnormal scores of the N-1 time points before it to form a window, and extract a feature vector from the window. This vector can include the current abnormal score, the average score in the window, the maximum score, the score variance, the growth slope of the score sequence, and other statistical quantities. Finally, these feature vectors and their corresponding "normal" / "abnormal" labels are input into the training algorithm of the SVM model. The training algorithm will automatically learn an optimal decision hyperplane that can best distinguish normal samples from abnormal samples in the feature space;

[0070] In the online inference (i.e. real-time monitoring) stage, the execution process of this step is as follows: first, the pre-warning analysis engine receives the abnormal score values calculated by the LSTM auto-encoder module in real time and one by one. Whenever a new abnormal score value arrives, the engine will immediately use the same sliding window and feature engineering method as in the training stage to construct the feature vector of the current time. Then, the feature vector is input into the already trained SVM model. The SVM model will judge which side of its internal decision hyperplane the feature vector is on. Finally, the model outputs a pre-defined class label, i.e. the pre-warning analysis result, according to the judgment result. In response to the output label being "abnormal" or a certain pre-warning level, the system will trigger the corresponding alarm mechanism, such as sending an alarm message to the management personnel, highlighting on the monitoring large screen, or starting the emergency plan.

[0071] In summary, the groundwater environment monitoring method based on big data according to the embodiments of the present application is illustrated, which firstly decomposes complex water quality time series data into trend, season and residual components by STL decomposition technology, so as to effectively separate the inherent periodicity of the groundwater system and random disturbance, on this basis, innovatively uses the trend and season components as a dynamic macro background to interpret the true meaning of the residual component, and then combines the normal disturbance mode of the LSTM autoencoder deep learning in a specific background, and finally makes a warning judgment by the SVM model. In this way, it can effectively adapt to the periodic changes of the groundwater environment, accurately identify the normal water quality fluctuations caused by natural factors, thereby significantly reducing the false positive rate and improving the monitoring sensitivity of real pollution events, and finally achieving accurate dynamic evaluation and reliable warning of the groundwater environment state.

[0072] Further, a groundwater environment monitoring system based on big data is also provided.

[0073] Figure 3 A block diagram of the groundwater environment monitoring system based on big data according to the embodiments of the present application is shown. As shown in Figure 3 The groundwater environment monitoring system based on big data 300 according to the embodiments of the present application includes: a data acquisition module 310 for acquiring water quality time series data from a specific monitoring well; an STL decomposition module 320 for inputting the water quality time series data into the STL decomposition module to obtain trend, season and residual components; a water quality residual feature extraction module 330 for extracting a water quality residual feature vector sequence from the residual component with the trend component and the season component as background information; an anomaly score calculation module 340 for performing anomaly score calculation on the water quality residual feature vector sequence based on the LSTM autoencoder to obtain an anomaly score value sequence; and a warning analysis module 350 for inputting the anomaly score value sequence into a warning analysis engine based on the SVM model to obtain a warning analysis result.

[0074] As described above, the groundwater environment monitoring system based on big data 300 according to the embodiments of the present application can be implemented in various wireless terminals, such as a server with a groundwater environment monitoring algorithm based on big data, etc. In one possible implementation, the groundwater environment monitoring system based on big data 300 according to the embodiments of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the groundwater environment monitoring system based on big data 300 can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the groundwater environment monitoring system based on big data 300 can also be one of the many hardware modules of the wireless terminal.

[0075] Alternatively, in another example, the big data based groundwater environment monitoring system 300 and the wireless terminal can also be separate devices, and the big data based groundwater environment monitoring system 300 can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in a conventional data format.

[0076] Embodiments of the present disclosure have been described above, the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles, practical application, or improvement of technology in the market of the embodiments, or to enable other ordinary skilled in the art to understand the embodiments disclosed herein.

Claims

1. A groundwater environment monitoring method based on big data, characterized in that, include: Acquire time-series water quality data from specific monitoring wells; Input the water quality time series data into the STL decomposition module to obtain the trend component, seasonal component and residual component; Using trend and seasonal components as background information, a water quality residual feature vector sequence is extracted from the residual component, including: concatenating the trend and seasonal components to obtain the background component; and performing feature implicit interaction on the background component and the residual component to obtain the water quality residual feature vector sequence. An anomaly score sequence is obtained by performing an anomaly score calculation on the water quality residual feature vector sequence using an LSTM autoencoder. The abnormal score sequence is input into the early warning analysis engine based on the SVM model to obtain the early warning analysis results; The process of performing feature implicit interaction on the background component and the residual component to obtain a water quality residual feature vector sequence includes: extracting local temporal features from the background component and the residual component to obtain a sequence of local temporal feature encoded vectors for the background component and a sequence of local temporal feature encoded vectors for the residual component; calculating the manifold transmission modulation matrix of the sequence of local temporal feature encoded vectors for the background component relative to the sequence of local temporal feature encoded vectors for the residual component; performing feature alignment on the sequence of local temporal feature encoded vectors for the background component based on the manifold transmission modulation matrix to obtain a sequence of reshaped local temporal feature encoded vectors for the background component; and performing feature fusion on each set of reshaped local temporal feature encoded vectors for the background component and the sequence of local temporal feature encoded vectors for the residual component to obtain the water quality residual feature vector sequence.

2. The groundwater environment monitoring method based on big data according to claim 1, characterized in that, Water quality time series data is input into the STL decomposition module to obtain trend components, seasonal components, and residual components. This includes: the STL decomposition module processes the water quality time series data using the following formula to obtain trend components, seasonal components, and residual components: in, For trend components, For seasonal portions, These are the residual components.

3. The groundwater environment monitoring method based on big data according to claim 1, characterized in that, An anomaly score sequence is obtained by performing an LSTM-based autoencoder on the water quality residual feature vector sequence, including: The water quality residual feature vector sequence is input into the trained LSTM autoencoder to obtain the reconstructed water quality residual feature vector sequence. The reconstruction error between the corresponding water quality residual feature vectors and the reconstructed water quality residual feature vectors in each group of the water quality residual feature vector sequence and the reconstructed water quality residual feature vector sequence is calculated to obtain the anomaly score sequence.

4. The groundwater environment monitoring method based on big data according to claim 3, characterized in that, Calculating the reconstruction error between the water quality residual feature vector sequence and the reconstructed water quality residual feature vector sequence for each group includes: calculating the mean square error between the water quality residual feature vector and the reconstructed water quality residual feature vector for each group as the reconstruction error.

5. A groundwater environment monitoring system based on big data, characterized in that, include: The data acquisition module is used to acquire water quality time series data from specific monitoring wells; The STL decomposition module is used to input water quality time series data into the STL decomposition module to obtain trend components, seasonal components and residual components. The water quality residual feature extraction module is used to extract a sequence of water quality residual feature vectors from the residual components, using trend components and seasonal components as background information. This includes: concatenating the trend components and seasonal components to obtain the background components; and performing feature implicit interaction on the background components and residual components to obtain a sequence of water quality residual feature vectors. The anomaly score calculation module is used to calculate the anomaly score of the water quality residual feature vector sequence based on LSTM autoencoder to obtain the anomaly score numerical sequence. The early warning analysis module is used to input the abnormal score sequence into the early warning analysis engine based on the SVM model to obtain the early warning analysis results; The process of performing feature implicit interaction on the background component and the residual component to obtain a water quality residual feature vector sequence includes: extracting local temporal features from the background component and the residual component to obtain a sequence of local temporal feature encoded vectors for the background component and a sequence of local temporal feature encoded vectors for the residual component; calculating the manifold transmission modulation matrix of the sequence of local temporal feature encoded vectors for the background component relative to the sequence of local temporal feature encoded vectors for the residual component; performing feature alignment on the sequence of local temporal feature encoded vectors for the background component based on the manifold transmission modulation matrix to obtain a sequence of reshaped local temporal feature encoded vectors for the background component; and performing feature fusion on each set of reshaped local temporal feature encoded vectors for the background component and the sequence of local temporal feature encoded vectors for the residual component to obtain the water quality residual feature vector sequence.

Images