CN115470850A - A water quality anomaly event recognition and early warning method based on water quality spatiotemporal data of pipe network - Google Patents

Abstract

A water quality abnormal event identification early warning method based on pipe network water quality time-space data belongs to the technical field of water supply pipe network water treatment. Firstly, adjacent monitoring stations with the same change trend in the water supply network are analyzed in groups, and data are prepared and processed. And secondly, constructing an antagonistic learning network model, and finding a stable state through model training and updating. Third, an anomaly score is constructed based on the generator and the discriminator of the model, and an appropriate anomaly score threshold is selected. Fourthly, calculating the probability of the water quality abnormal event, calculating the probability of the pollution event at each moment, and smoothing. And finally, respectively utilizing the single-site model and the multi-site model to calculate the probability of the water quality abnormal accident, and fusing to obtain the combined event probability reflecting the possibility of the pollution event. According to the invention, the detection accuracy of the pollution event is improved by fusing the results of the single-site and multi-site anomaly detection models; the model is constructed and trained only by water quality data of a water supply network in normal operation, and the model is wide in application range and high in practicability.

Description

A water quality anomaly event identification and early warning method based on water quality spatiotemporal data of pipe network

technical field

The invention belongs to the technical field of water treatment of water supply pipe networks, and relates to a water quality abnormal event identification and early warning method based on pipe network water quality spatio-temporal data.

Background technique

The water supply network is an important urban infrastructure that transports water safely and reliably to users. However, especially in developing countries like China, pollution accidents often occur in the water supply network due to the aging of the network, lack of dispatch and maintenance management, and poor construction quality. When a pollution accident occurs in the water supply network, unless the pollution can be found and removed quickly, the polluted water will quickly spread to the entire water supply network, forming a global risk. This will not only cause interruption of water supply and huge economic losses, but also cause environmental damage and public health safety problems. Therefore, rapid and accurate detection of pollution incidents is crucial to ensuring the safety of urban water supply, which will help the water affairs group formulate remedial measures, reduce losses caused by pollution incidents, and improve the water supply level and social acceptance of the water affairs group.

Traditionally, detection of specific substances in drinking water has been carried out through field sampling and laboratory analysis. This method can test various types of water quality parameters or directly identify pollutants. However, for large-scale water supply networks, this method is very time-consuming and labor-intensive, and most importantly, it cannot provide early warning of pollution events in time. At present, the water affairs groups in some cities in my country have established a relatively complete sensor monitoring network (SCADA system) in the local water supply system. Using water quality online monitoring sensors can monitor the changes of conventional water quality parameters at key nodes in the pipeline network in real time. Based on this In addition, it can quickly analyze and respond to water quality accidents in the water supply network, so as to improve the company's water supply service level.

The research on the identification of water pollution accidents can be mainly divided into methods based on statistical analysis models, methods based on hydraulic models, and methods based on machine learning models. In the method based on the statistical analysis model, the detection of pollution accidents is usually based on the distribution of water quality parameter data of the water supply network. Due to the nonlinear and non-stationary nature of water quality data, statistical methods are usually not suitable for detecting small abnormal changes in water quality in water supply networks. The hydraulic model-based approach detects pollution incidents by comparing observed real-time water quality data with predicted values obtained using a network hydraulic water quality model. However, due to the complexity of the pipe network topology and data limitations, it is difficult to obtain an accurate hydraulic water quality model in practical applications. Machine learning algorithms are considered as an alternative to predict changes in water quality parameters and identify pollution incidents in real time. Various machine learning algorithms have been applied to the detection of water quality anomalies in water supply networks, such as artificial neural network (ANN), support vector machine (SVM), long short-term memory (LSTM), etc. These models can capture the water quality of a single sensor site Characteristics of Time Series Data. However, these models do not take advantage of the spatial distribution relationship of sensor data from multiple stations, which will generate more false positives when water quality monitoring stations have large hydraulic operation changes during normal operation. In general, the existing models have the problems of low model accuracy, low detection rate of pollution detection accidents, and many false positives in practical applications.

Contents of the invention

For the above-mentioned deficiencies, the problem to be solved by the present invention is to provide a method for identifying and early warning of abnormal water quality accidents in the water supply pipe network, which can perform correlation analysis on multivariate water quality data in a spatial range in complex water supply systems where the hydraulic water quality model cannot be accurately obtained , to obtain the characteristics of the spatio-temporal distribution of water quality in the water supply network, realize the real-time automation of water pollution accident monitoring, improve the detection rate of the model for pollution accidents, and reduce the number of false alarms of the model.

In order to achieve the above-mentioned purpose, the technical scheme adopted in the present invention is:

A water quality abnormal event identification and early warning method based on pipe network water quality spatiotemporal data, comprising the following steps:

(1) Select the monitored water quality parameters and water quality sensor stations, and analyze the time interval and change trend of the sensors receiving polluted water, and group and analyze the adjacent monitoring stations with the same change trend in the water supply network. Divide N sensor detection stations into N+1 groups: in the first N groups, each group contains multi-parameter water quality monitoring data of a single sensor station, and another group contains multi-parameter water quality monitoring data of all N sensor stations.

其中

是t时刻叠加图像m^t中第i行，第j列的元素，计算为t时刻所收集的V个水质参数数据中第i个和第j个数据的总和。当N＝1时，叠加图像转换仅用于单个传感器站点的多参数数据融合；当N>1时，其叠加图像转换用于融合多个站点的时空分布数据。为了减少噪声的影响，每个时刻的叠加图像取过去d个时间长度的平均值。(2) Perform standardized preprocessing on the water quality data of the N+1 group, and convert the water quality time series data in each group after preprocessing into superimposed images. Assuming that each station collects Nr water quality parameters, for each moment, V water quality parameter data can be collected from the analyzed N sensor stations (V=Nr×N). At time t, the range image of each group can be expressed as

in

is the i-th row and j-th column element in the superimposed image m ^t at time t, calculated as the sum of the i-th and j-th data among the V water quality parameter data collected at time t. When N = 1, the superimposed image transformation is only used for multi-parameter data fusion of a single sensor site; when N > 1, its superimposed image transformation is used to fuse the spatio-temporal distribution data of multiple stations. In order to reduce the influence of noise, the superimposed image at each moment takes the average value of the past d time lengths.

(3) Construct an adversarial learning network model GAN.

和基于实测数据的叠加图像m输入到判别器网络，判别器用于区分实测数据的叠加图像m(真实)和由生成器G重构的叠加图像

(伪造)。The described confrontational learning network model is divided into a generator G and a discriminator D, and the generator G adopts a convolutional neural network including an encoder for compressing image information and a decoder for restoring image information; the discriminator D is a network used to compress and extract Convolutional Neural Networks for Classification of Image Features. First, use step (2) to convert the multivariate time series data of water quality monitoring stations analyzed under normal operation into superimposed images, and then use the historical superimposed images of the past K moments as the input of the GAN model, through the generator network G Generate an overlay image of the next moment. Next, the image generated by the generator

and the superimposed image m based on the measured data are input to the discriminator network, and the discriminator is used to distinguish the superimposed image m (real) of the measured data from the superimposed image reconstructed by the generator G

(forgery).

The training of the GAN model uses the improved W- _GAN loss as the training loss LD of the discriminator, which is used to stabilize the training process, specifically expressed as,

是由生成器重构的叠加图像；m是基于实测数据的叠加图像；

是由一对真实叠加图像和重构叠加图像组成的插值图像；

是对插值图像

求梯度；λ_GP是梯度惩罚项的系数，根据经验值取10。In the formula, E represents the mathematical expectation of the calculation; D( ) is the feature vector finally output by the discriminator D;

is the superimposed image reconstructed by the generator; m is the superimposed image based on the measured data;

is an interpolated image consisting of a pair of real overlay images and reconstructed overlay images;

is the interpolated image

Find the gradient; λ _GP is the coefficient of the gradient penalty item, which is 10 according to the empirical value.

The generator training uses the reconstruction loss between the real overlay image and the generated overlay image as the generator loss L _G to help the generator learn the distribution of normal water quality data. Its calculation is expressed as,

In the GAN model, two networks, G and D, are trained and updated simultaneously. The ultimate goal of model training is not to minimize the loss of any single network, but to find a steady state when both G and D losses converge to a stable state.

与基于实测数据构造的叠加图像m之间的重构损失来计算。基于判别器网络的异常识别是通过比较将基于实测数据的叠加图像和生成器生成的图像分别输入判别器网络后得到的最终输出的特征向量D(m)和

之间特征损失。最终基于GAN模型构造的异常分数具体表示为，(4) Construct an anomaly score ψ based on the generator and discriminator of the GAN model trained in step (3). GAN-based anomaly score calculation consists of two parts: anomaly identification with generator network and anomaly identification with discriminator network. Among them, the anomaly recognition based on the generator network is by comparing the images generated by the generator network

and the reconstruction loss between the superimposed image m constructed based on the measured data. The anomaly recognition based on the discriminator network is to compare the final output feature vector D(m) and

Between feature loss. The final abnormal score based on the GAN model is specifically expressed as,

Among them, ψ(t) represents the anomaly score calculated using the GAN model at time t. The closer the anomaly score is to 0, the more normal the current water quality is. On the contrary, the larger the anomaly score, the greater the difference between the water quality at the current moment and the water quality during normal operation, and the more likely the water quality is in an abnormal state at this moment. _λs is a weighting parameter that adjusts the relative importance of reconstruction loss and feature loss with respect to the anomaly score.

(5) Select an appropriate anomaly score threshold ψ _thre , and when the anomaly score obtained in step (4) exceeds ψ _thre , it is used as an initial outlier point recognition. Since the construction and training of the GAN model in step (3) are based on water quality data under normal operating conditions, the selection of the abnormal score threshold is mainly determined based on the abnormal score distribution obtained from the training GAN network dataset. After the water quality data set used for GAN model construction and training in step (3) is used to obtain abnormal scores in step (4), arrange the abnormal score values from small to large, and select 96%-99% of the distribution according to the actual situation of the data distribution. The % quantile point is used as its abnormal score threshold ψ _thre , so that most of the points in the water quality dataset used to construct the GAN model are within the normal range.

(6) Use the time-series Bayesian principle to calculate the probability of abnormal water quality events, and calculate the probability P(t) of pollution events at each time based on the identification results of abnormal points obtained in step (5). Specifically, it can be expressed by the following formula,

where TPR is the true positive rate, calculated as the ratio of the number of time steps correctly classified as abnormal when the water supply network is polluted to the total number of polluted time steps. When there is no prior knowledge of pollution events, TPR takes the value of 0.5. FPR is the false positive rate, calculated as the ratio of the number of time steps identified as outliers by the model to the total number of time steps in normal operation when the water supply network is operating normally without pollution, which is equivalent to the moment when the anomaly score threshold is exceeded in the training set The ratio of the number to the total number of training datasets.

The probability of occurrence of pollution events at the initial moment is given as P(0). Since pollution events are rare in real life, a smaller probability value P(0)∈[10 ^-6 ,10 ^-4 ] is taken. When the calculated probability P(t) exceeds a certain threshold P _thre , the model recognizes it as a pollution event. A low probability threshold can improve the event detection rate, but may increase the number of false alarms; a high probability threshold can improve the reliability of event alarms and reduce the number of false alarms, but at the same time, fewer pollution events may be detected. The probability threshold P _thre for identifying a pollution event is set according to the trade-off between the detection rate and the false positive rate of the pollution event by decision makers, and is usually set to a probability value exceeding 70%.

(7) The hydraulic changes in the routine operation of the water supply network will lead to sudden changes in water quality parameters in a short period of time. In order to distinguish between normal water quality changes and pollution events, a simple exponential smoothing model is used to smooth the calculated probability, which can be expressed by the following formula,

P(t)=αP(t)+(1-α)P(t-1) (6)

In the formula, α is a smoothing parameter, which determines the degree of emphasis on the latest updated event probability, α∈[0.3,0.9]. The smaller α is, the slower the update of the event probability is, and the more abnormal points need to be identified to reach the alarm threshold.

(8) Calculate the probability of abnormal water quality accidents by using the single-site model and the multi-site model respectively. The single-site model refers to applying the anomaly detection method proposed in steps (3)-(7) to the first N sets of water quality data in steps (1)-(2), constructing a GAN model for each site, and combining N The maximum value of the pollution event occurrence probabilities calculated by the GAN models is taken as the event probability P _single calculated by the single-site model. The multi-site model refers to applying the anomaly detection method proposed in steps (3)-(7) to the N+1th set of water quality data in steps (1)-(2), which includes N water quality monitoring stations Based on the monitored water quality data, a GAN model is constructed to obtain the pollution event probability P _multi calculated by the final multi-site model.

(9) In order to make full use of the water quality relationship among the monitoring stations and within the monitoring stations, the event probability calculated by the single-site model and the multi-site model is integrated, and based on the multivariate water quality parameters of all monitoring stations, the probability of pollution events is obtained The combined event probability P _all can be specifically expressed by the following expression:

P _all (t)=ηP _single (t)+(1-η)P _multi (t) (7)

In the formula, P _single (t) and P _multi (t) represent the probability of occurrence of pollution events calculated by using the single-site model and the multi-site model at time t, respectively. η is the importance weight to adjust the influence of single-site model and multi-site model on the synchronization decision-making of final pollution event identification. The single-site and multi-site models in the present invention are all unsupervised models, and the pollution information is not known in advance, so setting η=0.5 reflects the same importance of the single-site model and the multi-site model to the final pollution detection and alarm. When the combined event probability P _all exceeds the preset probability threshold P _thre in step (6), the alarm signal of the final model is given, and the occurrence probability of water quality abnormal accident P _all is given.

The beneficial effects of the present invention are:

(1) The method for detecting abnormal water quality accidents proposed by the present invention takes into account the water quality data information associated with multiple sites, and improves the detection accuracy of pollution events by fusing the results of single-site and multi-site anomaly detection models.

(2) The water supply pipe network water quality abnormal accident detection method that the present invention proposes has stronger robustness, can adapt to a certain degree of noise point and unsteady water quality data situation, for the water quality index kind of detection, the water quality site number of detection The absence of specific requirements increases the scope of application of the present invention.

(3) Most of the existing pipe network water pollution accident identification methods require actual pollution data for model training or setting of related parameters, and it is difficult to collect a large number of pollution events for training and learning in real life. Compared with these methods, the method for detecting abnormal water quality accidents of the water supply pipe network proposed by the present invention is an unsupervised learning method. The construction and training of the model only need the water quality data under the normal operation of the water supply pipe network, and the application range of the model is wider and the practicability stronger.

Description of drawings

Figure 1 is a flow chart of model building;

Figure 2 is a layout diagram of a city's water supply network and sensor monitoring stations;

Figure 3 is a schematic diagram of the construction of the confrontational learning model (GAN);

Figure 4 shows the final identification and early warning of abnormal water quality events obtained by using the single-site model, multi-site model, and single-site and multi-site combined model for the test set data.

detailed description

In order to present the technical solutions and advantages of the present invention more clearly, the present invention will be described in detail below in conjunction with the accompanying drawings and examples. It should be noted that the examples are only specific illustrations of the present invention, but the embodiments of the invention are not limited thereto.

Example 1.

With reference to accompanying drawing 1, concrete implementation steps of the present invention are as follows:

S1, data preparation and processing. The model builds and trains the model by using the water quality data of normal sensor monitoring stations, and evaluates the performance of the model by using the data containing water pollution events. The water quality data of multiple sensor monitoring stations in the water supply network are simulated by using the hydraulic water quality model of the pipe network, and the selected N sensors are divided into N+1 groups. The first N groups contain the water quality parameter data of their respective single stations, and the latter group Contains water quality parameter data of N stations. The required data include water quality data in normal operation and water quality data with pollution events, which are divided into the following two parts:

S11, normal water quality data. Input time-series values of multiple water quality indicators under normal operating conditions at the water source, where the water quality indicators include but not limited to residual chlorine, pH, temperature, conductivity, turbidity, TOC (total organic carbon), etc. The water quality model collects water quality data from multiple selected monitoring stations. Divide the original normal data into two parts, 70% training set and 30% test set.

S12, water quality data containing pollution events. In order to ensure that pollution events can affect the selected water quality monitoring stations, pollution events are set up near the water source. Since there are few records of water quality abnormal events during the operation of the pipeline network, the occurrence of water quality events is largely dependent on the environment of the pipeline network. The invention refers to the method of simulating the occurrence of events in related research, by simulating the change of water quality parameters in Gaussian shape distribution, setting the duration of pollution occurrence as 10 hours, and randomly simulating the number of water quality indicators (3-6) affected when different pollution times occur, Correspondingly monitor the change trend (increase or decrease) of water quality indicators, and the magnitude of change (1.0-2.5), add pollution to the data of the test set and use the hydraulic water quality model to obtain the change of water quality parameters at each station.

S2, superimposed image conversion, data image conversion is performed on the normal and polluted water quality data collected from sensor stations 1, 2, and 3. In order to ensure the size consistency of the superimposed images of single site and multi-site, after the operation of converting the time series data into superimposed images, the final superimposed images are set to the same size (32*32) by using the image filling method, The outermost circle of the image is filled with 0. In order to reduce the influence of noise, the superimposed image at each moment takes the average value of the past 5 time lengths.

S3. Construct an adversarial learning model, and use the trained adversarial learning model to calculate abnormal scores at each moment. For the superimposed images converted from N single-site and one multi-site water quality data, construct an adversarial learning model (GAN) for training. The structure and parameter settings of each adversarial learning model are the same, and the generator adopts the self-encoding structure and convolutional neural network. The network uses 30 historical images (2.5 hours) as input, and the output is the superimposed image estimated at the current moment. The discriminator D adopts a convolutional neural network structure, takes the 30 historical images as conditional input, and simultaneously inputs the real data structure at the current moment. The overlay image of and the overlay image generated by the generator, the output is a feature vector that evaluates the authenticity of the input image. Combining the reconstruction loss of the generator and the feature loss of the discriminator, Equation (4) calculates the anomaly score of the water quality at each moment.

S4, for the abnormal scores calculated by the adversarial learning model trained in S3, it is necessary to determine the normal value threshold range of the abnormal scores calculated by different models. For single-site and multi-site models, the selection of the threshold is related to the distribution of abnormal scores calculated in the training set. The general principle is to keep the calculated abnormal scores in most training sets in the normal range.

S5, the probability update and early warning of abnormal water quality events. Time-series Bayesian principle is used to update the probability of water quality abnormal events. When the probability exceeds a certain threshold, the model recognizes it as a pollution accident and alarms. The probability threshold of the pollution event alarm is set to P _thre =80%, the pollution event probability at the initial moment is set to a relatively low value P(0)=10 ^-5 , and the value of the smoothing parameter is α=0.6. Using formulas (5) and (6) to update the probability of occurrence of water quality abnormal events for N single-site models and a multi-site model respectively.

S6, counting the alarm results of multiple models at the same time. Based on the pollution event occurrence probability obtained by the multi-site model and the single-site model, the pollution event occurrence probability P _all of the final model is calculated. When it exceeds the pollution event alarm probability threshold P _thre = 80%, the final model alarm is issued, and can be Output the pollution event probability P _all and the sensor sites that have detected the pollution and the corresponding water quality parameters.

The method of the present invention is applied to a water supply network in a certain city in China (Fig. 2). A total of 33 sensor detection sites are placed in the network, and the sensor sites 1, 2, and 3 close to pollution are selected for the construction of the model method and performance evaluation , collect water quality data records of three sites at intervals of 5 minutes for 2 months, design water quality including residual chlorine, pH, temperature, conductivity, turbidity, total organic carbon, divide the data into training set and test set, The data in the training set is the normal operation data of the pipeline network, which is used for model construction and training, and the data in the test set is the data with pollution incidents added, which is used for the performance evaluation of the model to detect pollution accidents. Figure 4 compares the final detection results of the model with the single-site model based on a single site and the multi-site model. It can be seen that the proposed model combines the advantages of the single-site and multi-site models. In the seven pollution accidents, the model Six pollution events were successfully detected, and only two short-term false positives occurred, and the results were better than those using only single-site or multi-site models. The method proposed by the invention can improve the detection accuracy and increase the detection rate of pollution events. The application of this example also proves that the method proposed by the present invention has better practicability, high effective alarm rate and low false alarm rate, and can have better application value in actual water supply pipe network.

The above-mentioned embodiment only expresses the implementation mode of the present invention, but can not therefore be interpreted as the limitation of the scope of the patent of the present invention, it should be pointed out that, for those skilled in the art, under the premise of not departing from the concept of the present invention, Several modifications and improvements can also be made, all of which belong to the protection scope of the present invention.

Claims (5)

1. A water quality abnormal event identification and early warning method based on pipe network water quality space-time data is characterized by comprising the following steps:

(1) Selecting monitored water quality parameters and water quality sensor stations, analyzing the time interval and the change trend of the polluted water received by the sensor, and analyzing adjacent monitored stations with the same change trend in the water supply network in groups; dividing N sensor detection stations into N +1 groups: in the first N groups, each group contains multi-parameter water quality monitoring data of a single sensor station, and in addition, a group contains multi-parameter water quality monitoring data of all N sensor stations;

(2) Carrying out standardized pretreatment on the water quality data of the N +1 groups, and respectively converting the water quality time sequence data in each group into superposed images after the pretreatment; assuming that each site collects Nr water quality parameters, V water quality parameter data (V = Nr × N) may be collected from the N sensor sites analyzed for each time instant; at time t, the range image of each group can be represented as

Wherein

Is a superimposed image m at time t ^t Calculating the element of the ith row and the jth column as the sum of the ith data and the jth data in the V water quality parameter data collected at the time t; when N =1, the overlay image transformation is only used for multi-parameter data fusion of a single sensor site; when N is present>1, converting the superposed images to be used for fusing space-time distribution data of a plurality of sites; in order to reduce the influence of noise, the superposed image at each moment is taken as the average value of the past d time lengths;

(3) Constructing an antagonistic learning network model GAN;

the antagonistic learning network model comprises a generator G and a discriminator D, wherein the generator G adopts a convolutional neural network comprising an encoder for compressing image information and a decoder for restoring the image information; the discriminator D is a classification convolution neural network used for compressing and extracting image features; firstly, converting multivariate time sequence data of a water quality monitoring station analyzed in a normal operation state into a superposed image by using the step (2); then, taking the historical superposed images of the past K moments as the input of the GAN model, and generating the superposed image of the next moment through a generator network G; finally, the image generated by the generator is processed

And the superimposed image m based on the measured data is input to a network of discriminators for discriminating between the superimposed image m (true) of the measured data and the superimposed image reconstructed by the generator G

(forgery);

training of the GAN model employs the improved W-GAN loss as the training loss L of the arbiter _D For stabilizing the training process, in particular as,

wherein E represents the mathematical expectation of the calculation; d (-) is a feature vector finally output by the discriminator D;

is a superimposed image reconstructed by the generator; m is a superimposed image based on the measured data;

is an interpolation image composed of a pair of real superimposed images and a reconstructed superimposed image;

is a pair interpolation image

Calculating a gradient; lambda [ alpha ] _GP Is the coefficient of the gradient penalty term, and 10 is taken according to the empirical value;

generator training employs reconstruction loss between a true overlay image and a generated overlay image as generator loss L _G To help the generator learn the distribution of normal water quality data; the calculation of which is shown as,

in the GAN model, training and updating two networks G and D simultaneously; the final goal of model training is to find a steady state when the losses of G and D both converge to a steady state;

(4) Constructing an abnormal score psi based on the generator and the discriminator of the trained GAN model in the step (3);

(5) Selecting an appropriate anomaly score threshold psi _thre Setting most points in the water quality data set for constructing the GAN model within a normal range, and when the abnormal score obtained in the step (4) exceeds psi _thre Then as an initial anomalyPoint identification;

(6) Calculating the probability of the water quality abnormal event by using a time sequence Bayes principle, calculating the probability P (t) of the pollution event at each moment based on the identification result of the abnormal point obtained in the step (5), and specifically using the following formula,

wherein TPR is a true positive rate calculated as a ratio of time steps correctly classified as abnormal when the water supply pipe network is contaminated to total time steps contaminated; in the absence of prior knowledge of a contamination event, the TPR value is 0.5; FPR is a false positive rate, calculated as the ratio of the number of time steps of the model identified as an abnormal point to the total number of time steps of normal operation under the condition of normal operation and no pollution of the water supply network, which is equivalent to the ratio of the number of times of exceeding an abnormal score threshold value in the training set to the total number of the training data sets;

giving a probability of occurrence of a pollution event P (0) at an initial time, when the calculated probability P (t) exceeds a threshold P _thre Then the model is identified as a contamination event; the probability threshold value P of pollution event identification _thre Setting according to the balance consideration of a decision maker on the detection rate and the false alarm rate of the pollution event;

(7) Normal operation hydraulic changes in a water supply network can lead to sudden changes in water quality parameters in a short period; in order to distinguish normal water quality change and pollution events, an exponential smoothing model is used for smoothing the calculated probability;

(8) Respectively utilizing the single-site model and the multi-site model to calculate the probability of the water quality abnormal accident;

the single-site model is that the anomaly detection methods proposed in the steps (3) to (7) are respectively and independently applied to the first N groups of water quality data in the steps (1) to (2), a GAN model is constructed for each site, and the maximum value of the pollution event occurrence probability calculated by the N GAN models is used as the event probability P calculated by the single-site model _single ；

The multi-site model refers to the anomaly detection proposed by the steps (3) to (7)The method is applied to the (N + 1) th group of water quality data in the steps (1) - (2), wherein the group of data comprises water quality data monitored by N water quality monitoring sites, a GAN model is constructed, and the pollution event occurrence probability P calculated by the multi-site model is finally obtained _multi ；

(9) In order to fully utilize the water quality relations among monitoring stations and in the monitoring stations, the event probabilities calculated by the single-site model and the multi-site model are fused, and the combined event probability P reflecting the possibility of the pollution event is obtained based on the multivariate water quality parameters of all the monitoring stations _all Expressed using the following expression:

P _all (t)＝ηP _single (t)+(1-η)P _multi (t) (7)

in the formula, P _single (t) and P _multi (t) respectively representing the time t, and calculating the probability of occurrence of the pollution event by using a single-site model and a multi-site model; eta is the importance weight of adjusting the influence of the single-site model and the multi-site model on the final pollution event identification synchronous decision;

the single-site model and the multi-site model are both unsupervised models, and pollution information is not known in advance, so that eta =0.5 is set to reflect the same importance of the single-site model and the multi-site model to final pollution detection alarm; when combined event probability P _all Exceeding the probability threshold P preset in the step (6) _thre Then giving out alarm signal of final model and giving out probability P of abnormal water quality accident _all 。

2. The pipe network water quality space-time data-based water quality abnormal event identification and early warning method as claimed in claim 1, wherein in the step (4), the GAN-based abnormal score calculation comprises two parts: anomaly identification using a generator network and anomaly identification using a discriminator network; wherein the generator network based anomaly identification is by comparing images generated by the generator network

And the overlay image m constructed based on the measured dataCalculating the loss; the anomaly recognition based on the discriminator network is to compare the feature vectors D (m) and D (m) of the final output obtained by inputting the superimposed image based on the measured data and the image generated by the generator into the discriminator network

Loss of features between; the anomaly score finally constructed based on the GAN model is specifically expressed as,

where ψ (t) represents an abnormality score calculated at time t using a GAN model; the closer the abnormal score is to 0, the more normal the current water quality state is, otherwise, the larger the abnormal score is, the larger the difference between the water quality state at the current moment and the water quality in normal operation is, and the more likely the water quality at the moment is in an abnormal state; lambda [ alpha ] _s Are weighting parameters that adjust the relative importance of reconstruction loss and feature loss with respect to the anomaly score.

3. The pipe network water quality space-time data-based water quality abnormal event identification and early warning method as claimed in claim 1, wherein in step (5), the abnormal score threshold psi _thre Confirmation was performed as follows: because the construction and the training of the GAN model in the step (3) are carried out based on the water quality data in the normal operation state, the selection of the abnormal score threshold value is mainly determined based on the abnormal score distribution obtained by training the data set of the GAN network; after the abnormal scores of the water quality data set used for GAN model construction and training in the step (3) are obtained in the step (4), the abnormal score values are arranged from small to large, and 96% -99% score points in the distribution are selected as the abnormal score threshold psi according to the actual situation of the data distribution _thre So that most points in the water quality data set used for constructing the GAN model are within a normal range.

4. The method of claim 1A water quality abnormal event identification early warning method based on pipe network water quality space-time data is characterized in that in the step (6), the occurrence probability of a pollution event is P (0) and belongs to [10 ] ^-6 ,10 ^-4 ](ii) a The probability threshold value P of the pollution event identification _thre A probability value of over 70% is set.

5. The method for identifying and early warning the abnormal water quality events based on the pipe network water quality space-time data is characterized in that in the step (7), the smoothing treatment is carried out by adopting the following formula:

P(t)＝αP(t)+(1-α)P(t-1) (6)

wherein, alpha is a smooth parameter, determines the attention degree of the recently updated event probability, and belongs to [0.3,0.9]; the smaller alpha, the slower its event probability updates, and the more outliers that need to be identified to reach the alarm threshold.

Images