CN109359698A - A leak identification method based on long short-term memory neural network model - Google Patents

Abstract

The present disclosure provides a method for identifying leakage in a pipeline network based on a long-short-term memory neural network model, comprising the following steps: S1, acquiring DMA entry data; S2, cleaning the acquired DMA entry data, and constructing a multi-scale time Data set; S3, establishing a long-short-term memory neural network model; S4, identifying abnormal traffic points based on the constructed multi-scale time data set and the established long-short-term memory neural network model; S5, according to the identified abnormal traffic points Identify leaks in the pipe network. The disclosed method for identifying leakage of a pipeline network reduces the accident false alarm rate and increases the accuracy of identifying leakage.

Description

A leak identification method based on long short-term memory neural network model

technical field

The present disclosure relates to the field of neural networks, in particular to a leakage identification method based on a long-short-term memory neural network model.

Background technique

In recent years, the water supply pipeline network leakage accidents with sudden and high uncertainty have occurred frequently, which have received widespread attention from the society. Leakage accidents will not only cause waste of water and energy, economic losses, but also cause water pollution of the pipeline network in severe cases. , endangering public health. However, because the water supply pipe network is deeply buried in the ground, the system is huge, and there are many environmental interference factors, the status of the pipe network is not easy to monitor, which greatly increases the difficulty of leak detection of the pipe network.

In recent years, the global water supply industry has vigorously promoted and practiced the District Metering Area (DMA), which is used to effectively control the leakage of the pipe network. This method classifies the huge water supply pipe network and installs boundary valves to decompose it into A separate water intake area with clear boundaries. The commonly used DMA leakage monitoring mainly obtains the flow change in this area by monitoring the change of the minimum flow rate of DMA at night, but this method has problems such as low detection efficiency and long detection time. There is a time difference in the occurrence process of leakage accidents, that is, the length of time between the occurrence of leakage in the pipe network and the discovery of leakage accidents by the water supply company. The longer the time difference, the greater the amount of water leakage. Therefore, how to find leakage as soon as possible, reduce the time difference, and reduce the amount of water leakage is the key to the leakage control of the pipeline network.

With the attention and application of SCADA systems in various water supply companies, more and more water supply companies have begun to pay attention to collecting the huge monitoring data of the water supply pipe network. In this context, compared with the traditional leak detection hardware method, it has higher efficiency and lower cost The leakage identification method based on data mining has become a current research hotspot. The data mining method mainly uses the least squares method, polynomial regression, ARIMA model, support vector regression, BP neural network and other methods to build a prediction model, and then uses a fixed threshold to compare the difference between the predicted value and the measured value to identify the leakage accident of the pipeline network.

The traditional model can successfully identify some leakage accidents with large flow, but in the context of the increasing frequency of online monitoring of the pipeline network, more intensive data collection, and uneven data quality, it cannot process a large amount of flow data quickly and accurately. , Low fault tolerance. Application No. 201410507513.1 proposes a water pipe leakage detection method based on wavelet singularity analysis and ARMA model, and judges whether there is a slow leakage phenomenon of water pipe leakage through two-step prediction data analysis, which successfully solves the problem of slow leakage judgment, but the two-stage prediction of this method is too It is cumbersome, and the single-point alarm cannot continuously judge the leakage of the pipe network, and there is a problem of high false alarm rate. Application No. 201710998436.8 The predicted value of water volume is obtained through the neural network model prediction of the gated circulation unit, and the threshold value is compared with the measured value of water volume by the cosine angle method, so as to realize the leakage identification of the pipeline network; however, these methods do not consider the correction of the feedback of water volume With input, it is impossible to continuously identify pipe network bursts in real time, and a single threshold alarm will cause a large misjudgment, causing unnecessary loss of manpower and material resources of the water supply company, and the method has low reliability in actual operation. Considering that the leakage accident of the water supply network is a continuous process, it is urgent to provide a continuous leakage identification method with strong learning ability, high fault tolerance ability and accurate identification accuracy.

SUMMARY OF THE INVENTION

(1) Technical problems to be solved

In view of the above problems, the main purpose of the present disclosure is to provide a leakage identification method based on a long short-term memory neural network model, so as to solve at least one of the above problems.

(2) Technical solutions

In order to achieve the above object, as an aspect of the present disclosure, a method for identifying leakage in a pipeline network based on a long-short-term memory neural network model is provided, comprising the following steps:

S1, obtain DMA entry data;

S2, cleaning the acquired DMA entry data, and constructing a multi-scale time data set;

S3, establish a long-short-term memory neural network model;

S4, identifying abnormal traffic points based on the constructed multi-scale time data set and the constructed long-short-term memory neural network model;

S5, according to the identified abnormal flow points, identify the leakage of the pipeline network.

In some embodiments, in the step S1, when the DMA is a single entry, the flow data of the DMA single entry is obtained; when the DMA is multiple entries, the total flow of the DMA multiple entries is obtained; the flow data sampling interval is Minutes/time, the number of samples throughout the day is 24k times, that is, a total of 24k historical traffic data are obtained every day.

In some embodiments, the step S2 includes the following sub-steps:

S21, cleaning the acquired DMA entry data, and using the cleaned data to construct a continuous time series;

S22, dividing the continuous time series to construct a date series;

S23, constructing a multi-scale time data set based on the time series and the date series, where the multi-scale time data set is the input data of the long-short-term memory neural network model.

In some embodiments, the step of cleaning the acquired DMA entry data includes checking DMA entry data consistency and filling in missing values.

In some embodiments, the step S3 includes:

The constructed multi-scale time data set is used as the model input; wherein, the multi-scale time data set includes a training set and a verification set;

The predicted traffic at the next moment of the DMA entry as the model output; and

Select tanh, ReLU or Linear as the activation function of each network layer of the model, thereby constructing a long-term memory neural network model.

In some embodiments, in the step of using the constructed multi-scale temporal dataset as an input to the model,

Select the adjacent date sequence of the traffic at time t to be predicted, that is, the date sequence data at time t-1, time t+1, and time t as the model input, and set it as the first dimension time period, the second dimension time period, and the third dimension time. paragraph; and

Select the adjacent time series data of the traffic flow at time t to be predicted as the model input, that is, the first c times at time t, set as the fourth dimension time period.

In some embodiments, the step S4 includes:

S41, calculate the residual, and calculate the predicted flow output by the long-short-term memory neural network model in the training set data The residual error R _t with the measured value X _t is calculated as:

S42 , dividing the residuals, dividing the residual R _t according to time points, the same sampling time points constitute the same group of data, and a total of 24k groups of residual sequences are obtained.

S43: Calculate the residual error thresholds after the segmentation, respectively calculate 3σ intervals for the 24k groups of residuals, and the upper and lower bounds of the 3σ intervals constitute the 24k groups of upper and lower residual error thresholds corresponding to the 24k time points, and the mth day at the nth time The formula for calculating the 3σ threshold interval of the residual r _mn is:

In the formula, is the residual mean value corresponding to the n time; Var[R _n ] is the residual standard deviation corresponding to the n time;

S44, identify the abnormal flow point, calculate the flow residual at time t of the verification set, if the residual value at time t is greater than the upper residual threshold corresponding to the time t, determine the time point as an abnormal flow point, and record it as a high flow point; If the residual value at time t is less than the lower residual threshold corresponding to this time, it is determined that this time point is an abnormal flow point, and is recorded as a low flow point.

In some embodiments, the step S5 includes:

S51, after correcting the abnormal flow points, input the long-short-term memory neural network model for calculation;

S52 , according to the calculation result obtained by inputting the long-short-term memory neural network model after correction, count the number of continuous high-flow points, and identify the leakage of the pipe network by the number of continuous high-flow points.

In some embodiments, in the step S51, when the nth time of the mth day is determined to be an abnormal flow point, the corrected value of the flow data at this time in, is the traffic forecast value output by the long-short-term memory neural network corresponding to the nth time on the mth day; is the residual correction value at this moment, and is the average value of R _n in the residual sequence of the training set,

In some embodiments, in the step S52, the number q of continuous high-flow points is counted, when q is greater than or equal to the threshold Q, an early warning is started, and the leakage accident duration T _burst is calculated; the leakage accident The detection time is calculated as follows:

The formula for calculating the duration of the leakage accident is as follows:

(3) Beneficial effects

It can be seen from the above technical solutions that a leakage identification method based on a long-short-term memory neural network model of the present disclosure has at least one of the following beneficial effects:

(1) Different from the traditional machine learning method of leakage identification, the present disclosure adopts the deep learning method, and fuses the date series and time series data of the predicted traffic data through the LSTM model, and the predicted traffic is closer to the non-fault real traffic than the traditional algorithm. , and the long and short-term memory neural network has strong fault tolerance. In the case of abnormality in the input of the model training set, through the forget gate in LSTM, the abnormal data can be filtered out within a certain range, and the output is again guaranteed to be closer to non-faulty. The significance of training is not to predict real-time flow data, but to learn the trend of water flow changes under real non-fault conditions at the predicted moment, providing a prediction basis for subsequent leakage identification.

(2) Different from the traditional single-threshold leakage identification method, the present disclosure proposes a multi-threshold identification method based on time-based residuals. Divide the residual sequence to obtain 24k residual subsequences corresponding to the 24k sampling moments, calculate each group of residual subsequences, and take the upper bound of each set of 3σ intervals as the leakage losses corresponding to the 24k sampling moments. The identification threshold constitutes the refined threshold for different sampling moments. The multi-threshold leakage identification method based on time-based residuals refines the residual judgment, reduces the impact of changes in user water consumption on detection, greatly reduces the false alarm rate of the pipe network, and avoids false alarms caused by false alarms. The loss of manpower increases the trust of the model.

(3) Different from the traditional single abnormal point non-correction alarm identification method, the present disclosure proposes a feedback correction-based multi-point alarm method. By comparing the data characteristics of the true value and the abnormal value, an accurate replacement method is proposed to replace the abnormal value with the sum of the predicted value and the time-based residual deviation. The resulting prediction deviation at the next moment ensures the model's ability to identify continuous outliers. Compared with single-point abnormal judgment, the multi-point identification method greatly reduces the noise misjudgment caused by instrument failure, transmission error, etc., which not only reduces the accident false alarm rate, but also increases the accuracy of leakage identification, effectively identifying leakage accidents and damages. Its duration helps the water company to make intelligent decisions.

Description of drawings

FIG. 1 is a topological structure diagram of the present disclosure based on a long-short-term memory neural network.

FIG. 2 is a structural diagram of the disclosed multi-threshold leakage detection method based on timed residuals.

FIG. 3 is a topological structure diagram of the present disclosure based on a long-short-term memory neural network.

FIG. 4 is a flowchart of the leakage identification method of the present disclosure.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure will be further described in detail below with reference to the specific embodiments and the accompanying drawings.

In order to protect water resources, reduce the leakage rate of the pipeline network, reduce the economic losses and water quality safety risks caused by leakage accidents, and ensure the safe and reliable operation of water supply, it is necessary to use the existing technology to deeply excavate the hydraulic characteristic data that can be monitored in the water supply pipeline network, and establish a real-time monitoring system. , A model with strong fault tolerance, high precision, and low false positives to identify the leakage of the pipe network. In view of the deficiencies of the prior art, the present disclosure aims to provide a long-short-term memory-based time recursive cyclic neural network model, a training method and application thereof, and a multi-threshold-based feedback correction model for the continuous leakage accident of the water supply pipe network Identification methods and applications. By constructing a long-term memory neural network prediction model and a multi-threshold feedback correction and recognition model, a large amount of short-term data is processed in real time, and the characteristic relationship of the pipe network data is mined. The duration of the loss accident is realized, and the rapid alarm and continuous identification of the leakage accident can be realized. The present disclosure can perform high-precision diagnosis and early warning of water supply pipe network leakage, greatly shorten the response time of leakage accidents, improve the reliability of accident discrimination, and effectively assist water supply companies in diagnosing and early warning of pipe network leakage.

The present disclosure provides a long-short-term memory neural network-based model (LSTM), an input and a water supply pipe network leakage identification method, which is a multi-threshold water supply pipe network leakage accident identification method, which can be corrected based on feedback. Point alarm to comprehensively identify leakage accidents of water supply network, including the following steps:

S1, obtain DMA (District Metering Area, independent metering area) entry data;

S2, cleaning the acquired DMA entry data, and constructing a multi-scale time data set;

S3, establish a long and short-term memory lining network model;

S4, abnormal flow point identification;

S5, identify the leakage of the pipeline network.

The steps of the leakage identification method of the long-short-term memory neural network model of the present disclosure will be described in detail below with reference to FIGS. 1-4 .

S1, get DMA entry data:

Specifically, when the DMA cell is a single entry, obtain the DMA entry flow data; when the DMA cell is a multi-entry, obtain the DMA entry flow sum;

Among them, the flow data sampling interval is Minutes/time, the number of samples throughout the day is 24k times, that is, a total of 24k historical traffic data are obtained every day. Of course, the sampling interval and sampling number of the present disclosure are not limited to this.

In addition, the DMA entry data can also be user water quantity or other data.

S2, cleaning the acquired DMA entry data, and constructing a multi-scale time data set, which includes the following sub-steps:

S21. Clean the acquired DMA entry data, and use the cleaned data to construct a continuous time series.

Specifically, data cleaning includes filling in missing data values, and the method is interpolation. Of course, the data cleaning process and method of the present disclosure is not limited to this.

S22, dividing the continuous time series to construct a date series;

Obtain the historical flow data of the pipeline network of the data to be measured, and reconstruct the historical flow time series in a split manner. If the daily sampling frequency of the historical flow data is k times/hour, the number of sampling per day is 24k times, and a total of a day is sampled. The segmentation and reconstruction expression is:

f ₁₁ indicates the flow data collected for the first time on the first day, f 1, _24k indicates the flow data collected at the 24k time on the first day, f _a1 indicates the flow data collected for the first time on the a day, f _{a, 24k} represents the traffic data collected at the 24kth time on day a. Similarly, in Figure 1, f _{d-7, t-1} represent data corresponding to time t-1 7 days ago, and f _{d-6, t} represent data corresponding to time t 6 days ago.

The above is the reconstruction of the historical sequence with the date as the matrix row and the sampling time point as the matrix column, and the historical traffic sequence is reconstructed into a historical traffic matrix.

Among them, the longitudinal direction of the matrix is a sequence of dates in which the dates are arranged in order to represent periodic changes.

S23, a multi-scale time data set is formed based on the time series and the date series, and the data set is the input data of the long-short-term memory neural network model.

S3, establishing a long-short-term memory neural network model, which includes the following sub-steps:

S31, normalize the input data of the model. The formula is:

In the formula, y represents the normalized data, x represents the input original data, and x _max and x _min represent the maximum and minimum values of the input traffic data, respectively;

S32 , extracting model input features, and extracting input features of four time dimensions including a date sequence of three dimensions and a time sequence of one dimension. Taking into account the long and short-term memory characteristics of LSTM, the feature data that is closer to the feature at the moment to be predicted is input in turn.

Wherein, steps S2, S31 and S32 correspond to the feature extraction steps in FIG. 4 .

The front three-dimensional feature data of the model is a date sequence close to time t to be predicted, specifically including,

Select the adjacent date sequence of the traffic at time t to be predicted, that is, the date sequence data at time t-1, time t+1, and time t as the model input, and set it as the first dimension time period, the second dimension time period, and the third dimension time. segment, expressed as:

The first dimension time period t-1-24k*b, t-1-24k*(b-1)... t-1-24k time data set,

The data set at the second dimension time period t+1-24k*b, t+1-24k*(b-1)...t+1-24k,

The data set of the third dimension time period t-24k*b, t-24k*(b-1)...t-24k.

Among them, the data sets at the first time period t-1-24k*b, t-1-24k*(b-1)...t-1-24k represent the previous b days, the previous b- 1 day, ... to the data at time t-1 of the previous day; the second time period t+1-24k*b, t+1-24k*(b-1)...t+1-24k The data set of the time, respectively represents the data from the previous b days, the previous b-1 days, ... to the previous 1 day time t+1; the third time period t-24k*b, t-24k*(b-1). .....t-24k time data set, representing the data from the previous b days, the previous b-1 days, ... to the previous 1 day time t.

Preferably, b is an integer from 3 to 28, preferably 7.

The feature data of the fourth dimension of the model is the time series near the time t to be predicted, which specifically includes selecting the time series data of the traffic near the time t to be predicted as the model input, that is, the first c moments of the time t are set as the fourth dimension. Time period, denoted t-c, t-(c-1)...t-1, where c is an integer of 2-24, preferably 5-7. b=c.

S33, please refer to FIG. 1, the long-short-term memory neural network model includes a long-short-term memory unit layer and a plurality of fully connected neural unit layers, and each unit layer is connected in series to form a multi-level input transmission layer;

Each cycle of the long and short-term memory cell layer includes a plurality of long and short-term memory cells. The data transfer relationship between cells is that the data inputted by each cell passes through the input gate, forgetting gate, and output gate in the cell, and the output is the long-term memory and short-term memory of the next moment, entering the next cell, and looping in turn. The final output goes to the next neural unit layer.

The fully connected neural unit layer is a basic unit layer of the artificial neural network;

The model input is to build a dataset based on time series and date series traffic data;

The model output is the predicted flow of the DMA entry at the next moment;

The number of fully connected neural network layers is an integer greater than or equal to 2;

The activation function of each network layer selects tanh, ReLU or Linear.

Divide the acquired DMA traffic data (cleaned data) into a training set and a validation set in proportion, and input them into the long-term memory neural network model in turn, and output the traffic at the next moment (the next moment after the moment corresponding to the input data) data.

S4. Identification of abnormal traffic points:

The abnormal flow points include high flow points and low flow points;

S41, calculate the residual, and calculate the predicted flow output by the long-short-term memory neural network model in the training set data The residual R _t (t=(m-1)*24k+n) with the measured value X _t , the calculation formula is:

S42: Divide the residuals, divide the above-mentioned residual sequence R _t according to time points, the same sampling time points constitute the same group of data, and obtain a total of 24k groups of residual sequences, that is, time-based residual sequences R ₁ to R _24k .

S43, calculate the residual threshold after segmentation. The 3σ thresholds are calculated for the above 24k groups of residuals respectively. The calculation formula of the 3σ threshold interval of the residual r _mn at the nth time on the mth day is as follows:

In formula (4), R _n is the data of the nth column of the residual matrix R, is the residual mean value corresponding to the nth time; Var[ _Rn ] is the residual standard deviation corresponding to the nth time.

The upper and lower bounds of the 3σ interval constitute 24k groups of upper and lower residual thresholds corresponding to 24k time points

S44 , identifying abnormal flow points, and calculating the flow residual at time t of the verification set. The corresponding segmentation time at time t is the nth time on the mth day. When the residual value at time t is greater than the upper residual threshold corresponding to this time , determine the time point as an abnormal flow point and mark it as a high flow point. When the residual value at time t is less than the lower residual threshold corresponding to this time , determine the time point as an abnormal flow point and mark it as a low flow point, please refer to Figure 2.

S5, identify the leakage of the pipe network, please refer to Figure 3:

S51, correction of abnormal flow points

When a residual is marked as an abnormal flow point, the abnormal flow at the point needs to be replaced. The replaced value is used as input to enter the model again, and the flow calculation at the next moment is performed. When the nth time of the mth day is determined to be an abnormal flow point, the corrected value of the flow data at this time in, is the traffic forecast value output by the long-short-term memory neural network corresponding to the nth time on the mth day; is the residual correction value at the nth time on the mth day, and is the average value of _Rn in the training set residual sequence,

S52 , the accident identification and alarm step, count the number q of continuous high flow points, when q is greater than or equal to the alarm threshold Q (the leakage of the pipeline network occurs), start an early warning, and calculate the leakage accident duration T _burst . The quantity threshold Q is preferably 2-3.

The leakage accident detection time T _detect is calculated as formula (7)

The leakage accident duration T _burst is calculated as formula (8)

In the following embodiments, Python 3.6 software is used as the model development platform, Numpy and Pandas libraries are used to read, store, and analyze data, Matplotlib library is used to visualize data, and Keras library is used to build neural network models. Improved development efficiency.

Taking a DMA water supply pipe network in SX City in my country as an example, the specific steps of a leakage identification method based on long-short-term memory neural network model (LSTM) and multi-threshold feedback correction are introduced in detail:

1. Get DMA entry data:

Obtain the inlet flow data of a DMA water supply network in SX City, the data date is from July 10, 2015 to November 21, 2015, a total of 135 days. The sensor sampling interval is to record the instantaneous flow data every 5min (12 times/h), that is, 288 times per day. When the DMA cell is a single entry, the flow data of the DMA entry is obtained; when the DMA cell is multi-entry, the total flow of the DMA entry is obtained. This example is a multi-entry DMA with 3 entries, so the sum of the DMA entry traffic is calculated as the historical traffic data.

2. Data cleaning and construction of multi-scale temporal datasets:

1) Preprocess DMA entry data. In this example, the sampling frequency of the instantaneous flow data of the DMA entrance obtained is 12 times/hour, then the sampling times per day is 288 times, and a total of 135 days are collected, and the total amount of data is 38880. Clean the above data, check data consistency, fill Missing value. Among them, the interpolation method is selected as the method of filling missing values. The cleaned historical flow data are arranged in chronological order to form a time series {f ₁ , f ₂ , ..., f ₃₈₈₈₀ }, which represents the trend change characteristics of flow.

2) Divide the continuous time series to construct a date series. The data after data cleaning is reconstructed in a split way, and the time series is reconstructed with the date as the matrix row and the sampling time point as the matrix column, and the time series is reconstructed into a historical traffic matrix.

Then the longitudinal direction of the matrix is a sequence of dates arranged in sequence at the same time point, which characterizes the periodic variation characteristics of the flow.

3) The above time series and date series form a model dataset, which provides basic data for the input of the long-short-term memory neural network.

3. Establish long and short-term memory neural network model

1) Perform feature extraction on the above model data set, extract the date sequence within 7 days of the traffic data at time t and the time sequence within 35 minutes (7 sampling time points), and perform normalization processing according to the aforementioned formula (2), as the feature input model, where,

The first time period is the date sequence traffic data at time t-1, including t-1-24h*7, t-1-24h*6, t-1-24h*5, t-1-24h*4, t- Traffic data at 1-24h*3, t-1-24h*2, t-1-24h*1;

The second time period is the date sequence flow data at time t+1, including t+1-24h*7, t+1-24h*6, t+1-24h*5, t+1-24h*4, t+ Traffic data at 1-24h*3, t+1-24h*2, t+1-24h*1;

The third time period is the date sequence traffic data at time t, including t-24h*7, t-24h*6, t-24h*5, t-24h*4, t-24h*3, t-24h*2, Traffic data at time t-24h*1;

The fourth time period is the time series flow data near time t, including the flow data of t-35min, t-30min, t-25min, t-20min, t-15min, t-10min, t-5min, and time;

The above four time periods are respectively input into the long-short-term memory neural network model with four dimensions and seven moments.

2) The Long Short-Term Memory Neural Network (LSTM) provided by the present disclosure is an evolutionary model of the recurrent neural network, which is more suitable for learning from experience to classify time series, processing and forecasting. It is characterized by the ability to compress the input vector representation in the recurrent neural network and update the prediction output of the model with long and short-term memory.

The model consists of 3 LSTM layers and 3 fully connected neural network layers as input layers, forming a multi-level input transfer layer, which is configured to receive the above-mentioned four-dimensional traffic feature data. The number of nodes in the three LSTM layers is set to 128, 64, and 48, respectively, and the activation function of the LSTM layer is tanh. The activation function of the first and second fully connected neural network layers is ReLU, and the activation function of the third fully connected neural network layer is Linear. The third fully connected neural network layer is connected to the output layer. The activation function of the output layer is Linear, the learning rate is 0.002, and the batch_size is 60.

Among them, 28,000 samples are selected for the training set, and the input samples of historical traffic characteristics are randomly scrambled. Among them, 25,200 samples are randomly selected as training samples, and 5,184 samples are selected for the verification set.

The input of the model is the normalized traffic characteristic data of the above four dimensions.

The output of the model is the flow of the next period of the DMA entry.

4. Identification of abnormal traffic points

1) Calculate as formula (3) Calculate the measured value sequence X _t and the predicted value sequence of the model training set The residual sequence R _t between .

2) The training set residual sequence R _t is subjected to periodization and segmentation processing. The model training set contains a total of 87 days of traffic data, and the residual sequence R _t is divided into 288 groups at intervals of 5 minutes, each group contains 87 data, that is, each group corresponds to the change of data at the same time with the number of days.

3) Calculate the 3σ interval of formula (4) for each group of residuals above, and take the upper and lower bounds of the 3σ interval as the upper and lower residual thresholds at the moment.

4) Calculate the residual of the validation set, and mark an abnormal flow point when the residual value is outside the 3σ interval. Among them, when the residual value is greater than the upper residual threshold, it is marked as a high traffic point. Table 1 is a partial example of the timed residual threshold (upper residual threshold). This threshold can be regarded as the minimum flow rate that can be identified by the present disclosure, that is, the minimum leakage range that can be identified by the present disclosure is between 2.85% and 13.14% of the percentage of the average daily flow rate, which has strong practicality in the actual pipe network and high accident sensitivity. characteristic.

Table 1 Residual thresholds

5. Identification of leakage of pipeline network:

1) Correction of abnormal flow points.

The 288 groups of residual sequences in the training set in 4(2) are averaged to form a residual correction sequence, corresponding to 288 moments.

Use the validation set data to verify the model effect. That is, during the period from November 3, 2015 to November 21, 2017, when a residual is detected as an abnormal traffic point, the traffic at this point needs to be marked and replaced in time. The replaced value is used as input to enter the model again, and the flow calculation at the next moment is performed. The replacement method is to replace the abnormal value at time t with the sum of the flow prediction value at time t and the residual correction sequence, and use it as the input to predict the flow at time t+1. Taking accident 1 in Table 2 as an example, after the residual is greater than the upper residual threshold at 2:20 on November 17, it is marked as a high-flow point, and the abnormal flow value at this moment is replaced as above.

2) Accident identification and alarm steps. In this example, the threshold Q of the number of alarms is set to 2, that is, when the number of continuous high-flow points is greater than 2, the alarm is activated, and it is recognized that the leakage accident has occurred at this moment. Inside, with fast detection characteristics.

This implementation case includes the real simulated leakage accident in the pipeline network, that is, the flow change of the leakage accident is simulated by turning on the fire hydrant. The included accidents and identification results are shown in Table 2. The present disclosure successfully identifies the flow change of the real leakage experiment, and uses the commonly used evaluation method confusion matrix method to evaluate the model recognition effect. The accurate alarm rate is as high as 100%, and the false alarm rate is as low as 0.19%. reporting rate and other characteristics.

Table 2 Identification results

The above results show that the leakage identification method based on the long-short-term memory neural network model and the multi-threshold feedback correction model can identify leakage accidents in the pipeline network quickly, accurately, with high accuracy and low false alarm rate, and the method is practical Strong performance, strong fault tolerance for data quality. The present disclosure expands and strengthens the research content of the existing water supply network identification model, and provides new ideas for water companies to make scientific and reasonable decisions.

In addition, the above definitions of various elements and methods are not limited to various specific structures, shapes or manners mentioned in the embodiments, and those of ordinary skill in the art can simply modify or replace them.

It should be noted that the directional terms mentioned in the embodiments, such as "up", "down", "front", "rear", "left", "right", etc., only refer to the directions of the drawings, not to limit the scope of protection of the present disclosure. Throughout the drawings, the same elements are denoted by the same or similar reference numbers. Conventional structures or constructions will be omitted when it may lead to obscuring the understanding of the present disclosure. Moreover, the shapes and sizes of the components in the figures do not reflect the actual size and proportion, but merely illustrate the contents of the embodiments of the present disclosure. Furthermore, in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim.

Furthermore, the word "comprising" or "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements.

The ordinal numbers such as "first", "second", "third", etc. used in the description and the claims are used to modify the corresponding elements, which themselves do not mean that the elements have any ordinal numbers, nor do they Representing the order of a certain element and another element, or the order in the manufacturing method, the use of these ordinal numbers is only used to clearly distinguish an element with a certain name from another element with the same name.

Similarly, it will be appreciated that in the above description of exemplary embodiments of the disclosure, various features of the disclosure are sometimes grouped together into a single embodiment, figure, or its description. However, this method of disclosure should not be interpreted as reflecting an intention that the claimed disclosure requires more features than are expressly recited in each claim. Rather, as the following claims reflect, disclosed aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the Detailed Description are hereby expressly incorporated into this Detailed Description, with each claim standing on its own as a separate embodiment of the present disclosure.

The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present disclosure in detail. It should be understood that the above-mentioned specific embodiments are only specific embodiments of the present disclosure, and are not intended to limit the present disclosure. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure should be included within the protection scope of the present disclosure.

Claims (10)

1. A pipe network leakage identification method based on a long-time memory neural network model comprises the following steps:

s1, obtaining DMA entry data;

s2, cleaning the obtained DMA entry data, and constructing a multi-scale time data set;

s3, establishing a long-term and short-term memory neural network model;

s4, identifying abnormal flow points based on the built multi-scale time data set and the built long-time and short-time memory neural network model;

and S5, identifying the leakage of the pipe network according to the identified abnormal flow points.

2. The method according to claim 1, wherein, in step S1, when the DMA is a single entry, the DMA single entry traffic data is obtained; when the DMA is a multi-entry, acquiring the flow sum of the DMA multi-entry; the flow data sampling interval isThe sampling number is 24k times in minutes/time and all the day, namely 24k pieces of historical flow data are acquired each day.

3. The method according to claim 2, wherein the step S2 includes the sub-steps of:

s21, cleaning the acquired DMA entry data, and constructing a continuous time sequence by using the cleaned data;

s22, segmenting the continuous time sequence to construct a date sequence;

and S23, constructing a multi-scale time data set based on the time sequence and the date sequence, wherein the multi-scale time data set is input data of the long-time and short-time memory neural network model.

4. The method of claim 3, wherein the step of flushing the obtained DMA entry data comprises checking DMA entry data for consistency and padding missing values.

5. The method according to claim 3, wherein the step S3 includes:

taking the constructed multi-scale time data set as model input; wherein the multi-scale temporal data set comprises a training set and a validation set;

taking the predicted flow of the DMA inlet at the next moment as a model to be output; and

selecting tanh, ReLU or Linear as an activation function of each network layer of the model, and constructing the long-term memory neural network model.

6. The method of claim 5, wherein, in the step of inputting the constructed multi-scale temporal dataset as a model,

selecting an adjacent date sequence of t-time flow to be predicted, namely date sequence data of t-1 time, t +1 time and t time as model input, and setting the date sequence data as a first dimension time period, a second dimension time period and a third dimension time period; and

and selecting the adjacent time sequence data of the flow at the t moment to be predicted as model input, namely setting the previous c moments of the t moment as a fourth dimension time period.

7. The method according to claim 6, wherein the step S4 includes:

s41, calculating residual error, and calculating the predicted flow output by the long-time memory neural network model in the training set dataAnd measured value X_tResidual error R of_tThe calculation formula is as follows:

s42, cutting the residual error, and dividing the residual error R_tAnd (4) segmenting according to time points, forming the same group of data by the same sampling time point, and obtaining 24k groups of residual sequences.

S43, calculating residual threshold after segmentation, calculating 3 sigma intervals for the 24k groups of residual errors respectively, wherein the upper and lower boundaries of the 3 sigma intervals form 24k groups of upper and lower residual threshold corresponding to 24k time points, and the nth time residual r of the mth day_mnThe 3 σ threshold interval calculation formula is as follows:

in the formula,is the residual mean value corresponding to the moment n; var [ R ]_n]Is the residual standard deviation corresponding to the moment n;

s44, identifying abnormal flow points, calculating flow residual errors of the verification set at the moment t, and if the residual error value at the moment t is greater than an upper residual error threshold corresponding to the moment t, determining the time point as an abnormal flow point and marking as a high flow point; and if the residual value at the moment t is smaller than the lower residual threshold corresponding to the moment, determining that the time point is an abnormal flow point and recording as a low flow point.

8. The method according to claim 7, wherein the step S5 includes:

s51, correcting the abnormal flow point and inputting the abnormal flow point into the long-time and short-time memory neural network model for calculation;

and S52, counting the number of continuous high-flow points according to the calculation result obtained by inputting the corrected long-time and short-time memory neural network model, and identifying the leakage of the pipe network according to the number of the continuous high-flow points.

9. The method according to claim 8, wherein in the step S51, when the nth time on the mth day is determined as the abnormal flow rate point, the correction value of the flow rate data at that time isWherein,the flow prediction value output by the long-time memory neural network corresponding to the nth time of the mth day is obtained;is the residual error correction value at the moment, which is R in the residual error sequence of the training set_nThe average value of the values is calculated,

10. the method as claimed in claim 9, wherein in the step S52, the number Q of continuous high flow points is counted, when Q is greater than or equal to a threshold Q, a pre-warning is started, and the leakage accident duration T is calculated_burst(ii) a The leakage accident detection time calculation formula is as follows:

the leakage accident duration calculation formula is as follows: