A semi-supervised iterative training method for rainfall-inflow event recognition in sewage networks

Abstract

This invention discloses a semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks, comprising the following steps: S1: acquiring real-time liquid level data, historical liquid level data, and real-time rainfall data; S2: preprocessing the two types of liquid level data respectively; S3: obtaining rainfall sequences, real-time liquid level sequences, and historical liquid level sequences at a unified time scale; S4: obtaining trend terms using STL; S5: obtaining a state recognition model; S6: obtaining candidate rainfall-inflow events; S7: determining whether the time interval threshold has been reached; if not, proceeding to the next step; if it has been reached, determining whether to adjust the period based on the frequency of candidate rainfall-inflow events; if not adjusted, proceeding to the next step; if adjusted, returning to S4; S8: outputting the results in real time. This invention combines STL and HMM models to identify the hidden state sequences of liquid level changes, achieving long-term, stable, and low-labor-cost identification of rainfall-inflow events in sewage pipe networks.

Description

Technical Field

[0001] This invention belongs to the field of pipeline monitoring technology, and in particular relates to a semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipelines. Background Technology

[0002] Urban drainage systems are critical infrastructure for ensuring the safe operation of cities and the quality of life for residents. With the continuous advancement of urbanization, problems such as combined sewer overflows, aging pipes, and illegal connections are becoming increasingly prominent, leading to significant inflows into drainage networks during rainfall. These inflow events not only increase the treatment load on sewage treatment plants but also pose environmental risks such as sewage overflows and water pollution. Therefore, timely and accurate identification of abnormal increases in pipe network levels caused by rainfall is of great importance for urban water management, drainage system maintenance, and risk warning.

[0003] Currently, the identification of rainfall inflow events mainly relies on human experience or threshold-based judgment methods. For example, preliminary judgments are made by observing whether a sudden increase in liquid level coincides with the rainfall period or whether the flow rate change exceeds the empirical range. However, such methods have obvious limitations: on the one hand, human judgment is inefficient and highly subjective, making it difficult to adapt to large-scale, multi-site online monitoring data; on the other hand, static thresholds cannot cope with diverse drainage characteristics and dynamically changing environmental conditions, resulting in a high misjudgment rate and poor adaptability. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks that is highly accurate, adaptable, and capable of automatic identification.

[0005] Technical solution: This invention discloses a semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks, comprising the following steps:

[0006] S1: Obtain real-time and historical liquid level data of the target monitoring station, as well as real-time rainfall data of the area corresponding to the target monitoring station;

[0007] S2: Perform outlier removal and missing value imputation preprocessing on real-time liquid level data and historical liquid level data respectively;

[0008] S3: Resample the real-time rainfall data, preprocessed real-time liquid level data, and historical liquid level data to obtain rainfall sequences, real-time liquid level sequences, and historical liquid level sequences at a unified time scale;

[0009] S4: The STL time series decomposition method is used to decompose the real-time liquid level series and the historical liquid level series respectively, and the trend term T of the real-time liquid level series is obtained respectively. t实Trend term T of historical liquid level series t历 ;

[0010] S5: Construct a Hidden Markov Model and utilize the trend term T t历 Using the time series as the observation sequence, the Baum-Welch expectation-maximization algorithm is used to train the hidden Markov model in unsupervised training to obtain a state recognition model for predicting the hidden state sequence of liquid level.

[0011] S6: The trend term T of the real-time liquid level data... t实 Input the state recognition model, and the state recognition model outputs the liquid level hidden state sequence; filter the "high liquid level" state segments in the liquid level hidden state sequence, and identify the "high liquid level" state segments that meet the preset conditions as candidate rainfall-inflow events;

[0012] S7: Set the time interval threshold for calibrating the state recognition model. If the time interval threshold is not reached, proceed directly to the next step. When the time interval threshold is reached, calculate the occurrence frequency of the candidate rainfall-inflow event within a specified period and compare it with the frequency of human experience or the frequency of manually labeled samples. If the difference is less than the preset frequency error threshold, proceed to the next step. If the difference is greater than or equal to the frequency error threshold, increment the period parameter in the STL time series decomposition method at equal intervals and repeat steps S4 to S7 to achieve adaptive adjustment and dynamic iterative training of the state recognition model until the difference is less than the preset frequency error threshold.

[0013] S8: Output the finally identified candidate rainfall-inflow events in a structured record format in real time for subsequent display, analysis or decision support.

[0014] Furthermore, in step S2, the preprocessing method for real-time liquid level data and historical liquid level data is as follows: an adaptive 3σ anomaly detection algorithm based on a sliding time window is used to identify outliers in the two types of liquid level data respectively. After removing the identified outliers, a linear interpolation method within the nearest time window is used to fill in the outliers and missing values.

[0015] Furthermore, in step S2, the adaptive 3σ anomaly detection algorithm based on a sliding time window is used to identify outliers in the liquid level data as follows:

[0016] In the formula, LB represents the lower limit of the normal value of the liquid level data in the past week, UB represents the upper limit of the normal value of the liquid level data in the past week, μ represents the mean of the liquid level data of the target monitoring station in the past week, and σ represents the standard deviation of the liquid level data of the target monitoring station in the past week.

[0017] If the monitored value of the liquid level data is lower than LB or higher than UB, it is judged as an abnormal value;

[0018] The method used to impute outliers and missing values ​​using linear interpolation within a nearby time window is as follows:

[0019] In the formula, S t D represents the missing or outlier values ​​in the liquid level data at time t. t S represents t timestamp; S t-1 Indicates the distance S in the liquid level data t The most recent normal value, D t-1 S represents t-1 timestamp; S t+1 Indicates the distance S in the liquid level data t The most recent normal value, D t+1 S represents t+1 Timestamp.

[0020] Furthermore, the resampling method in step S3 is the sliding window averaging method, and the sampling method is as follows: determine the sampling frequency, the initial period of the sampling time series, and the final period of the sampling time series, and calculate the normalized time point series based on this; for each target sampling point, take the data within a range of T minutes in front of it as the sliding window, and calculate the average liquid level within the window, and use the average value as the resampling value of the target sampling point.

[0021] Furthermore, in step S4, the STL time series decomposition method, based on the local weighted regression algorithm, decomposes the two liquid level sequences into the following equations:

[0022] In the formula, y t S represents liquid level sequence data. t T represents the seasonal term. t R represents the trend term. t This represents the residual term.

[0023] Furthermore, the steps for training a Hidden Markov Model are as follows:

[0024] S51: The Hidden Markov Model (HMM) with two hidden states is selected for training. The number of hidden states of the HMM is set to 2. Hidden state S1 represents the "normal fluctuation" state of the liquid level under normal operation, and hidden state S2 represents the "high liquid level" state of the liquid level under abnormal operation. The set of hidden states is S={S1,S2}.

[0025] Using the trend term T from historical liquid level data t历 The liquid level monitoring value represents observation state O, and the set of observation states is... ;

[0026] S52: Using the K-means clustering method to analyze the trend term T of historical liquid level data. t历 Perform clustering to divide the data into two clusters;

[0027] Based on the clustering results, an initial observation probability distribution parameter set {μ1, σ1, μ2, σ2} is constructed, where μ1 is the observation mean of hidden state S1, σ1 is the observation standard deviation σ1 of hidden state S1, μ2 is the observation mean of hidden state S2, and σ2 is the observation standard deviation σ2 of hidden state S2.

[0028] The initial state probability vector is obtained based on the clustering results, and the trend term T contained in the two data clusters is calculated respectively. t历 Quantity in the overall trend term T t历 The proportion π in the vector gives the initial state probability vector π = [p1, p2], where p1 refers to a certain trend term T. t历 The probability of belonging to S1, p2 refers to the trend term T. t历 The probability of belonging to S2, and p1+p2=1;

[0029] S53: Construct the initial state transition probability matrix A. And set A as a high self-transformation probability structure; where a ij This represents the probability that when the hidden state is i at time t, the hidden state changes to j at time t+1.

[0030] S54: Calculate the observed state probability matrix B. The observation probability density function in each hidden state follows a one-dimensional normal distribution. , Indicates that in the hidden state S i Below, the liquid level value o was observed at time t. N The observation probability density; μ i Hidden state S i The mean parameter, σ i Hidden state S i The standard deviation parameter;

[0031] S55: The Baum-Welch expectation-maximization algorithm is used for unsupervised iterative training of the HMM, through maximization of the trend term T. t历 The log-likelihood function of the corresponding observation sequence is used to optimize the HMM model parameters;

[0032] S56: Set the stopping condition for unsupervised iterative training: the log-likelihood increment is lower than the preset convergence threshold or the number of iterations reaches the preset upper limit;

[0033] S57: After unsupervised iterative training stops, the resulting trained HMM is the intermediate model, and the intermediate model outputs the corresponding trend term T. t历 The hidden state time sequence of the observation sequence is obtained, and then the difference between the observation means of two hidden states is compared. If the difference is greater than or equal to a preset threshold, the intermediate model converges and is used as the state recognition model; if the difference is less than the preset threshold, the intermediate model is optimized and trained again until the difference is greater than or equal to the preset threshold.

[0034] Furthermore, the preset threshold in step S57 is set as follows:

[0035] Let T be the trend term of the historical liquid level series. t历 The sequence of liquid level monitoring values ​​is {x1, x2, x3, ..., x Q};

[0036] Calculate {x1,x2,x3,…,x} Q The mean of} ;

[0037] Based on mean Calculate {x1,x2,x3,…,x} Q The global standard deviation σ of} global and the global standard deviation σ global As a preset threshold , where q∈[1,Q].

[0038] Furthermore, the process of optimizing the intermediate model training in step S57 is as follows:

[0039] S571: If the difference between the observed means of two hidden states is less than a preset threshold, adjust the elements in the initial state transition probability matrix A; the automatic adjustment method is: increase The conversion probability is reduced. The conversion probability, keep The transition probability;

[0040] S572: Obtain historical liquid level data under conditions of no heavy rainfall interference, and use the STL time series decomposition method to obtain the trend term of the historical liquid level data. Then, use the probability of this trend term being assigned to hidden states S1 and S2 to reset the initial state probability vector π.

[0041] If the proportion of consecutive high liquid level segments in the training set exceeds 20%, increase p2 to the range of 0.1~0.2;

[0042] S573: After adjusting A and π, return to step S55 to perform unsupervised iterative training on the HMM again until the difference between the observed means of the two hidden states in step S57 is greater than or equal to the preset threshold. If the condition that the difference between the observed means of the two hidden states is greater than or equal to the preset threshold is not met after 3 iterations, the intermediate model obtained in the last iteration is used as the state recognition model.

[0043] Furthermore, the preset conditions in step S6 are: ① The duration of the latent state time series of the "high liquid level" state is not less than T1 hours and the cumulative rainfall of a single rainfall is not less than R millimeters; ② The interval between the latent state time series of the "high liquid level" state and the previous rainfall is not less than T2 hours; if conditions ① and ② are met at the same time, it is identified as a candidate rainfall-inflow event.

[0044] In step S7, the seasonal period parameter (period) of the STL time series decomposition method is adjusted as follows: initially, the seasonal period parameter (period) is set to period0; during the nth iteration, the seasonal period parameter (period) is period0 + nΔ. period , where Δ period The increment step size for the periodic parameter.

[0045] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention effectively separates trend, periodic, and noise components in liquid level data using the STL time-series decomposition algorithm. Combined with a Gaussian Hidden Markov Model (HMM) to identify the hidden state sequence of liquid level changes, it can accurately determine whether liquid level anomalies are caused by rainfall, improving the accuracy of rainfall inflow event identification. In this invention, the training of the HMM constitutes the inner iteration, ensuring full convergence in each training iteration. Adjusting the seasonal period parameter (period) of the STL time-series decomposition method based on the matching degree constitutes the outer iteration; this dual iteration ensures the robustness and accuracy of the state identification model. Furthermore, this invention combines STL and HMM models to identify the hidden state sequence of liquid level changes without requiring manual annotation, possesses self-learning capabilities, and is suitable for deployment on intelligent monitoring platforms, enabling long-term, stable, and low-manual-cost identification of rainfall inflow events.

[0046] This invention progressively expands the seasonal period parameter (period) of the STL time series decomposition method, which can suppress the influence of atypical short-term fluctuations, realize multi-scale trend extraction from short-term to medium- and long-term, and introduce an adjustable and adaptive time scale dimension for Markov model to identify hidden states. Unlike the existing technology that decomposes time series statically with a fixed period, this invention dynamically adjusts the seasonal period parameter (period) of the STL time series decomposition method to adapt to the differences in the liquid level response scale of urban drainage systems under different seasons, weather conditions, or operating conditions, avoiding the situation where abnormal signals are buried or misjudged due to the fixed-period static decomposition of time series. Attached Figure Description

[0047] Figure 1 This is a flowchart of the present invention;

[0048] Figure 2 This is a comparison chart of trend terms after adjusting the seasonal period parameter (period) according to an embodiment of the present invention.

[0049] Figure 3 This is a diagram showing the real-time liquid level data recognition results of an embodiment of the present invention;

[0050] Figure 4 This is a result output diagram of an embodiment of the present invention. Detailed Implementation

[0051] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0052] Example 1

[0053] This invention discloses a semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks, such as... Figure 1 As shown, it includes the following steps:

[0054] S1: Obtain real-time and historical liquid level data of the target monitoring station, as well as real-time rainfall data of the area corresponding to the target monitoring station.

[0055] The sampling frequency for real-time liquid level data is no less than five minutes each time; real-time rainfall data is the actual rainfall data measured at the target monitoring station or grid rainfall data that has been inverted and corrected.

[0056] S2: Perform outlier removal and missing value imputation preprocessing on real-time liquid level data and historical liquid level data respectively.

[0057] The preprocessing method for real-time and historical liquid level data is as follows: an adaptive 3σ anomaly detection algorithm based on a sliding time window is used to identify outliers in both types of liquid level data. After removing the identified outliers, a linear interpolation method within the nearest time window is used to fill in the outliers and missing values.

[0058] The method for identifying outliers in liquid level data using an adaptive 3σ anomaly detection algorithm based on a sliding time window is as follows:

[0059] In the formula, LB represents the lower limit of the normal value of the liquid level data in the past week, UB represents the upper limit of the normal value of the liquid level data in the past week, μ represents the mean of the liquid level data of the target monitoring station in the past week, and σ represents the standard deviation of the liquid level data of the target monitoring station in the past week.

[0060] If the monitored value of the liquid level data is lower than LB or higher than UB, it is judged as an abnormal value.

[0061] To address common issues in drainage monitoring data such as missing values ​​and abnormal fluctuations, this invention employs a sliding window 3σ method to dynamically identify outliers and uses a time interpolation algorithm for repair, significantly improving data availability and the input quality of subsequent HMM models, and avoiding misjudgments caused by abnormal data.

[0062] The method used to impute outliers and missing values ​​using linear interpolation within a nearby time window is as follows:

[0063] In the formula, S t D represents the missing or outlier values ​​in the liquid level data at time t. t S represents t timestamp; S t-1 Indicates the distance S in the liquid level data t The most recent normal value, D t-1 S represents t-1 timestamp; S t+1 Indicates the distance S in the liquid level data t The most recent normal value, D t+1 S represents t+1 The timestamp. A timestamp is represented as 2023-11-30 11:00:00.

[0064] S3: Resample the real-time rainfall data, preprocessed real-time liquid level data, and historical liquid level data to obtain rainfall sequences, real-time liquid level sequences, and historical liquid level sequences at a unified time scale.

[0065] The resampling method is a sliding window averaging method, and the sampling procedure is as follows: The sampling frequency, the initial time period of the sampling time series, and the final time period of the sampling time series are determined, and a normalized time point series is calculated based on these. For each target sampling point, data within a preceding T minutes is taken as a sliding window, and the average liquid level within this window is calculated. This average value is used as the resampling value for the target sampling point. Using the window average as the value for this time point allows for the aggregation of high-frequency data, thereby obtaining rainfall sequences, real-time liquid level sequences, and historical liquid level sequences at regular time intervals.

[0066] S4: The STL time series decomposition method is used to decompose the real-time liquid level series and the historical liquid level series respectively, and the trend term T of the real-time liquid level series is obtained respectively. t实 Trend term T of historical liquid level series t历 .

[0067] The STL time series decomposition method, based on the local weighted regression algorithm, decomposes the liquid level sequence into the following formula:

[0068] In the formula, y t S represents liquid level sequence data. t T represents the seasonal term. t R represents the trend term. t The residual terms are represented by the trend term, which reflects long-term changes in liquid level, the seasonal term captures periodic fluctuations (such as diurnal or seasonal variations), and the residual terms retain random fluctuations and anomalies. The STL time series decomposition method can remove periodic normal fluctuations, making abnormal changes more significant in the residuals. Furthermore, using the trend term instead of the liquid level sequence as input to the state recognition model in subsequent steps can reduce the adverse effects of fluctuations and noise in pipeline liquid level data on the recognition performance.

[0069] S5: Construct a Hidden Markov Model and utilize the trend term T t历 Using the time series as the observation sequence, the hidden Markov model is trained unsupervised using the Baum-Welch expectation-maximization algorithm to obtain a state recognition model for predicting the hidden state sequence of liquid level.

[0070] This invention constructs an automated identification process using the STL temporal decomposition method and a hidden Markov model, enabling accurate judgment of potential rainfall inflow events in multi-source monitoring data, which is beneficial to improving overall identification efficiency and engineering adaptability.

[0071] The steps for training a Hidden Markov Model are as follows:

[0072] S51: The double-hidden-state Gaussian HMM in the Hidden Markov Model is selected for training. The characteristics of the double-hidden-state Gaussian HMM are more suitable for the continuity and fluctuation characteristics of liquid level change data. The number of hidden states of the HMM is set to 2. Hidden state S1 represents the "normal fluctuation" state of the liquid level under normal operation, and hidden state S2 represents the "high liquid level" state of the liquid level under abnormal operation. The set of hidden states is S={S1,S2}. The "high liquid level" state is commonly seen in the state of "continuous rise" or "high liquid level" caused by external disturbances.

[0073] Using the trend term T from historical liquid level data t历 The liquid level monitoring value represents observation state O, and the set of observation states is... .

[0074] S52: Using the K-means clustering method to analyze the trend term T of historical liquid level data. t历 Clustering is performed to divide the data into two clusters. The purpose of this step is to pre-segment the data using the unsupervised K-means clustering method (with a set number of categories of 2) based on the global statistical characteristics of historical liquid level data, and to obtain a preliminary state division.

[0075] Based on the clustering results, an initial observation probability distribution parameter set {μ1, σ1, μ2, σ2} is constructed, where μ1 is the observation mean of hidden state S1, σ1 is the observation standard deviation of hidden state S1, μ2 is the observation mean of hidden state S2, and σ2 is the observation standard deviation of hidden state S2. Furthermore, μ1 is the mean of the data cluster corresponding to hidden state S1, σ1 is the variance of the data cluster corresponding to hidden state S1, and μ2 is the mean of the data cluster corresponding to hidden state S2, and σ2 is the variance of the data cluster corresponding to hidden state S2.

[0076] The initial state probability vector is obtained based on the clustering results, and the trend term T contained in the two data clusters is calculated respectively. t历 Quantity in the overall trend term T t历 The proportion π in the vector gives the initial state probability vector π = [p1, p2], where p1 refers to a certain trend term T. t历 The probability of belonging to S1, p2 refers to the trend term T. t历 The probability of belonging to S2 is p1 + p2 = 1.

[0077] S53: Construct the initial state transition probability matrix A. And set A as a high self-transformation probability structure; where a ij Let represent the probability that the hidden state at time t+1 changes to j when the hidden state is i at time t; where a high self-transition probability structure refers to the existence of some states i such that a ii Much greater than a ij (and j ≠ i), meaning the probability of staying in the current state is much greater than the probability of transitioning to another state; for example, if the diagonal element of A is greater than 0.8, such as... This is to enhance the temporal continuity and stability of the state.

[0078] S54: Calculate the observed state probability matrix B. The observation probability density function in each hidden state follows a one-dimensional normal distribution. , Indicates that in the hidden state S i Below, the liquid level value o was observed at time t. N The observation probability density; this observation probability is defined by the above observation probability density function (Gaussian distribution, i.e., one-dimensional normal distribution), where μ i Hidden state S i The mean parameter, σ i Hidden state S i The standard deviation parameter; this function can be used to calculate the likelihood of any observation in a given hidden state, which is used for inference of hidden state sequences and model training, that is, when o is determined. N After determining the hidden state, the observed values ​​are generated based on the observation probability density function (Gaussian distribution, i.e., one-dimensional normal distribution) corresponding to the hidden state.

[0079] S55: The Baum-Welch expectation-maximization algorithm is used for unsupervised iterative training of the HMM, through maximization of the trend term T. t历 The log-likelihood function of the corresponding observation sequence is used to optimize the HMM model parameters.

[0080] S56: Set the stopping condition for unsupervised iterative training: the log-likelihood increment is lower than a preset convergence threshold or the number of iterations reaches a preset upper limit; in this embodiment, the convergence threshold is set to 10⁻ 4 Set the upper limit of the number of iterations to 100.

[0081] S57: After unsupervised iterative training stops, the resulting trained HMM is the intermediate model, and the intermediate model outputs the corresponding trend term T. t历 The hidden state time series of the observed sequence is obtained, and then the difference between the observed means of two hidden states is compared. If the difference is greater than or equal to a preset threshold, the intermediate model converges and is used as the state recognition model; if the difference is less than the preset threshold, the intermediate model is optimized and trained again until the difference is greater than or equal to the preset threshold. When the difference between the observed means of two hidden states is less than the preset threshold, it indicates that "normal state" and "high liquid level" are difficult to separate, and the structure of the state recognition model needs to be adjusted.

[0082] Preferably, the preset threshold is set as follows: let the trend term T of the historical liquid level sequence be... t历 The sequence of liquid level monitoring values ​​is {x1, x2, x3, ..., x Q};

[0083] Calculate {x1,x2,x3,…,x} Q The mean of} ;

[0084] Based on mean Calculate {x1,x2,x3,…,x} Q The global standard deviation σ of} global and the global standard deviation σ global As a preset threshold , where q∈[1,Q].

[0085] The process of further optimizing the training of the intermediate model is as follows:

[0086] S571: If the difference between the observed means of two hidden states is less than a preset threshold, adjust the elements in the initial state transition probability matrix A; adjusting the elements in the initial state transition probability matrix A is to enhance the model's discriminative ability, and the automatic adjustment method is: increase The conversion probability is reduced. The conversion probability, keep The conversion probability. Increase This involves increasing the self-transfer probability of the high liquid level state, for example, by increasing S2 to 0.95, to enhance the stability of the "high liquid level" state; and reducing... This reduces the probability of transitioning from normal to high liquid level, for example, to 0.05, to avoid slight fluctuations being misjudged as abnormal; maintaining... The transition probability is set in order to maintain the natural continuity of the "normal state", for example, by keeping it around 0.9.

[0087] S572: Acquire historical liquid level data under conditions of no heavy rainfall interference, and use the STL time series decomposition method to obtain the trend term of this historical liquid level data. Then, use the probability of this trend term being assigned to hidden states S1 and S2 to reset the initial state probability vector π. Under conditions of no heavy rainfall interference, the liquid level time series exhibits a low-frequency trend pattern characterized by daily periodic fluctuations, stable changes, and no drastic increases or decreases. By decomposing the liquid level series using the STL time series decomposition method and extracting this trend component as the basis for modeling, the interference of short-term fluctuations on the identification of hidden states can be eliminated.

[0088] In step S52, the initial state probability vector π is set to π=[0.95,0.05] by default, meaning the model is more likely to start from the "normal state".

[0089] If the proportion of consecutive high liquid level segments in the training set exceeds 20%, increase p2 to the range of 0.1~0.2;

[0090] After adjusting A and π, return to step S55 to perform unsupervised iterative training of the HMM again until the difference between the observed means of the two hidden states in step S57 is greater than or equal to a preset threshold. If the condition that the difference between the observed means of the two hidden states is greater than or equal to the preset threshold is not met after 3 iterations, the intermediate model obtained in the last iteration is used as the state recognition model. Setting an upper limit of 3 iterations helps to ensure a balance between efficiency and performance in training the HMM.

[0091] S6: The trend term T of the real-time liquid level data... t实 Input the state recognition model, and the state recognition model outputs the liquid level hidden state sequence; filter the "high liquid level" state segments in the liquid level hidden state sequence, and identify the "high liquid level" state segments that meet the preset conditions as candidate rainfall-inflow events.

[0092] When the trend term T of the real-time liquid level data is... t实Using the observed sequence as input to the state recognition model, the corresponding hidden state time series can be obtained. In the liquid level data, this hidden state time series represents the current liquid level status, such as high, normal, or low. In pipeline rainfall inflow situations, the liquid level typically exhibits persistently high values. Therefore, the Hidden State Model (HMM) can be used to identify the hidden state of the liquid level data, thereby automatically separating the time periods of high liquid level values.

[0093] The preset conditions in step S6 are: ① The duration of the hidden state time series of the "high liquid level" state is not less than T1 hours and the cumulative rainfall of a single rainfall is not less than R millimeters; ② The interval between the hidden state time series of the "high liquid level" state and the previous rainfall is not less than T2 hours; if conditions ① and ② are met at the same time, it is identified as a candidate rainfall-inflow event.

[0094] In this embodiment, T1 is set to 3 hours, R to 2 mm, and T2 to 12 hours. In practical applications, T1, R, and T2 can be adjusted according to the regional hydrological characteristics. For the rainfall time series of the target monitoring point area, effective rainfall events are divided according to two conditions: the cumulative rainfall of a single rainfall event is greater than 2 mm and the interval between the previous rainfall event and the previous event is greater than 12 hours. For the hidden state time series of high liquid levels identified by the state recognition model, continuous high liquid level periods with a continuous repetition time of 3 hours or more are selected. Based on the effective rainfall periods and continuous high liquid level periods obtained in the above two steps, overlapping periods between the two periods are first selected, and then the intersection of the effective rainfall periods and the continuous high liquid level periods is the identified candidate rainfall-inflow event.

[0095] S7: Set the time interval threshold for calibrating the state recognition model. If the time interval threshold is not reached, proceed directly to the next step. When the time interval threshold is reached, calculate the occurrence frequency of the candidate rainfall-inflow event within a specified period and compare it with the difference between the frequency of human experience or the frequency of manually labeled samples. If the difference is less than the preset frequency error threshold, proceed to the next step. If the difference is greater than or equal to the frequency error threshold, increment the period parameter value in the STL time series decomposition method at equal intervals and repeat steps S4 to S7 to achieve adaptive adjustment and dynamic iterative training of the state recognition model until the difference is less than the preset frequency error threshold.

[0096] Preferably, the frequency error threshold is set to 20%.

[0097] Preferably, the time interval threshold is set to half a month, meaning that the frequency of occurrence of candidate rainfall-inflow events within a specified period is evaluated every half month against the frequency of human experience or manually labeled samples, and it is determined whether the seasonal period parameter (period) of the STL time series decomposition method needs to be adjusted. In practical applications, users can set the time interval threshold according to their actual needs. Setting the time interval threshold for calibrating the state recognition model not only avoids the problem of low operating efficiency of the state recognition model caused by real-time evaluation of the frequency of occurrence of candidate rainfall-inflow events within a specified period against the frequency of human experience or manually labeled samples, but also does not affect the normal application of the state recognition model.

[0098] Human experience frequency: By acquiring a small amount of prior human knowledge, the empirical frequency of inflow events during the dry and rainy seasons is constructed. For example, based on years of monitoring experience and three years of historical data, it is calculated that inflow events occur on average once every 20 days during a typical dry season and once every 5 days during the rainy season. This frequency can be regarded as a reference benchmark for the event recognition accuracy of the state recognition model.

[0099] The labeled sample frequency refers to the process of manually identifying and labeling some rainfall-inflow events under certain conditions, based on the output results of the preliminary model, to establish a real sample library for further evaluation of the accuracy of the state recognition model. The preliminary model refers to the state recognition model trained in step S5 that has not yet been input with real-time liquid level data.

[0100] For example, the initial iteration of the state recognition model showed events occurring approximately every 5 days during the dry season and every 2 days during the rainy season. However, based on historical experience, the actual frequency of events during the dry season is approximately once every 30 days, and during the rainy season, approximately once every 10 days. The state recognition model's results are too high, possibly due to oversensitivity caused by setting the STL parameter too small. Therefore, it is necessary to increase the seasonal period parameter (period) in subsequent iterations, i.e., increase the seasonal smoothing window.

[0101] The seasonal period parameter `period` in the STL time series decomposition method is adjusted as follows: initially, `period` is set to `period0`. In the nth iteration, `period` becomes `period0 + nΔ`. In this embodiment, `period0` is set to 24 hours, and `Δ` is set to 1 hour, meaning `period` increases by 1 hour with each iteration. Gradually expanding the seasonal period parameter `period` in the STL time series decomposition method also gradually increases the smoothing window width of the trend term, suppressing the influence of atypical short-term fluctuations and thus strengthening the expressive power of the "continuous liquid level rise" feature. Gradually expanding the seasonal period parameter `period` in the STL time series decomposition method not only achieves multi-scale trend extraction from short-term to medium-to-long-term, but also introduces an adjustable and adaptive time scale dimension for Markov model latent state identification, significantly different from existing technologies that only statically decompose time series with a fixed period. Figure 2 As shown in the figure, the seasonal period parameter period of the STL time series decomposition method is adjusted to 24 hours and 96 hours respectively. It can be directly seen from the figure that the trend time series of the same original liquid level are different after adjusting the seasonal period parameter period. Figure 2 In this context, "original liquid level" refers to the liquid level data, and "trend time series" refers to the trend term obtained after the liquid level data is decomposed by STL.

[0102] This process is repeated iteratively through steps S4-S7, continuously adjusting the parameters of the state recognition model to gradually bring the frequency of identified dry and rainy season events closer to empirical values. When the difference between the candidate rainfall-inflow event density output by the state recognition model and the frequency of manually experienced or manually labeled samples is less than the frequency error threshold, the parameters are considered to be reasonably selected, and the state recognition model has reached calibration equilibrium. In this embodiment, the training of the Hidden Markov Model constitutes the inner iteration, ensuring full convergence each time; the adjustment of the seasonal period parameter (period) of the STL time series decomposition method based on the matching degree constitutes the outer iteration. This dual iteration ensures the robustness and accuracy of the state recognition model.

[0103] Traditional STL time series decomposition methods are mostly based on fixed periods, which are difficult to adapt to the differences in the liquid level response scale of urban drainage systems under different seasons, weather or operating conditions, and are prone to burying or misjudging abnormal signals. This invention constructs a "trend density-aware time series decomposition strategy" by combining "equal interval period increment + outer layer trend density feedback iteration + inner layer HMM state recognition retraining", which can achieve the following effects: (1) Dynamic matching of trend scale: the gradual increment of the seasonal period parameter period can realize the smooth evolution of the trend term from daily changes to weekly scale, so that the HMM model can automatically capture the real response period of liquid level affected by rainfall; (2) Preventing overfitting and oversensitivity: a small seasonal period parameter period may misidentify occasional fluctuations as events. The increment strategy can effectively filter out short-term anomalies and improve the identification of stability. Qualitative; (3) Enhance the adaptive capability of HMM model: HMM model automatically completes self-calibration of trend extraction scale without human intervention; (4) Introduce structural feedback loop: guide the period increase by the feedback of the matching deviation between candidate rainfall-inflow event density and prior frequency, realize the collaborative optimization of outer trend adjustment and inner model training; (5) Break through the conventional static setting paradigm: existing STL time series decomposition methods mostly set the period based on experience. This invention proposes a clear dynamic adjustment logic and automatic convergence mechanism, which can adapt to the differences in liquid level response scale of urban drainage system under different seasons, weather or working conditions.

[0104] During the iterative steps S4-S7, the STL time-series decomposition method in step S4 involves re-decomposing the liquid level data based on the updated period value, using a ternary structure of seasonal, trend, and residual terms. The trend term T... t This is used in the next step, HMM modeling. Internally, the STL uses the LOESS local regression method to estimate the trend curve; the trend bandwidth, influenced by the period, determines the trend response time scale. Step S5 uses the updated trend sequence T. t The Hidden Markov Model (HMM) is retrained. The Baum-Welch algorithm is used to update the state transition matrix A, the initial observation probability distribution parameter set {μ1, σ1, μ2, σ2}, and the initial state probability vector π in each iteration until the log-likelihood function converges, ensuring that the HMM achieves a local optimum fit under each new trend. In steps S6 and S7, event recognition rules are applied to the "high liquid level" state sequence predicted by the state recognition model to re-extract rainwater inflow events and statistically analyze the event density during the dry and rainy seasons.

[0105] S8: Output the finally identified candidate rainfall-inflow events in a structured record format in real time for subsequent display, analysis or decision support.

[0106] Preferably, the final results are output in real time in a structured recording format, including the start and end times of rainfall-inflow events, corresponding rainfall amounts, and liquid level changes. Presenting these in a structured format facilitates integration into smart water management systems for automatic early warning, operation and maintenance scheduling, or model calibration. These results can be displayed intuitively on the monitoring platform or used by operation and maintenance personnel for further analysis and decision-making. Furthermore, since rainfall-inflow characteristics (start and end times of rainfall-inflow events, corresponding rainfall amounts, and liquid level changes) may change over time (e.g., due to climate change or pipeline renovation impacts), and the setting of time interval thresholds, this invention supports periodic retraining: after accumulating more unlabeled data, the above iterative calibration process is repeated periodically to update the state recognition model parameters.

[0107] This invention's process for identifying inflow events can be fully automated without manual intervention, making it suitable for deployment in urban drainage information platforms or water management systems. It achieves real-time response at the minute level and is applicable to the automatic processing of large-scale, multi-site liquid level data. Furthermore, the key identification parameters (such as T1, T2, and R) can be flexibly configured according to the characteristics of different cities or regions, exhibiting strong adaptability and scalability, and supporting refined management needs. This invention enables accurate identification and early warning of rainfall inflow events, effectively preventing urban risks such as pipe network overflows and water pollution, reducing unnecessary loads on sewage treatment systems, lowering manual inspection costs and maintenance manpower, and facilitating the transformation of urban drainage systems towards digitalization and intelligence.

[0108] Example 2

[0109] This invention discloses a semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks, applied to the SCADA00182 station in a drainage system of a certain urban area in Beijing. The method includes the following steps:

[0110] Using SCADA00182 as the target monitoring station, step S1 in Example 1 is executed, wherein the time resolution of the real-time liquid level data is 5 minutes, the acquisition period is not less than 30 days, and the time resolution of the real-time rainfall data is 5 minutes. The real-time liquid level data acquired in this example is as follows: Figure 3 As shown in b and c in the figure, the real-time rainfall data collected in this embodiment is as follows: Figure 3 As shown in 'a'.

[0111] Perform step S2 in Example 1 to preprocess the real-time liquid level data and historical liquid level data by removing outliers and filling in missing values.

[0112] Perform step S3 in Example 1, and normalize the rainfall data, preprocessed real-time liquid level data and historical liquid level data into a time series of 1 data point per hour. Calculate the mean of each time point and the window value of the previous 60 minutes, and use the mean of the window value as the value of that time point.

[0113] In step S4 of Example 1, the STL time series decomposition method is used to decompose the real-time liquid level sequence and the historical liquid level sequence to obtain the trend term T of the real-time liquid level sequence. t实 Trend term T of historical liquid level series t历 .

[0114] Perform step S5 in Example 1 to construct a Hidden Markov Model and utilize the trend term T. t历 Using the time series as the observation sequence, the hidden Markov model is trained unsupervised using the Baum-Welch expectation-maximization algorithm to obtain a state recognition model for predicting the hidden state sequence of liquid level.

[0115] S6: The trend term T of the real-time liquid level data... t实 Input the state recognition model obtained in step S5 to obtain the hidden states of the corresponding observation sequence, such as... Figure 3 As shown in b in the figure.

[0116] The state recognition model outputs a hidden state time series indicating a high liquid level. Filtering conditions are set, and the hidden state time series indicating a "high liquid level" state are filtered. Those meeting the filtering conditions are identified as candidate rainfall-inflow events. For the rainfall time series in the target monitoring area, effective rainfall events are divided based on two conditions: a cumulative rainfall of more than 2 mm per event and an interval of more than 12 hours between events. For the hidden state time series indicating a "high liquid level" state identified by the state recognition model, periods of continuous high liquid level with a repeating time of 3 hours or more are selected. Based on the effective rainfall periods and continuous high liquid level periods obtained in the above two steps, overlapping periods are first selected, and then the intersection of the effective rainfall periods and the continuous high liquid level periods is identified as the candidate rainfall-inflow events. The final candidate rainfall-inflow events obtained in this embodiment are as follows: Figure 3 As shown in c in the figure. By analyzing... Figure 3 As can be clearly seen from 'a' and 'b' in the diagram, the method proposed in this invention effectively eliminates fluctuations and does not identify temporary numerical increases caused by fluctuations as high-value states. When a continuous high-value state is identified within the past day, it further combines real-time rainfall data to identify the time period of inflow events caused by rainfall, and the identification effect is very close to that of manual identification.

[0117] Perform step S7 in Example 1 to obtain the historical frequency of inflow events, and obtain the following historical frequencies: dry season (November to March of the following year): approximately one significant rainwater inflow event occurs on average every 30 days; rainy season (June to September): approximately one event occurs on average every 10 days.

[0118] Based on the state recognition model, the frequency of candidate rainfall-inflow events is obtained from the recognition results. In the dry season, the average number of events is once every 5 days; in the rainy season, the average number of events is once every 2.3 days.

[0119] The frequency of the above candidate rainfall-inflow events deviates significantly from the historical frequency obtained based on human experience. The difference between the two is too large, indicating that there are too many misjudgments. It is initially speculated that the STL smoothing period is too small, causing too many short-term fluctuations to be identified as high liquid level states. It is necessary to adjust the seasonal period parameter period of the STL time series decomposition method and return to step S4. Repeat steps S4-S7 until the difference between the two is less than the preset value of 5%.

[0120] In this embodiment, a total of 3 loops are performed, that is, the seasonal period parameter is adjusted 3 times, and the 3 adjustments to the seasonal period parameter are as follows:

[0121] Round 1 adjustment: Increase the period to 48 hours, execute steps S4-S7, and the state recognition model outputs the recognition results to obtain the frequency of candidate rainfall-inflow events as 1 time / 10 days in the dry season and 1 time / 5 days in the rainy season;

[0122] Second round of adjustments: Increase the period to 72 hours, execute steps S4-S7, and the state recognition model outputs the recognition results to obtain the frequency of candidate rainfall-inflow events as 1 time / 25 days in the dry season and 1 time / 11 days in the rainy season;

[0123] Third round of fine-tuning: Increase the period to 96 hours, execute steps S4-S7, and the state recognition model outputs the recognition results to obtain the frequency of candidate rainfall-inflow events as 1 time / 29 days in the dry season and 1 time / 9.5 days in the rainy season.

[0124] The output of the state recognition model in the third cycle is less than 5% different from the historical frequency, indicating that the parameters of the state recognition model have converged and the semi-supervised calibration is complete.

[0125] The final status recognition model was permanently deployed in the real-time data flow monitoring system of the monitoring station for a period of 96 hours, and the recognition algorithm ran at a rolling frequency of 5 minutes. Subsequent six months of field testing showed that the status recognition model achieved an average recognition accuracy of 91.4%, significantly reducing false alarms and substantially lowering the cost of manual verification.

[0126] S8: Output the finally identified candidate rainfall-inflow events in real time in a structured record format. Specifically, this includes outputting the identified rainfall-inflow event time periods in JSON, CSV, or GeoJSON format, including the site number, event start and end times, duration of high liquid level, and corresponding rainfall information, for use by external information systems. The output results of this embodiment are as follows: Figure 4 As shown.

Claims

1. A semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks, characterized in that: Includes the following steps, S1: Obtain real-time and historical liquid level data of the target monitoring station, as well as real-time rainfall data of the area corresponding to the target monitoring station; S2: Perform outlier removal and missing value imputation preprocessing on real-time liquid level data and historical liquid level data respectively; S3: Resample the real-time rainfall data, preprocessed real-time liquid level data, and historical liquid level data to obtain rainfall sequences, real-time liquid level sequences, and historical liquid level sequences at a unified time scale; S4: The STL time series decomposition method is used to decompose the real-time liquid level series and the historical liquid level series respectively, and the trend term T of the real-time liquid level series is obtained respectively. t实 Trend term T of historical liquid level series t历 ; S5: Construct a Hidden Markov Model and utilize the trend term T t历 Using the time series as the observation sequence, the Baum-Welch expectation-maximization algorithm is used to train the hidden Markov model in unsupervised training to obtain a state recognition model for predicting the hidden state sequence of liquid level. S6: The trend term T of the real-time liquid level data... t实 Input the state recognition model, and the state recognition model outputs the liquid level hidden state sequence; filter the "high liquid level" state segments in the liquid level hidden state sequence, and identify the "high liquid level" state segments that meet the preset conditions as candidate rainfall-inflow events; S7: Set the time interval threshold for the state recognition model calibration. If the time interval threshold is not reached, proceed directly to the next step. When the time interval threshold is reached, calculate the occurrence frequency of the candidate rainfall-inflow event within a specified period and compare it with the difference between the frequency of human experience or the frequency of manually labeled samples. If the difference is less than the preset frequency error threshold, proceed to the next step. If the difference is greater than or equal to the frequency error threshold, increment the period parameter value in the STL time series decomposition method at equal intervals and repeat steps S4 to S7 until the difference is less than the preset frequency error threshold. S8: Output the finally identified candidate rainfall-inflow events in real time in a structured record format.

2. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 1, characterized in that: In step S2, the preprocessing method for real-time liquid level data and historical liquid level data is as follows: an adaptive 3σ anomaly detection algorithm based on sliding time window is used to identify outliers in the two types of liquid level data respectively. After removing the identified outliers, a linear interpolation method within the nearest time window is used to fill in the outliers and missing values.

3. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 2, characterized in that: In step S2, the adaptive 3σ anomaly detection algorithm based on a sliding time window is used to identify outliers in the liquid level data as follows: In the formula, LB represents the lower limit of the normal value of the liquid level data in the past week, UB represents the upper limit of the normal value of the liquid level data in the past week, μ represents the mean of the liquid level data of the target monitoring station in the past week, and σ represents the standard deviation of the liquid level data of the target monitoring station in the past week. If the monitored value of the liquid level data is lower than LB or higher than UB, it is judged as an abnormal value; The method for imputing outliers and missing values ​​using linear interpolation within a nearby time window is as follows: In the formula, S t D represents the missing or outlier values ​​in the liquid level data at time t. t S represents t timestamp; S t-1 Indicates the distance S in the liquid level data t The most recent normal value, D t-1 S represents t-1 timestamp; S t+1 Indicates the distance S in the liquid level data t The most recent normal value, D t+1 S represents t+1 Timestamp.

4. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 1, characterized in that: The resampling method in step S3 is the sliding window averaging method, and the sampling method is as follows: determine the sampling frequency, the initial period of the sampling time series and the end period of the sampling time series, and calculate the normalized time point series based on this; for each target sampling point, take the data within the range of T minutes in front of it as the sliding window, and calculate the average liquid level within the window, and use the average value as the resampling value of the target sampling point.

5. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 1, characterized in that: In step S4, the STL time series decomposition method, based on the local weighted regression algorithm, decomposes the two liquid level sequences into the following equations: In the formula, y t S represents liquid level sequence data. t T represents the seasonal term. t R represents the trend term. t This represents the residual term.

6. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 1, characterized in that: The steps for training a Hidden Markov Model are as follows: S51: The hidden state Gaussian HMM in the Hidden Markov Model is selected for training. The number of hidden states of the HMM is set to 2. Hidden state S1 represents the "normal fluctuation" state of the liquid level under normal operation, and hidden state S2 represents the "high liquid level" state of the liquid level under abnormal operation. The set of hidden states is S={S1,S2}. Using the trend term T from historical liquid level data t历 The liquid level monitoring value represents observation state O, and the set of observation states is... ; S52: Using the K-means clustering method to analyze the trend term T of historical liquid level data. t历 Perform clustering to divide the data into two clusters; Based on the clustering results, an initial observation probability distribution parameter set {μ1, σ1, μ2, σ2} is constructed, where μ1 is the observation mean of hidden state S1, σ1 is the observation standard deviation σ1 of hidden state S1, μ2 is the observation mean of hidden state S2, and σ2 is the observation standard deviation σ2 of hidden state S2. The initial state probability vector is obtained based on the clustering results, and the trend term T contained in the two data clusters is calculated respectively. t历 Quantity in the overall trend term T t历 The proportion π in the vector gives the initial state probability vector π = [p1, p2], where p1 refers to a certain trend term T. t历 The probability of belonging to S1, p2 refers to the trend term T. t历 The probability of belonging to S2, and p1+p2=1; S53: Construct the initial state transition probability matrix A. And set A as a high self-transformation probability structure; where a ij This represents the probability that when the hidden state is i at time t, the hidden state changes to j at time t+1. S54: Calculate the observed state probability matrix B. The observation probability density function in each hidden state follows a one-dimensional normal distribution. , Indicates that in the hidden state S i Below, the liquid level value o was observed at time t. N Observation probability density; μ i Hidden state S i The mean parameter, σ i Hidden state S i The standard deviation parameter; S55: The Baum-Welch expectation-maximization algorithm is used for unsupervised iterative training of the HMM, through maximization of the trend term T. t历 The log-likelihood function of the corresponding observation sequence is used to optimize the HMM model parameters; S56: Set the stopping condition for unsupervised iterative training: the log-likelihood increment is lower than the preset convergence threshold or the number of iterations reaches the preset upper limit; S57: After unsupervised iterative training stops, the resulting trained HMM is the intermediate model, and the intermediate model outputs the corresponding trend term T. t历 The hidden state time sequence of the observation sequence is obtained, and then the difference between the observation means of two hidden states is compared. If the difference is greater than or equal to a preset threshold, the intermediate model converges and is used as the state recognition model; if the difference is less than the preset threshold, the intermediate model is optimized and trained again until the difference is greater than or equal to the preset threshold.

7. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 6, characterized in that: The preset threshold in step S57 is set as follows: Let T be the trend term of the historical liquid level series. t历 The sequence of liquid level monitoring values ​​is {x1, x2, x3, ..., x Q }; Calculate {x1,x2,x3,…,x} Q The mean of} ; Based on mean Calculate {x1,x2,x3,…,x} Q The global standard deviation σ of} global and the global standard deviation σ global As a preset threshold , where q∈[1,Q].

8. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 6, characterized in that: The process of further optimizing the intermediate model training in step S57 is as follows: S571: If the difference between the observed means of two hidden states is less than a preset threshold, adjust the elements in the initial state transition probability matrix A; the automatic adjustment method is: increase The conversion probability is reduced. The conversion probability, keep The transition probability; S572: Obtain historical liquid level data under conditions of no heavy rainfall interference, and use the STL time series decomposition method to obtain the trend term of the historical liquid level data. Then, use the probability of this trend term being assigned to hidden states S1 and S2 to reset the initial state probability vector π. If the proportion of consecutive high liquid level segments in the training set exceeds 20%, increase p2 to the range of 0.1~0.2; S573: After adjusting A and π, return to step S55 to perform unsupervised iterative training on the HMM again until the difference between the observed mean values ​​of the two hidden states in step S57 is greater than or equal to the preset threshold. If the condition that the difference between the observed mean values ​​of the two hidden states is greater than or equal to the preset threshold is not met after 3 iterations, then the intermediate model obtained in the last iteration is used as the state recognition model.

9. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 1, characterized in that: The preset conditions in step S6 are: ① The duration of the hidden state time series of the "high liquid level" state is not less than T1 hours and the cumulative rainfall of a single rainfall is not less than R millimeters; ② The interval between the hidden state time series of the "high liquid level" state and the previous rainfall is not less than T2 hours; if conditions ① and ② are met at the same time, it is identified as a candidate rainfall-inflow event.

10. The semi-supervised iterative training method for identifying rainfall-inflow events in sewage pipe networks according to claim 1, characterized in that: In step S7, the seasonal period parameter (period) of the STL time series decomposition method is adjusted as follows: initially, the seasonal period parameter (period) is set to period0; during the nth iteration, the seasonal period parameter (period) is period0 + nΔ. period , where Δ period The increment step size for the periodic parameter.

Images