Water leakage monitoring method based on multi-domain feature collaborative perception

Abstract

The invention provides a water leakage monitoring method based on multi-domain feature collaborative perception, and belongs to the field of water leakage monitoring of a water purification system. The method comprises the steps of performing feature engineering processing on original sensor data to generate multi-domain features so as to construct a feature engineering system; training a long-short-term memory self-encoder model by utilizing water leakage time period data and non-water-leakage time period data which are marked manually, predicting the data of the water leakage time period through the long-short-term memory self-encoder model, and calculating a reconstruction error; the threshold value is dynamically adjusted by calculating the difference value between the real-time reconstruction error and the static threshold value, the threshold value offset is predicted according to parameters, and the threshold value adjustment strategy is refined; when the real-time reconstruction error exceeds a dynamic threshold value, water leakage early warning is triggered; and retraining the long and short-term memory auto-encoder model and the linear regression model by using latest data regularly. Through deep fusion of data driving and a physical mechanism, an accurate, intelligent and low-cost water leakage monitoring solution is provided for a water purification system.

Description

Technical Field

[0001] This invention belongs to the field of water purification system leakage monitoring, and particularly relates to a leakage monitoring method based on multi-domain feature collaborative sensing. Background Technology

[0002] The current mainstream technology for leak detection in water purification systems is based on physical protection devices, with typical solutions including: 1) Short-circuit detection method based on a leak protection plate: This method involves laying a protective plate with a conductive coating under the pipe. When a leak causes a short circuit in the coating, a power-off protection is triggered. This method has the following characteristics: Highly dependent on hardware: requires additional installation of conductive ground planes or sensor arrays; Low detection delay: It can respond within 1-3 seconds after a leak occurs, but it cannot provide early warning of potential risks; High false alarm rate: Affected by humidity and condensation, the false alarm rate can reach 12%-18%.

[0003] 2) Flow balance-based monitoring method: Leakage is determined by comparing the difference between the influent and effluent flow rates. Its limitations include: Sensitive to transient disturbances: Flow fluctuations (±15%) caused by water start-up and shutdown may trigger false alarms; Micro-leak detection blind zone: When the leakage amount is less than 5% of the rated flow, the detection success rate is less than 62%.

[0004] 3) Pressure decay-based detection method: Leakage is detected by monitoring the rate of decrease in static pressure in the pipeline. Problems: The testing cycle is long: the pressure needs to stabilize for 30-60 minutes before testing can be performed. Significant environmental interference: Temperature changes (±5℃) can cause pressure reference drift, affecting detection accuracy. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a multi-domain feature collaborative sensing method for leak detection, which solves the following core problems existing in current water purification system leak detection technologies: 1) Limitations in monitoring dimensions: existing physical protection devices (such as leak protection plates) and single-parameter monitoring schemes (such as flow threshold methods); 2) Insufficient dynamic adaptability: fixed thresholds or simple dynamic thresholds (such as moving averages) cannot adapt to operating condition disturbances such as water quality fluctuations and temperature changes, resulting in a persistently high false alarm rate.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a leakage detection method based on multi-domain feature collaborative sensing, comprising the following steps: Data collection and preprocessing: Collect raw sensor data and perform feature engineering to generate multi-domain features in order to build a feature engineering system; Model training and reconstruction error calculation: Based on the feature engineering system, a long short-term memory autoencoder model is trained using manually labeled leakage time period data and non-leakage time period data. The long short-term memory autoencoder model is then used to predict leakage time period data and calculate the reconstruction error. Dynamic threshold learning and adjustment: By calculating the difference between the real-time reconstruction error and the static threshold, the threshold is dynamically adjusted, and a linear regression model is used to predict the threshold offset based on the parameters, thus refining the threshold adjustment strategy for dynamic threshold learning and adjustment. Threshold comparison and early warning: The real-time reconstruction error is compared with the dynamic threshold. If the real-time reconstruction error exceeds the dynamic threshold, a leakage warning is triggered; otherwise, it is judged as a normal state. Model updates: The long short-term memory autoencoder model and the linear regression model are retrained regularly using the latest data to maintain the accuracy and adaptability of the models and complete the monitoring of water leaks.

[0007] Furthermore, the feature engineering system encompasses original physical quantity monitoring parameters, physical constraint features, time-domain statistical features, and frequency-domain features.

[0008] Furthermore, the raw physical quantity monitoring parameters include the source water flow rate. , water flow rate Ambient humidity Ambient temperature TDS of raw water TDS of purified water and water production time ; The physical constraint features include: Desalination rate : ; in, This indicates a smoothing term to prevent division by zero; Recovery rate : ; Flow difference : ; The time-domain statistical features include: Average flow : ; in, Indicates the length of the time window. t Indicates the end time of the sliding window. i This represents the index of all time points within the sliding window. This indicates the source water flow rate at each point in time within the window; Flow standard deviation : ; in, F Indicates flow rate; Recovery rate variance : ; in, R Indicates recovery rate. This represents the recycling rate at each point in time within the window. This represents the average recovery rate within the window; Humidity series : .

[0009] Furthermore, the calculation of the model training and reconstruction error includes the following steps: By manually annotating patterns to mark the time periods of leakage, the feature engineering system is divided into leakage time period data and non-leakage time period data. The long short-term memory autoencoder model was used to learn data from non-leaking time periods in order to train the long short-term memory autoencoder model. A trained long short-term memory autoencoder model is used to predict leakage time periods and calculate the reconstruction error.

[0010] Furthermore, the expression for the reconstruction error is as follows: ; in, Indicates reconstruction error, T This indicates the number of samples in the input sequence. W Indicates the length of the time window. Indicates the first i The sample at the th j The original feature vectors at each time step, Indicates the first i The sample at the th j Reconstructed feature vectors at each time step L2 norm operator; Reconstruction error in data for the period of leakage mean and standard deviation as follows: ; ; in, M This represents the number of samples in the dataset during the period of leakage. This represents the reconstruction error at each point in time within the window.

[0011] Furthermore, the dynamic threshold learning and adjustment includes the following steps: A static threshold is set based on the confidence level as a preliminary basis for judging the leakage situation; Considering the impact of changes in ambient temperature and humidity on the data, the threshold is dynamically adjusted by calculating the difference between the real-time reconstruction error and the static threshold. By using a linear regression model to predict threshold offsets based on parameters, the threshold adjustment strategy can be further refined to enable dynamic threshold learning and adjustment.

[0012] Furthermore, the expression for the static threshold is as follows: ; in, Indicates the static threshold. k This represents the confidence level parameter. Represents reconstruction error The mean, Represents reconstruction error Standard deviation; The expression for the threshold offset is as follows: ; in, Indicates the threshold offset. This indicates the reconstruction error.

[0013] Furthermore, the dynamic threshold is calculated as follows: ; in, Indicates a dynamic threshold. Indicates the static threshold. This represents the threshold predicted by the Long Short-Term Memory Autoencoder model using data from the time period of water leakage.

[0014] The beneficial effects of this invention are: (1) By constructing a multi-domain feature collaborative sensing system and a long short-term memory autoencoder model, the present invention achieves the following technical effects: 1) Multi-parameter fusion monitoring: integrate physical constraint features, time domain statistical features and frequency domain energy features to improve the detection sensitivity of small leaks (<0.5L / min); 2) Low-cost deployment: deploy inexpensive sensors such as TDS, flow rate and humidity, and combine them with software algorithms to achieve the desired results.

[0015] (2) Multidimensional feature engineering: This invention constructs a comprehensive feature engineering system by combining multiple features, covering original physical quantity monitoring parameters, physical constraint features, time-domain statistical features, and frequency-domain features. This multidimensional feature extraction method can more comprehensively reflect the leakage situation, improve the ability to identify leakage, and thus improve the accuracy of leakage detection.

[0016] (3) Dynamic threshold adjustment mechanism: The present invention adopts a dynamic threshold adjustment mechanism, which takes into account the influence of environmental factors (such as temperature and humidity) on the data. By dynamically adjusting the threshold, the system can better adapt to the leakage detection needs under different environments, and improve the accuracy and robustness of the early warning system.

[0017] (4) Linear regression model predicts threshold offset: This invention uses a linear regression model to predict threshold offset based on changes in environmental parameters, further refining the threshold adjustment strategy. This refined threshold adjustment method enables the system to detect water leakage more sensitively and improves the sensitivity of early warning.

[0018] (5) Regular model update: This invention proposes a strategy of regular model update. By regularly retraining the long short-term memory autoencoder model and the linear regression model with the latest data, this strategy ensures that the model can effectively detect leaks in the long term, extends the service life of the system, maintains the accuracy and adaptability of the model, and ensures the long-term effectiveness of the early warning system.

[0019] In summary, this invention significantly improves the accuracy of leak detection, the robustness of the system, the sensitivity of early warning, and the long-term effectiveness of the model through new technologies and measures such as multidimensional feature engineering, dynamic threshold adjustment mechanism, linear regression model prediction of threshold offset, and periodic model updates. This gives the invention significant advantages and broad application prospects in the field of leak detection, effectively solves the problems existing in the prior art, and improves the accuracy and robustness of leak early warning systems. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0022] Example like Figure 1 As shown, this invention provides a leakage detection method based on multi-domain feature collaborative sensing, the implementation method of which is as follows: Data collection and preprocessing: Collect raw sensor data and perform feature engineering to generate multi-domain features in order to build a feature engineering system; In this embodiment, raw sensor data is collected, including key parameters such as source water flow rate, purified water flow rate, ambient humidity, ambient temperature, source water TDS, purified water TDS, and water production time.

[0023] In this embodiment, feature engineering is performed on the original sensor data to generate 34-dimensional features. The 34-dimensional features cover 7-dimensional original features, 3-dimensional physical features (desalination rate, recovery rate, flow difference), 4-dimensional time-domain statistical features (flow mean, flow standard deviation, recovery rate variance, humidity sequence), and 20-dimensional frequency-domain features (extracted through fast Fourier transform), thereby constructing a comprehensive feature engineering system.

[0024] In this embodiment, the feature engineering system covers the original physical quantity monitoring parameters, physical constraint features, time-domain statistical features, and frequency-domain features.

[0025] In this embodiment, the raw physical quantity monitoring parameters include the source water flow rate. The unit is ml, net water flow rate. The unit is ml, and the ambient humidity is... The unit is %RH, ambient temperature. The unit is TDS of raw water The unit is ppm, and the TDS of purified water is... The units are ppm and water production time. ; The physical constraint features include: Desalination rate : ; in, This indicates a smoothing term to prevent division by zero; Recovery rate : ; Flow difference : ; In this embodiment, the time-domain statistical features are defined as follows: A sliding window statistic is defined (window size W=12, i.e., starting from the current time, 12 consecutive sampling points are traced back to form a time window). The time-domain statistical features include: Average flow : ; in, Indicates the length of the time window (window size) W =12, that is: taking the current time as the end point, backtracking 12 consecutive sampling points to form a time window). t Indicates the end time of the sliding window. i This represents the index of all time points within the sliding window. This indicates the source water flow rate at each point in time within the window.

[0026] Flow standard deviation : ; in, F Indicates flow rate; Recovery rate variance : ; in, R Indicates recovery rate. This represents the recycling rate at each point in time within the window. This represents the average recovery rate within the window; Humidity series : .

[0027] In this embodiment, the frequency domain features are obtained as follows: a Fast Fourier Transform (FFT) is performed on each time domain feature (FFT and amplitude spectrum calculation), and frequency domain energy is extracted. The frequency components corresponding to the top N=5 largest amplitudes are selected. ,in, N This indicates the number of frequency points selected. This represents the frequency index after the amplitude is sorted.

[0028] The Fast Fourier Transform is as follows: ; The amplitude spectrum is calculated as follows: ; The final matrix of multi-domain features: ; in, Indicates original features, Indicates physical constraint characteristics, Represents time-domain statistical characteristics. It represents the frequency domain features (obtained by fast Fourier transform of 4 statistical features, with the first 5 maximum amplitudes of each feature being taken).

[0029] Model training and reconstruction error calculation: Based on the feature engineering system, a long short-term memory autoencoder model is trained using manually labeled leakage time period data and non-leakage time period data. The model is then used to predict leakage time period data, and the reconstruction error is calculated. Specifically: By manually annotating patterns to mark the time periods of leakage, the feature engineering system is divided into leakage time period data and non-leakage time period data. The long short-term memory autoencoder model was used to learn data from non-leaking time periods in order to train the long short-term memory autoencoder model. A trained long short-term memory autoencoder model is used to predict leakage time periods and calculate the reconstruction error.

[0030] In this embodiment, an LSTM-AE (Long Short-Term Memory Autoencoder) model is trained using manually labeled leakage time period data and non-leakage time period data to learn the features of non-leakage time period (normal time period) data. The LSTM-AE model is then used to predict leakage time period data, and the reconstruction error is calculated. L This is used for subsequent threshold comparisons and early warnings.

[0031] In this embodiment, the LSTM-AE model is a general model, including an encoder, decoder, forget gate, input gate, and output gate.

[0032] In this embodiment, the LSTM-AE model is trained using data from non-leaking periods. After training, the model is then used to predict using data from leaking periods to identify the differences between the two datasets.

[0033] In this embodiment, the expression for the reconstruction error is as follows: ; in, Indicates reconstruction error, T Indicates the number of samples in the input sequence (integer). W Indicates the length of the time window (in hours, fixed at 12). Indicates the first i The sample at the th j The original feature vectors at each time step, Indicates the first i The sample at the th j Reconstructed feature vectors at each time step L2 norm operator; Reconstruction error in data for the period of leakage mean and standard deviation as follows: ; ; Where M represents the number of samples in the dataset for the time period of leakage. This represents the reconstruction error at each point in time within the window.

[0034] Dynamic threshold learning and adjustment: By calculating the difference between the real-time reconstruction error and the static threshold, the threshold is dynamically adjusted. A linear regression model is used to predict the threshold offset based on parameters, refining the threshold adjustment strategy for dynamic threshold learning and adjustment. The implementation method is as follows: A static threshold is set based on the confidence level as a preliminary basis for judging the leakage situation; Considering the impact of changes in ambient temperature and humidity on the data, the threshold is dynamically adjusted by calculating the difference between the real-time reconstruction error and the static threshold. Based on the predicted threshold offset from the parameters, the threshold adjustment strategy is further refined to enable dynamic threshold learning and adjustment.

[0035] In this embodiment, a static threshold is set based on confidence level as a preliminary basis for judging water leakage. The impact of changes in ambient temperature and humidity on the data is considered, and the real-time reconstruction error is calculated. L The threshold is dynamically adjusted based on the difference between the static threshold and the actual threshold to improve the accuracy and robustness of the early warning system. A linear regression model is used to predict the threshold offset based on parameters such as changes in flow rate, humidity, and temperature, further refining the threshold adjustment strategy and enhancing the sensitivity of the early warning system.

[0036] In this embodiment, the expression for the static threshold is as follows: ; in, Indicates the static threshold. k This represents the confidence level parameter. Represents reconstruction error The mean, Represents reconstruction error The standard deviation.

[0037] In this embodiment, the threshold is dynamically adjusted because the pipeline experiences thermal expansion and contraction under different ambient temperatures, which may cause changes in flow readings. Simultaneously, changes in humidity may also cause slight fluctuations in the overall water circuit data. Based on real-time reconstruction error... L With static threshold The difference is used to calculate the threshold offset. This allows for dynamic adjustment of the threshold. The expression for the threshold offset is as follows: ; in, Indicates the threshold offset. This indicates the reconstruction error.

[0038] If real-time reconstruction error L Less than the static threshold ,but =0 indicates that no threshold adjustment is needed. but This indicates that the threshold needs to be increased. To cover the new normal range.

[0039] In this embodiment, a linear regression model is used to train the changes in each parameter and predict the threshold offset. ,include: Flow change : ; ; Where M represents the number of samples in the dataset, This indicates the current traffic value. Represents the traffic in the dataset. This represents the average flow rate within the window.

[0040] Humidity change : ; ; in, This indicates the current humidity value. This represents the average humidity level within a window. For the humidity values ​​in the dataset, i This represents the index of all time points within the sliding window (M is 12, there are 12 points in a window, and each i represents the data of a point).

[0041] Temperature change : ; ; in, This is the current temperature value. This represents the average temperature of the window. These are the temperature values ​​in the dataset.

[0042] In this embodiment, the input feature for predicting the threshold offset is: The target value is (Threshold offset).

[0043] In this embodiment, the expression for the linear regression model is as follows: ; in, This represents the threshold predicted by the Long Short-Term Memory Autoencoder model using data from the leakage period. , and All represent weights.

[0044] Optimize model parameters by minimizing mean squared error (MSE) , The initial value is set randomly and then gradually adjusted using the gradient descent optimization algorithm.

[0045] In this embodiment, the dynamic threshold is calculated as follows: ; in, Indicates dynamic threshold. Indicates the static threshold. This represents the threshold predicted by the Long Short-Term Memory Autoencoder model using data from the time period of water leakage.

[0046] Threshold comparison and early warning: The real-time reconstruction error is compared with the dynamic threshold. If the real-time reconstruction error exceeds the dynamic threshold, a leakage warning is triggered; otherwise, it is judged as a normal state. In this embodiment, the real-time reconstruction error of LSTM-AE is... L With dynamic threshold Comparison: If If a leak warning is triggered, If so, it is considered normal.

[0047] Model updates: The long short-term memory autoencoder model and the linear regression model are retrained regularly using the latest data to maintain the accuracy and adaptability of the models and complete the monitoring of water leaks.

[0048] In this embodiment, the LSTM-AE and linear regression models are retrained periodically (e.g., weekly) with the latest data to maintain the accuracy and adaptability of the models and ensure the long-term effectiveness of the early warning system.

Claims

1. A leakage monitoring method based on multi-domain feature collaborative sensing, characterized in that, Includes the following steps: Data collection and preprocessing: Collect raw sensor data and perform feature engineering to generate multi-domain features in order to build a feature engineering system; Model training and reconstruction error calculation: Based on the feature engineering system, a long short-term memory autoencoder model is trained using manually labeled leakage time period data and non-leakage time period data. The long short-term memory autoencoder model is then used to predict leakage time period data and calculate the reconstruction error. Dynamic threshold learning and adjustment: By calculating the difference between the real-time reconstruction error and the static threshold, the threshold is dynamically adjusted, and a linear regression model is used to predict the threshold offset based on the parameters, thus refining the threshold adjustment strategy for dynamic threshold learning and adjustment. Threshold comparison and early warning: The real-time reconstruction error is compared with the dynamic threshold. If the real-time reconstruction error exceeds the dynamic threshold, a leakage warning is triggered; otherwise, it is judged as a normal state. Model updates: The long short-term memory autoencoder model and the linear regression model are retrained regularly using the latest data to maintain the accuracy and adaptability of the models and complete the monitoring of water leaks.

2. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 1, characterized in that, The feature engineering system covers original physical quantity monitoring parameters, physical constraint features, time-domain statistical features, and frequency-domain features.

3. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 2, characterized in that, The original physical quantity monitoring parameters include the source water flow rate. , water flow rate Ambient humidity Ambient temperature TDS of raw water TDS of purified water and water production time ; The physical constraint features include: Desalination rate : ； in, This indicates a smoothing term to prevent division by zero. Recovery rate : ； Flow difference : ； The time-domain statistical features include: Average flow : ； in, Indicates the length of the time window. t Indicates the end time of the sliding window. i This represents the index of all time points within the sliding window. This indicates the source water flow rate at each point in time within the window; Flow standard deviation : ； in, F Indicates flow rate; Recovery rate variance : ； in, R Indicates recovery rate. This represents the recycling rate at each point in time within the window. This represents the average recovery rate within the window; Humidity ranking : 。 4. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 3, characterized in that, The matrix expression for the generated multi-domain features is as follows: ； in, Indicates original features, Indicates physical constraint characteristics, Represents time-domain statistical characteristics. It represents the frequency domain characteristics.

5. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 1, characterized in that, The calculation of model training and reconstruction error includes the following steps: By manually annotating patterns to mark the time periods of leakage, the feature engineering system is divided into leakage time period data and non-leakage time period data. The long short-term memory autoencoder model was used to learn data from non-leaking time periods in order to train the long short-term memory autoencoder model. A trained long short-term memory autoencoder model is used to predict leakage time periods and calculate the reconstruction error.

6. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 5, characterized in that, The expression for the reconstruction error is as follows: ； in, Indicates the reconstruction error. T This indicates the number of samples in the input sequence. W Indicates the length of the time window. Indicates the first i The sample at the th j The original feature vectors at each time step, Indicates the first i The sample at the th j Reconstructed feature vectors at each time step L2 norm operator; Reconstruction error in data for the period of leakage mean and standard deviation as follows: ； ； in, M This indicates the number of samples in the dataset during the period of leakage. This represents the reconstruction error at each point in time within the window.

7. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 6, characterized in that, The dynamic threshold learning and adjustment includes the following steps: A static threshold is set based on the confidence level as a preliminary basis for judging the leakage situation; Considering the impact of changes in ambient temperature and humidity on the data, the threshold is dynamically adjusted by calculating the difference between the real-time reconstruction error and the static threshold. By using a linear regression model to predict threshold offsets based on parameters, the threshold adjustment strategy can be further refined to enable dynamic threshold learning and adjustment.

8. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 7, characterized in that, The expression for the static threshold is as follows: ； in, Indicates the static threshold. k This represents the confidence level parameter. Represents reconstruction error The mean, Represents reconstruction error Standard deviation; The expression for the threshold offset is as follows: ； in, Indicates the threshold offset. This indicates the reconstruction error.

9. The leakage monitoring method based on multi-domain feature collaborative sensing according to claim 7, characterized in that, The dynamic threshold is calculated as follows: ； in, Indicates dynamic threshold. Indicates the static threshold. This represents the threshold predicted by the Long Short-Term Memory Autoencoder model using data from the time period of water leakage.

Images