Anomaly detection based adaptive water dosing device and system

Abstract

This invention discloses an adaptive quantitative water supply device and system based on anomaly detection, comprising: a data acquisition module for acquiring and preprocessing internal data of the adaptive quantitative water supply device; a fluid-structure coupling modeling module for constructing the fluid-structure coupling relationship of a standard operating state sequence; a residual fingerprint construction module for constructing a residual fingerprint representation; a fluid-structure coupling autoencoder training module for training a fluid-structure coupling autoencoder; an anomaly detection module for generating anomaly judgment results and anomaly intensity sequences; a defense strategy generation module for generating defense strategies; and a water supply defense control execution module for executing corresponding flow interception, flow restriction, or pressure compensation control actions. This invention achieves adaptive defense control of quantitative water supply anomalies by constructing a fluid-structure coupling residual fingerprint and introducing constrained anomaly detection, possessing advantages such as accurate identification, stable response, and high operational safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent fluid control technology, and in particular to an adaptive quantitative water supply device and system based on anomaly detection. Background Technology

[0002] Existing quantitative water supply devices typically rely on mechanical structures in conjunction with changes in water supply pressure to achieve quantitative control, and monitor the water supply process through pressure sensors, displacement sensors, or flow meters. In actual operation, most water supply systems only collect data on a single operating parameter or a small number of state variables, and use threshold judgment, rule logic, or empirical models to identify abnormal states. This makes it difficult to reflect the coupling relationship between water supply pressure, piston displacement, elastic element deformation, and the vacuum state of the siphon cavity, resulting in limited ability to perceive abnormal evolution processes under complex operating conditions.

[0003] Meanwhile, even when data-driven anomaly detection methods are introduced in existing technologies, they mostly use raw sensor data as the analysis object, lacking prediction comparisons and residual structure modeling based on physical mechanisms, making it difficult to distinguish between normal fluctuations and structural anomalies. Moreover, the anomaly identification results are usually only used for alarm or shutdown processing, failing to form a closed-loop linkage with the water supply control process, and unable to dynamically generate defensive control strategies such as flow interception, flow restriction, or pressure compensation according to the degree of anomaly, thus affecting the operational safety and water supply stability of quantitative water supply devices under abnormal operating conditions. Summary of the Invention

[0004] One objective of this invention is to propose an adaptive quantitative water supply device and system based on anomaly detection. This invention achieves adaptive defense control of quantitative water supply anomalies by constructing a fluid-structure coupled residual fingerprint and introducing constrained anomaly detection, and has the advantages of accurate identification, stable response and high operational safety.

[0005] An adaptive quantitative water supply system based on anomaly detection according to an embodiment of the present invention includes:

[0006] The data acquisition module is used to synchronously acquire and preprocess the water supply pressure, piston displacement, elastic element deformation and siphon cavity vacuum status inside the adaptive quantitative water supply device, and generate a standard operating status sequence.

[0007] The fluid-structure coupling modeling module is used to construct the fluid-structure coupling relationship of the standard operating state sequence under normal operating conditions of the water supply device, and generate a set of predicted response sequences.

[0008] The residual fingerprint construction module is used to compare the predicted response sequence set with the actual collected response data during water supply operation, calculate the residual sequence between the predicted response and the actual response, and construct the residual fingerprint representation.

[0009] A fluid-structure coupled autoencoder training module is used to train a fluid-structure coupled autoencoder based on residual fingerprint representation.

[0010] The anomaly detection module is used during the online operation phase to input the real-time residual fingerprint representation into the trained fluid-structure coupled autoencoder, generate anomaly judgment results based on the residual reconstruction sequence, and generate anomaly intensity sequences.

[0011] The defense strategy generation module is used to generate defense strategies based on the abnormal strength sequence;

[0012] The water supply defense control execution module is used to drive the water supply device to perform corresponding flow interception, flow restriction or pressure compensation control actions based on the defense strategy, and maintain the corresponding control state according to the action duration parameter to realize abnormal water supply defense control.

[0013] Optionally, the vacuum state of the siphon cavity is a pressure state parameter inside the siphon cavity used to maintain the siphon effect, which characterizes the degree of negative pressure formed inside the siphon cavity relative to the external atmospheric pressure. The preprocessing includes time alignment, dimension normalization and noise suppression.

[0014] Optionally, the fluid-structure coupling modeling module includes:

[0015] Obtain the standard operating state sequence formed by the adaptive quantitative water supply device under normal operating conditions;

[0016] For each time sampling point in the standard operating state sequence, the water supply pressure and siphon cavity vacuum state at the corresponding time are extracted as input state vectors characterizing the fluid action state, and the piston displacement and elastic element deformation at the corresponding time are extracted as output state vectors characterizing the structural response state.

[0017] The input state vector and output state vector corresponding to each time sampling point are paired and organized to form a time-coupled sample set;

[0018] Based on the temporally coupled sample set, the influence of changes in water supply pressure and vacuum state of the siphon cavity on piston displacement and deformation of elastic elements is modeled and processed to construct a fluid-structure coupling relationship.

[0019] By inputting the water supply pressure and the vacuum state of the siphon cavity from the standard operating state sequence into the fluid-structure coupling relationship, the corresponding predicted displacement response sequence and predicted deformation response sequence are obtained.

[0020] The predicted displacement response sequence and the predicted deformation response sequence are arranged and stored in chronological order to form a set of predicted response sequences.

[0021] Optionally, the construction of the fluid-structure coupling relationship includes:

[0022] For each time sampling point in the time-series coupled sample set, the corresponding water supply pressure and siphon cavity vacuum state are extracted to form a fluid action state vector. At the same time, the piston displacement and elastic element deformation at the corresponding time sampling point are extracted to form a structural response state vector.

[0023] According to the time sampling order, the fluid action state vector and the structural response state vector corresponding to each time sampling point are paired point by point to form a mapping sample set;

[0024] Based on the mapped sample set, differential processing is performed on the water supply pressure, siphon cavity vacuum state, piston displacement and elastic element deformation of adjacent sampling points according to the time sampling order to obtain the water supply pressure change sequence, vacuum state change sequence, piston displacement change sequence and elastic element deformation change sequence.

[0025] Calculate the pressure-displacement response coefficient for the change in water supply pressure and piston displacement in each time period, and calculate the vacuum-deformation response coefficient for the change in vacuum state and deformation of elastic element in each time period.

[0026] Calculate the pressure-deformation coupling coefficient for the change in water pressure and the change in deformation of the elastic element in each time period, and calculate the vacuum-displacement coupling coefficient for the change in vacuum state and the change in piston displacement in each time period.

[0027] The pressure-displacement response coefficient value, vacuum-deformation response coefficient value, pressure-deformation coupling coefficient value, and vacuum-displacement coupling coefficient value are combined according to the time period index to obtain the coupling parameter set;

[0028] Perform statistical summarization processing on the set of coupling parameters to generate a coupling feature parameter vector, which is then associated and stored with the corresponding fluid action state vector and structural response state vector to form a fluid-structure coupling relationship.

[0029] Optionally, the residual fingerprint construction module includes:

[0030] During water supply operation, the water supply pressure and the vacuum state of the siphon cavity are collected in real time, and a set of corresponding predicted response sequences is generated based on the fluid-structure coupling relationship.

[0031] The predicted displacement response and predicted deformation response in the set of piston displacement, elastic element deformation and predicted response sequences collected at the same sampling point are aligned point by point. The displacement residual value between the actual displacement and the predicted displacement, and the deformation residual value between the actual deformation and the predicted deformation are calculated respectively. The displacement residual sequence and the deformation residual sequence are formed in time order.

[0032] Within a preset sliding time window, temporal abrupt structure extraction processing is performed on the displacement residual sequence and the deformation residual sequence respectively. The residual change amount, change direction identifier and change duration of adjacent sampling points are calculated to generate a temporal abrupt feature sequence.

[0033] Within the same time window, frequency domain transformation is performed on the displacement residual sequence and the deformation residual sequence respectively to extract the energy distribution value, dominant frequency position and frequency band energy concentration of the residuals, forming a frequency domain energy distribution feature sequence.

[0034] Perform time-frequency joint analysis on the displacement residual sequence and the deformation residual sequence, calculate the local concentration of residual energy on the time-frequency plane, extract the index of residual energy surge interval and the corresponding time-frequency energy intensity value, and generate a time-frequency evolution feature sequence;

[0035] Statistical structure modeling is performed on the displacement residual sequence and deformation residual sequence within the corresponding time window. The mean and variance of the residual sequence are calculated to generate a statistical structure feature sequence.

[0036] The residual fingerprint unit is constructed by jointly organizing the time-domain abrupt change feature sequence, the frequency-domain energy distribution feature sequence, the time-frequency evolution feature sequence, and the statistical structure feature sequence. The residual fingerprint units are then combined in chronological order to form a residual fingerprint representation.

[0037] Optionally, the fluid-structure coupled autoencoder training module includes:

[0038] At each time window index position, the residual fingerprint unit corresponding to the time window is extracted from the residual fingerprint representation, combined to form a residual fingerprint sequence, and the vacuum state value sequence is extracted. The residual fingerprint sequence and the vacuum state value sequence are aligned and normalized to generate a standard residual fingerprint sequence and a standard vacuum state value sequence.

[0039] The standard residual fingerprint sequence and the standard vacuum state value sequence are input into the fluid-structure coupled autoencoder;

[0040] The fluid-structure coupled autoencoder includes a residual fingerprint encoding layer, a vacuum state encoding layer, a coupling latent constraint layer, and a reverse mapping decoding layer;

[0041] In the residual fingerprint coding layer, temporal coding is performed on the standard residual fingerprint sequence to obtain the residual latent state sequence; in the vacuum state coding layer, temporal coding is performed on the standard vacuum state value sequence to obtain the vacuum latent state sequence.

[0042] In the coupled latent constraint layer, the vacuum latent state sequence is used as the constraint reference, and the latent space consistency constraint is applied to the residual latent state sequence to generate the coupled latent state sequence.

[0043] In the reverse mapping decoding layer, a predictive reconstruction reverse mapping is performed based on the coupled latent state sequence to generate a residual reconstruction sequence;

[0044] The parameters of the fluid-structure coupled autoencoder are iteratively updated until the residual reconstruction sequence converges or the training is completed by reaching the maximum number of training iterations.

[0045] Optionally, the anomaly detection module includes:

[0046] During the online operation phase, real-time residual fingerprint representations are acquired;

[0047] The real-time residual fingerprint representation is input into the trained fluid-structure coupled autoencoder to obtain the residual reconstruction sequence;

[0048] The residual reconstruction sequence is compared with the normal operation reconstruction error reference range to obtain the anomaly judgment result, which is used to characterize whether there is an abnormal state in the current time window.

[0049] Within the time window where an abnormal state is identified, an abnormality intensity value is generated based on the degree of deviation of the reconstruction error value corresponding to that time window from the normal operation reconstruction error reference range.

[0050] The anomaly determination results and anomaly intensity values ​​generated in each time window are combined in the order of the time window index to form an anomaly determination sequence and an anomaly intensity sequence.

[0051] Optionally, the normal operation reconstruction error reference range is obtained by statistically summarizing the reconstruction error values ​​corresponding to each time window under historical normal operation conditions, extracting their stable distribution interval, and using it as a reference range for anomaly judgment during the online operation phase.

[0052] Optionally, the defense strategy generation module includes:

[0053] Using the abnormal intensity sequence as the control input, a hierarchical mapping is performed according to the abnormal intensity score corresponding to each time window in the abnormal intensity sequence, and the defense control parameters under the corresponding time window are determined respectively.

[0054] The defense control parameters are combined in chronological order to form a defense strategy, which includes flow throttling parameters, flow limiting intensity parameters, pressure compensation parameters, and action duration parameters.

[0055] An adaptive quantitative water supply device based on anomaly detection according to an embodiment of the present invention includes a plug body, a displacement sensor, a support frame, and a water outlet pipe. The water outlet pipe is connected to the upper end of the plug body. A water supply channel is formed inside the plug body. A piston, a spring, and a cone are arranged axially inside the water supply channel. The cone is positioned above the piston and corresponds to the inlet position of the water outlet pipe. The piston moves axially along the plug body under the action of water supply pressure. The spring is positioned between the piston and the plug body. The cone changes its relative distance to the inlet of the water outlet pipe as the piston moves axially. A sealing plate is provided at the upper end of the plug body. The sealing plate is connected to the plug body through a sealing screw. The support frame supports the plug body. A cavity structure is provided inside the plug body and is connected to the water supply channel. The displacement sensor is installed inside the plug body and is connected to a metering and storage device through a signal line.

[0056] The beneficial effects of this invention are:

[0057] This invention synchronously collects and uniformly preprocesses multi-source operating parameters such as water supply pressure, piston displacement, elastic element deformation, and siphon cavity vacuum state. Under normal operating conditions, it establishes a fluid-structure coupling relationship between the fluid action state and the structural response state, and generates a set of predicted response sequences based on this coupling relationship. This enables the water supply device to predict its own structural response behavior during operation. By comparing the predicted response sequences with the actual collected response data, a residual sequence reflecting the degree of deviation from the water supply device's operation is formed. Furthermore, a residual fingerprint representation is constructed, incorporating time-domain abrupt change characteristics, frequency-domain energy distribution characteristics, time-frequency evolution characteristics, and statistical structure characteristics. This ensures that anomalies are no longer merely instantaneous fluctuations of a single parameter, but are systematically characterized as changes in the residual structure morphology, thereby improving the accuracy and stability of identifying potential abnormal states under complex operating conditions.

[0058] Building upon this foundation, this invention introduces a fluid-structure coupled autoencoder to model the residual fingerprint representation. By introducing latent constraints corresponding to the vacuum state of the siphon cavity into the latent space, the evolution of the residual latent state is made consistent with the actual physical operating state, thereby avoiding the misjudgment problems that may arise from anomaly detection methods that rely solely on data distribution. This structure enables the anomaly determination result to not only reflect the magnitude of the residual reconstruction error but also the degree of deviation of the anomaly state from the normal physical operating law, thus generating an anomaly intensity sequence with continuous quantitative significance, providing a reliable basis for subsequent control decisions.

[0059] Furthermore, this invention directly incorporates the anomaly detection results into the water supply control process. Based on the anomaly intensity sequence, it generates a defense strategy that includes flow throttling opening, flow limiting intensity, pressure compensation, and action duration, and drives the water supply device to execute the corresponding defense control actions. This enables the system to have adaptive adjustment capabilities under abnormal conditions, achieving closed-loop linkage between anomaly identification and water supply control. Compared to existing technologies that only focus on anomaly alarms or simple protection, this invention can reduce the impact of abnormal conditions on the device structure and water supply stability while ensuring the continuity of quantitative water supply function, thereby improving the operational safety, reliability, and intelligence level of the quantitative water supply device under complex operating conditions. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a schematic diagram of the structure of an adaptive quantitative water supply system based on anomaly detection proposed in this invention;

[0062] Figure 2 This is a schematic diagram of the fluid-structure coupled autoencoder for an adaptive quantitative water supply system based on anomaly detection proposed in this invention.

[0063] Figure 3 This is a front sectional view of an adaptive quantitative water supply device based on anomaly detection proposed in this invention.

[0064] Reference numerals in the attached diagram: 1. Meter and storage device; 2. Signal line; 3. Displacement sensor; 4. Sealing screw; 5. Sealing plate; 6. Water outlet pipe; 7. Conical body; 8. Piston; 9. Spring; 10. Support frame; 11. Bolt. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figures 1-2 An adaptive quantitative water supply system based on anomaly detection includes:

[0067] The data acquisition module is used to synchronously acquire and preprocess the water supply pressure, piston displacement, elastic element deformation and siphon cavity vacuum status inside the adaptive quantitative water supply device, and generate a standard operating status sequence.

[0068] The fluid-structure coupling modeling module is used to construct the fluid-structure coupling relationship of the standard operating state sequence under normal operating conditions of the water supply device, and generate a set of predicted response sequences.

[0069] The residual fingerprint construction module is used to compare the predicted response sequence set with the actual collected response data during water supply operation, calculate the residual sequence between the predicted response and the actual response, and construct the residual fingerprint representation.

[0070] A fluid-structure coupled autoencoder training module is used to train a fluid-structure coupled autoencoder based on residual fingerprint representation.

[0071] The anomaly detection module is used during the online operation phase to input the real-time residual fingerprint representation into the trained fluid-structure coupled autoencoder, generate anomaly judgment results based on the residual reconstruction sequence, and generate anomaly intensity sequences.

[0072] The defense strategy generation module is used to generate defense strategies based on the abnormal strength sequence;

[0073] The water supply defense control execution module is used to drive the water supply device to perform corresponding flow interception, flow restriction or pressure compensation control actions based on the defense strategy, and maintain the corresponding control state according to the action duration parameter to realize abnormal water supply defense control.

[0074] In this embodiment, the vacuum state of the siphon cavity is a gas pressure state parameter inside the siphon cavity used to maintain the siphon effect, which characterizes the degree of negative pressure formed inside the siphon cavity relative to the external atmospheric pressure. The preprocessing includes time alignment, dimension normalization and noise suppression.

[0075] In this embodiment, the fluid-structure coupling modeling module includes:

[0076] Obtain the standard operating state sequence formed by the adaptive quantitative water supply device under normal operating conditions;

[0077] For each time sampling point in the standard operating state sequence, the water supply pressure and siphon cavity vacuum state at the corresponding time are extracted as input state vectors characterizing the fluid action state, and the piston displacement and elastic element deformation at the corresponding time are extracted as output state vectors characterizing the structural response state.

[0078] The input state vector and output state vector corresponding to each time sampling point are paired and organized to form a time-coupled sample set, which reflects the correspondence between the fluid action state and the structural response state.

[0079] Based on the time-series coupled sample set, the influence of changes in water supply pressure and vacuum state of siphon cavity on piston displacement and elastic element deformation is modeled and processed to construct fluid-structure coupling relationship and describe the mapping relationship between fluid action state and structural response state under normal operating conditions.

[0080] By inputting the water supply pressure and the vacuum state of the siphon cavity from the standard operating state sequence into the fluid-structure coupling relationship, the corresponding predicted displacement response sequence and predicted deformation response sequence are obtained.

[0081] The predicted displacement response sequence and the predicted deformation response sequence are arranged and stored in chronological order to form a set of predicted response sequences.

[0082] In this embodiment, the fluid-structure coupling relationship is constructed, including:

[0083] For each time sampling point in the time-series coupled sample set, the corresponding water supply pressure and siphon cavity vacuum state are extracted to form a fluid action state vector. At the same time, the piston displacement and elastic element deformation at the corresponding time sampling point are extracted to form a structural response state vector.

[0084] According to the time sampling order, the fluid action state vector and the structural response state vector corresponding to each time sampling point are paired point by point to form a mapping sample set;

[0085] Based on the mapped sample set, differential processing is performed on the water supply pressure, siphon cavity vacuum state, piston displacement and elastic element deformation of adjacent sampling points according to the time sampling order to obtain the water supply pressure change sequence, vacuum state change sequence, piston displacement change sequence and elastic element deformation change sequence.

[0086] Calculate the pressure-displacement response coefficient for the change in water supply pressure and piston displacement in each time period, and calculate the vacuum-deformation response coefficient for the change in vacuum state and deformation of elastic element in each time period.

[0087] Calculate the pressure-deformation coupling coefficient for the change in water pressure and the change in deformation of the elastic element in each time period, and calculate the vacuum-displacement coupling coefficient for the change in vacuum state and the change in piston displacement in each time period.

[0088] The pressure-displacement response coefficient, vacuum-deformation response coefficient, pressure-deformation coupling coefficient, and vacuum-displacement coupling coefficient are combined according to the time period index to obtain a set of coupling parameters, which is used to characterize the degree of combined influence of water supply pressure and siphon cavity vacuum state on piston displacement and elastic element deformation.

[0089] Statistical summarization processing is performed on the set of coupling parameters to generate a vector of coupling characteristic parameters, which is associated with and stored with the corresponding fluid action state vector and structural response state vector to form a fluid-structure coupling relationship. This is used to input the water supply pressure and the vacuum state of the siphon cavity and output the predicted displacement response and predicted deformation response.

[0090] In the time-coupled sample set, differential processing is performed on the water pressure of adjacent sampling points according to the time sampling order to obtain a sequence of water pressure change. Differential processing is also performed on the piston displacement within the same time period to obtain a sequence of piston displacement change. The water pressure change sequence and the piston displacement change sequence are mapped one-to-one by time, and the ratio of piston displacement change to water pressure change in each corresponding time period is calculated to form a pressure-displacement response coefficient value sequence. The pressure-displacement response coefficient value sequences are then statistically summarized to obtain the pressure-displacement response coefficient. Similarly, following the same time sampling order, differential processing is performed on the vacuum state of the siphon cavity at adjacent sampling points to obtain a sequence of vacuum state change. Differential processing is also performed on the deformation of the elastic element within the same time period to obtain a deformation change sequence. The vacuum state change sequence and the deformation change sequence are mapped one-to-one by time, and the ratio of elastic element deformation change to vacuum state change in each corresponding time period is calculated to form a vacuum-deformation response coefficient value sequence. The vacuum-deformation response coefficient value sequences are then statistically summarized to obtain the vacuum-deformation response coefficient. Furthermore, sampling at the same time... At the sampling points, the sequence of water pressure changes and the sequence of elastic element deformation changes are correlated one-to-one over time. The ratio of the elastic element deformation change to the water pressure change in each corresponding time period is calculated to form a pressure-deformation coupling coefficient value sequence. The pressure-deformation coupling coefficient value sequence is then statistically summarized to obtain the pressure-deformation coupling coefficient. Simultaneously, at the same sampling points at the same time, the sequence of siphon cavity vacuum state changes and the sequence of piston displacement changes are correlated one-to-one over time. The ratio of the piston displacement change to the vacuum state change in each corresponding time period is calculated to form a vacuum... - A sequence of displacement coupling coefficient values ​​is obtained, and the vacuum-displacement coupling coefficient value sequence is statistically summarized to obtain the vacuum-displacement coupling coefficient. The pressure-displacement response coefficient, vacuum-deformation response coefficient, pressure-deformation coupling coefficient, and vacuum-displacement coupling coefficient are combined to generate a set of fluid-structure coupling parameters to characterize the combined influence of feedwater pressure and siphon cavity vacuum state on piston displacement and elastic element deformation. Based on this, a fluid-structure coupling relationship is constructed to output the predicted displacement response and predicted deformation response when the feedwater pressure and siphon cavity vacuum state are input.

[0091] In this embodiment, the residual fingerprint construction module includes:

[0092] During water supply operation, the water supply pressure and the vacuum state of the siphon cavity are collected in real time, and a set of corresponding predicted response sequences is generated based on the fluid-structure coupling relationship.

[0093] The predicted displacement response and predicted deformation response in the set of piston displacement, elastic element deformation and predicted response sequences collected at the same sampling point are aligned point by point. The displacement residual value between the actual displacement and the predicted displacement, and the deformation residual value between the actual deformation and the predicted deformation are calculated respectively. The displacement residual sequence and the deformation residual sequence are formed in time order.

[0094] Within a preset sliding time window, temporal abrupt structure extraction processing is performed on the displacement residual sequence and the deformation residual sequence respectively. The residual change amount, change direction identifier and change duration of adjacent sampling points are calculated to generate a temporal abrupt feature sequence, which characterizes the location and intensity of residual abrupt changes.

[0095] Temporal mutation structure extraction is a process in which the residual values ​​of adjacent time sampling points in the residual sequence are differentially calculated according to the time sampling order to extract the magnitude change, direction of change, and duration of change of the residual. This process is used to characterize the mutation location and intensity of the residual in the time dimension. The duration of change is the number of time sampling points in the residual sequence that continuously maintain the same direction of change from the time sampling point where the direction of change of the residual first changes to the time sampling point where the direction of change of the residual changes again.

[0096] Within the same time window, frequency domain transformation is performed on the displacement residual sequence and the deformation residual sequence respectively to extract the energy distribution value, dominant frequency position and frequency band energy concentration of the residuals, forming a frequency domain energy distribution feature sequence.

[0097] Perform time-frequency joint analysis on the displacement residual sequence and the deformation residual sequence, calculate the local concentration of residual energy on the time-frequency plane, extract the index of residual energy surge interval and the corresponding time-frequency energy intensity value, and generate a time-frequency evolution feature sequence;

[0098] The generation of the time-frequency evolution feature sequence includes: segmenting the displacement residual sequence and deformation residual sequence according to the time sampling order to obtain residual subsequences within the corresponding time window; performing time-frequency transformation on each residual subsequence to obtain the set of energy distribution values ​​of the residual at different frequency positions within the time window; statistically summarizing the residual energy distribution values ​​corresponding to each frequency position within the same time window, calculating the local concentration degree of residual energy on the time-frequency plane, and obtaining the time-frequency energy concentration value; comparing the time-frequency energy concentration values ​​in adjacent time windows in chronological order to identify continuous time windows where the time-frequency energy concentration value changes significantly, and determining the index of the residual energy surge interval; statistically analyzing the residual energy distribution values ​​of the corresponding time window within each residual energy surge interval, and calculating the time-frequency energy intensity value within that interval; arranging the residual energy surge interval index and the corresponding time-frequency energy intensity value in chronological order to generate a time-frequency evolution feature sequence for characterizing the joint evolution characteristics of residual energy with time and frequency.

[0099] Within a preset time window, time-frequency transformation is performed on the displacement residual sequence and the deformation residual sequence to obtain the residual energy value corresponding to each frequency position within the time window. Statistical processing is performed on the set of residual energy values ​​within the same time window to calculate the proportion of energy distribution concentrated at the dominant frequency position, generating the time-frequency energy concentration value corresponding to the time window. After determining the index of the residual energy surge interval, the set of residual energy values ​​corresponding to each time window within the interval is extracted, statistically summarized, and the cumulative intensity of residual energy within the surge interval is calculated to generate the time-frequency energy intensity value corresponding to the residual energy surge interval.

[0100] Statistical structure modeling is performed on the displacement residual sequence and deformation residual sequence within the corresponding time window. The mean and variance of the residual sequence are calculated, and a statistical structure feature sequence is generated to characterize the stability of the residual distribution structure.

[0101] The residual fingerprint unit is constructed by jointly organizing the time-domain abrupt change feature sequence, the frequency-domain energy distribution feature sequence, the time-frequency evolution feature sequence, and the statistical structure feature sequence. The residual fingerprint units are then combined in chronological order to form a residual fingerprint representation.

[0102] In this embodiment, the fluid-structure coupled autoencoder training module includes:

[0103] At each time window index position, the residual fingerprint unit corresponding to the time window is extracted from the residual fingerprint representation, combined to form a residual fingerprint sequence, and the vacuum state value sequence is extracted. The residual fingerprint sequence and the vacuum state value sequence are aligned and normalized to generate a standard residual fingerprint sequence and a standard vacuum state value sequence.

[0104] The standard residual fingerprint sequence and the standard vacuum state value sequence are input into the fluid-structure coupled autoencoder;

[0105] The fluid-structure coupled autoencoder includes a residual fingerprint coding layer, a vacuum state coding layer, a coupling latent constraint layer, and a reverse mapping decoding layer;

[0106] In the residual fingerprint coding layer, temporal coding is performed on the standard residual fingerprint sequence to obtain the residual latent state sequence; in the vacuum state coding layer, temporal coding is performed on the standard vacuum state value sequence to obtain the vacuum latent state sequence.

[0107] The standard residual fingerprint sequence is encoded using a time-series neural network in the order of the time window to generate a residual latent state sequence. The standard vacuum state value sequence is encoded using the same type of time-series neural network to generate a vacuum latent state sequence.

[0108] In the coupled latent constraint layer, the vacuum latent state sequence is used as the constraint reference, and the latent space consistency constraint is applied to the residual latent state sequence to generate the coupled latent state sequence, so that the evolution of the coupled latent state sequence between adjacent time window indices is consistent with the vacuum latent state sequence.

[0109] The generation of the coupled latent state sequence includes: aligning the vacuum latent state sequence and the residual latent state sequence according to the time window index; obtaining the corresponding vacuum latent state vector and residual latent state vector under each time window; using the vacuum latent state vector corresponding to the current time window as a constraint reference, extracting the values ​​of the vacuum latent state vector in each latent dimension of the latent space to form a constraint reference vector characterizing the evolution features of the vacuum state; within the same time window, aligning the residual latent state vector and the constraint reference vector dimension-by-dimensional in the latent space, matching the values ​​of the residual latent state vector in each latent dimension with the constraint reference vector. The values ​​of the corresponding latent dimensions are compared for consistency to obtain the consistency deviation in each latent dimension. The consistency deviation in each latent dimension is summarized to generate a constraint deviation value that characterizes the consistency between the residual latent state and the vacuum latent state in the current time window. Based on the constraint deviation value, the latent space consistency constraint is adjusted on the residual latent state vector to obtain the adjusted residual latent state vector that satisfies the vacuum latent state evolution constraint. The adjusted residual latent state vectors obtained in each time window are used as the coupled latent state vectors of the corresponding time window and arranged in order according to the time window index to generate a coupled latent state sequence.

[0110] In the reverse mapping decoding layer, a predictive reconstruction reverse mapping is performed based on the coupled latent state sequence to generate a residual reconstruction sequence;

[0111] When generating the residual reconstruction sequence, the coupled latent state vectors corresponding to each time window in the coupled latent state sequence are read sequentially according to the time window index order. The coupled latent state vector corresponding to the current time window is input into the inverse mapping decoding layer to obtain the predicted residual fingerprint unit under that time window. The predicted residual fingerprint unit and the residual fingerprint representation are consistent in dimensional structure. The predicted residual fingerprint units obtained under each time window are arranged in order according to the time window index order to generate the predicted residual fingerprint sequence. At the same time window index position, the predicted residual fingerprint sequence and the standard residual fingerprint sequence are compared and aligned dimension by dimension to obtain the reconstruction amount corresponding to each dimension. The reconstruction amounts obtained under each time window are combined according to the time window index order to generate the residual reconstruction sequence. The inverse mapping decoding layer adopts a network symmetrically set with the encoding structure to map the coupled latent state sequence back to the residual fingerprint space layer by layer.

[0112] The parameters of the fluid-structure coupled autoencoder are iteratively updated until the residual reconstruction sequence converges or the training is completed by reaching the maximum number of training iterations.

[0113] In this embodiment, the anomaly detection module includes:

[0114] During the online operation phase, real-time residual fingerprint representations are acquired;

[0115] The real-time residual fingerprint representation is input into the trained fluid-structure coupled autoencoder to obtain the residual reconstruction sequence;

[0116] The residual reconstruction sequence is compared with the normal operation reconstruction error reference range to obtain the anomaly judgment result, which is used to characterize whether there is an abnormal state in the current time window.

[0117] Within the time window where an abnormal state is identified, an abnormality intensity value is generated based on the degree of deviation of the reconstruction error value corresponding to that time window from the normal operation reconstruction error reference range.

[0118] The anomaly determination results and anomaly intensity values ​​generated in each time window are combined in the order of the time window index to form an anomaly determination sequence and an anomaly intensity sequence.

[0119] In this embodiment, the normal operation reconstruction error reference range is obtained by statistically summarizing the reconstruction error values ​​corresponding to each time window under historical normal operation conditions, extracting their stable distribution interval, and using it as a reference range for anomaly judgment during the online operation phase.

[0120] In this embodiment, the defense strategy generation module includes:

[0121] Using the abnormal intensity sequence as the control input, a hierarchical mapping is performed according to the abnormal intensity score corresponding to each time window in the abnormal intensity sequence, and the defense control parameters under the corresponding time window are determined respectively.

[0122] The defense control parameters are combined in chronological order to form a defense strategy, which includes parameters for flow throttling opening, flow limiting intensity, pressure compensation, and action duration.

[0123] refer to Figure 3 An adaptive quantitative water supply device based on anomaly detection includes a valve body 11, a displacement sensor 3, a support frame 10, and an outlet pipe 6. The outlet pipe 6 is connected to the upper end of the valve body 11. A water supply channel is formed inside the valve body 11. A piston 8, a spring 9, and a cone 7 are axially arranged inside the water supply channel. The cone 7 is positioned above the piston 8 and corresponds to the inlet position of the outlet pipe 6. The piston 8 moves axially along the valve body 11 under the action of water supply pressure. The spring 9 is located between the piston 8 and the valve body 11 to apply a restoring force to the piston 8, causing the piston 8 to form different axial positions under different water supply pressures. The cone 7 changes its relative distance to the inlet of the outlet pipe 6 as the piston 8 moves axially, thereby forming a variable throttling effect. The channel is used to regulate the water flow rate through the outlet pipe 6 to achieve quantitative water supply. The upper end of the plug body 11 is provided with a sealing plate 5, which is used to seal the lower end of the water supply channel and provide an installation and limiting reference for the piston 8 and the spring 9. The sealing plate 5 is connected to the plug body 11 through a sealing screw 4, which is used to fix the sealing plate 5 to the plug body 11 and form a detachable sealing structure. The support frame 10 supports the plug body 11 to ensure the structural stability of the plug body 11 during the water supply process. The plug body 11 is also provided with a cavity structure, which is connected to the water supply channel. The displacement sensor 3 is installed in the plug body 11 to collect the displacement information of the piston 8 and is connected to the metering and storage device 1 through the signal line 2.

[0124] The cavity structure is a closed space set inside the plug body 11, which is formed by the inner wall of the plug body 11 and the sealing plate 5, and is connected to the water supply channel. It is used to form a negative pressure state during the water supply process. The signal line 2 is used to transmit the piston displacement signal collected by the displacement sensor 3 to the electrical connection line of the external control device. The metering and storage device 1 is used to measure and store the displacement information to realize the flow recording and management of the water supply process.

[0125] Example 1: To verify the feasibility of this invention in practice, it was applied to a long-running adaptive quantitative water supply device in a city's public water supply network. This device serves residential water use, operating in an environment characterized by frequent switching between peak and off-peak water usage, significant water pressure fluctuations, and continuous long-term operation. Previously, in this scenario, the water supply device was prone to unstable water supply or abnormal water discharge when elastic elements fatigued, the siphon cavity experienced negative pressure fluctuations, or there was slight internal structural jamming. Furthermore, traditional monitoring methods struggled to identify the abnormal evolution process in a timely manner, often only discovering the anomaly after it had already affected water supply stability, placing significant pressure on maintenance.

[0126] In this application scenario, the system of this invention is integrated into the existing water supply device. A data acquisition module continuously and synchronously collects data on water supply pressure, piston displacement, elastic element deformation, and the vacuum state of the siphon chamber. The collected data undergoes time alignment and normalization to form a continuous sequence of standard operating states. During normal operation, the system first constructs a fluid-structure coupling relationship between the fluid action state and the structural response state based on historical stable operating data. This relationship characterizes the impact of changes in water supply pressure and vacuum state on piston displacement and elastic element deformation. Subsequently, during actual operation, the system generates a predicted response sequence in real time based on this coupling relationship and compares it with the actually collected response data, automatically forming a residual sequence reflecting the degree of operational deviation. By performing time-domain, frequency-domain, and time-frequency joint analysis on the residual sequence, the system continuously constructs a residual fingerprint representation, which is then input into a fluid-structure coupled autoencoder for online anomaly detection.

[0127] During operation, when the water supply device gradually deviates from its original normal operating pattern due to increased internal friction or fluctuations in negative pressure, the structural characteristics of the residual fingerprint change significantly. The system can identify the abnormal trend before the anomaly causes obvious water supply problems and generate a corresponding abnormal intensity sequence. Based on the changes in abnormal intensity, the defense strategy generation module automatically generates corresponding defense control strategies and drives the water supply device to perform control actions such as flow limiting or pressure compensation, gradually restoring the device's operating state to stability. The entire process continues throughout the actual operating time span, and the system maintains a stable operating state under different operating periods and different water load conditions. To verify the performance of this invention in practice, it was compared with a traditional system.

[0128] Table 1. Performance Comparison of Different Quantitative Water Supply Systems in Anomaly Monitoring and Prevention Control Capabilities

[0129]

[0130] As shown in Table 1, traditional quantitative water supply monitoring systems have a low accuracy rate in anomaly identification. This is mainly because these systems typically rely on single operating parameters such as water supply pressure or piston displacement for status judgment, lacking a characterization of the coupling relationship between the fluid dynamics and structural response within the water supply device. When operating conditions change, the system struggles to distinguish between normal operating fluctuations and abnormal states, resulting in a high false alarm rate. Furthermore, the anomaly response often lags behind the actual occurrence of the anomaly.

[0131] While water supply systems based on single anomaly identification have improved accuracy in anomaly detection, their analysis still focuses on raw sensor data and lacks a predictive comparison mechanism based on physical mechanisms. Therefore, when water supply pressure, vacuum status, or structural response change synchronously, the system is still prone to misjudging normal operation as anomaly, resulting in significant deficiencies in false alarm rate and anomaly recurrence rate. Furthermore, such systems typically lack close coordination with the water supply control process after an anomaly occurs, leaving the anomaly at risk of recurrence in subsequent operations.

[0132] The system of this invention exhibits advantages. The improved accuracy of anomaly identification stems from the system's construction of a fluid-structure coupling relationship under normal operating conditions and the generation of a predicted response sequence based on this coupling relationship. This allows the system to identify abnormal states by observing the residual structural changes between the predicted and actual responses, thus avoiding judgment based solely on fluctuations in a single parameter. The reduced false alarm rate is attributed to the fluid-structure coupled autoencoder's introduction of the evolution constraint of the siphon cavity's vacuum state in the potential space. This ensures that the anomaly determination results simultaneously conform to data characteristics and physical operating laws, reducing misjudgments caused by operational fluctuations.

[0133] Regarding anomaly response lag time, the system of this invention, through continuous construction and online analysis of residual fingerprints, can identify anomaly trends before they significantly impact water supply stability and promptly generate defensive control strategies, thereby shortening the anomaly response time. Simultaneously, since the defensive strategies directly drive the water supply device to perform control actions such as flow limiting or pressure compensation, the system can effectively suppress the continued amplification of abnormal states, reduce the probability of recurrence of abnormal states in subsequent operation, and improve the overall stability of the water supply process.

[0134] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive quantitative water supply system based on anomaly detection, characterized in that, include: The data acquisition module is used to synchronously acquire and preprocess the water supply pressure, piston displacement, elastic element deformation and siphon cavity vacuum status inside the adaptive quantitative water supply device, and generate a standard operating status sequence. The fluid-structure coupling modeling module is used to construct the fluid-structure coupling relationship of the standard operating state sequence under normal operating conditions of the water supply device, and generate a set of predicted response sequences. The residual fingerprint construction module is used to compare the predicted response sequence set with the actual collected response data during water supply operation, calculate the residual sequence between the predicted response and the actual response, and construct the residual fingerprint representation. A fluid-structure coupled autoencoder training module is used to train a fluid-structure coupled autoencoder based on residual fingerprint representation. The anomaly detection module is used during the online operation phase to input the real-time residual fingerprint representation into the trained fluid-structure coupled autoencoder, generate anomaly judgment results based on the residual reconstruction sequence, and generate anomaly intensity sequences. The defense strategy generation module is used to generate defense strategies based on the abnormal strength sequence; The water supply defense control execution module is used to drive the water supply device to perform corresponding flow interception, flow restriction or pressure compensation control actions based on the defense strategy, and maintain the corresponding control state according to the action duration parameter to realize abnormal water supply defense control.

2. The adaptive quantitative water supply system based on anomaly detection according to claim 1, characterized in that, The vacuum state of the siphon cavity is a pressure state parameter inside the siphon cavity used to maintain the siphon effect, which characterizes the degree of negative pressure formed inside the siphon cavity relative to the external atmospheric pressure. The preprocessing includes time alignment, dimension normalization and noise suppression.

3. The adaptive quantitative water supply system based on anomaly detection according to claim 1, characterized in that, The fluid-structure coupling modeling module includes: Obtain the standard operating state sequence formed by the adaptive quantitative water supply device under normal operating conditions; For each time sampling point in the standard operating state sequence, the water supply pressure and siphon cavity vacuum state at the corresponding time are extracted as input state vectors characterizing the fluid action state, and the piston displacement and elastic element deformation at the corresponding time are extracted as output state vectors characterizing the structural response state. The input state vector and output state vector corresponding to each time sampling point are paired and organized to form a time-coupled sample set; Based on the temporally coupled sample set, the influence of changes in water supply pressure and vacuum state of the siphon cavity on piston displacement and deformation of elastic elements is modeled and processed to construct a fluid-structure coupling relationship. By inputting the water supply pressure and the vacuum state of the siphon cavity from the standard operating state sequence into the fluid-structure coupling relationship, the corresponding predicted displacement response sequence and predicted deformation response sequence are obtained. The predicted displacement response sequence and the predicted deformation response sequence are arranged and stored in chronological order to form a set of predicted response sequences.

4. The adaptive quantitative water supply system based on anomaly detection according to claim 3, characterized in that, The construction of the fluid-structure coupling relationship includes: For each time sampling point in the time-series coupled sample set, the corresponding water supply pressure and siphon cavity vacuum state are extracted to form a fluid action state vector. At the same time, the piston displacement and elastic element deformation at the corresponding time sampling point are extracted to form a structural response state vector. According to the time sampling order, the fluid action state vector and the structural response state vector corresponding to each time sampling point are paired point by point to form a mapping sample set; Based on the mapped sample set, differential processing is performed on the water supply pressure, siphon cavity vacuum state, piston displacement and elastic element deformation of adjacent sampling points according to the time sampling order to obtain the water supply pressure change sequence, vacuum state change sequence, piston displacement change sequence and elastic element deformation change sequence. Calculate the pressure-displacement response coefficient for the change in water supply pressure and piston displacement in each time period, and calculate the vacuum-deformation response coefficient for the change in vacuum state and deformation of elastic element in each time period. Calculate the pressure-deformation coupling coefficient for the change in water pressure and the change in deformation of the elastic element in each time period, and calculate the vacuum-displacement coupling coefficient for the change in vacuum state and the change in piston displacement in each time period. The pressure-displacement response coefficient value, vacuum-deformation response coefficient value, pressure-deformation coupling coefficient value, and vacuum-displacement coupling coefficient value are combined according to the time period index to obtain the coupling parameter set; Perform statistical summarization processing on the set of coupling parameters to generate a coupling feature parameter vector, which is then associated and stored with the corresponding fluid action state vector and structural response state vector to form a fluid-structure coupling relationship.

5. The adaptive quantitative water supply system based on anomaly detection according to claim 1, characterized in that, The residual fingerprint construction module includes: During water supply operation, the water supply pressure and the vacuum state of the siphon cavity are collected in real time, and a set of corresponding predicted response sequences is generated based on the fluid-structure coupling relationship. The predicted displacement response and predicted deformation response in the set of actual piston displacement, elastic element deformation and predicted response sequence collected at the same sampling point are aligned point by point. The displacement residual value between the actual displacement and the predicted displacement, and the deformation residual value between the actual deformation and the predicted deformation are calculated respectively. The displacement residual sequence and the deformation residual sequence are formed in time order. Within a preset sliding time window, temporal abrupt structure extraction processing is performed on the displacement residual sequence and the deformation residual sequence respectively. The residual change amount, change direction identifier and change duration of adjacent sampling points are calculated to generate a temporal abrupt feature sequence. Within the same time window, frequency domain transformation is performed on the displacement residual sequence and the deformation residual sequence respectively to extract the energy distribution value, dominant frequency position and frequency band energy concentration of the residuals, forming a frequency domain energy distribution feature sequence. Perform time-frequency joint analysis on the displacement residual sequence and the deformation residual sequence, calculate the local concentration of residual energy on the time-frequency plane, extract the index of residual energy surge interval and the corresponding time-frequency energy intensity value, and generate a time-frequency evolution feature sequence; Statistical structure modeling is performed on the displacement residual sequence and deformation residual sequence within the corresponding time window. The mean and variance of the residual sequence are calculated to generate a statistical structure feature sequence. The residual fingerprint unit is constructed by jointly organizing the time-domain abrupt change feature sequence, the frequency-domain energy distribution feature sequence, the time-frequency evolution feature sequence, and the statistical structure feature sequence. The residual fingerprint units are then combined in chronological order to form a residual fingerprint representation.

6. The adaptive quantitative water supply system based on anomaly detection according to claim 1, characterized in that, The fluid-structure coupled autoencoder training module includes: At each time window index position, the residual fingerprint unit corresponding to the time window is extracted from the residual fingerprint representation, combined to form a residual fingerprint sequence, and the vacuum state value sequence is extracted. The residual fingerprint sequence and the vacuum state value sequence are aligned and normalized to generate a standard residual fingerprint sequence and a standard vacuum state value sequence. The standard residual fingerprint sequence and the standard vacuum state value sequence are input into the fluid-structure coupled autoencoder; The fluid-structure coupled autoencoder includes a residual fingerprint encoding layer, a vacuum state encoding layer, a coupling latent constraint layer, and a reverse mapping decoding layer; In the residual fingerprint coding layer, temporal coding is performed on the standard residual fingerprint sequence to obtain the residual latent state sequence; in the vacuum state coding layer, temporal coding is performed on the standard vacuum state value sequence to obtain the vacuum latent state sequence. In the coupled latent constraint layer, the vacuum latent state sequence is used as the constraint reference, and the latent space consistency constraint is applied to the residual latent state sequence to generate the coupled latent state sequence. In the reverse mapping decoding layer, a predictive reconstruction reverse mapping is performed based on the coupled latent state sequence to generate a residual reconstruction sequence; The parameters of the fluid-structure coupled autoencoder are iteratively updated until the residual reconstruction sequence converges or the training is completed by reaching the maximum number of training iterations.

7. The adaptive quantitative water supply system based on anomaly detection according to claim 1, characterized in that, The anomaly detection module includes: During the online operation phase, real-time residual fingerprint representations are acquired; The real-time residual fingerprint representation is input into the trained fluid-structure coupled autoencoder to obtain the residual reconstruction sequence; The residual reconstruction sequence is compared with the normal operation reconstruction error reference range to obtain the anomaly judgment result, which is used to characterize whether there is an abnormal state in the current time window. Within the time window where an abnormal state is identified, an abnormality intensity value is generated based on the degree of deviation of the reconstruction error value corresponding to that time window from the normal operation reconstruction error reference range. The anomaly determination results and anomaly intensity values ​​generated in each time window are combined in the order of the time window index to form an anomaly determination sequence and an anomaly intensity sequence.

8. The adaptive quantitative water supply system based on anomaly detection according to claim 7, characterized in that, The normal operation reconstruction error reference range is obtained by statistically summarizing the reconstruction error values ​​corresponding to each time window under historical normal operation conditions, extracting their stable distribution interval, and using it as a reference range for anomaly judgment during the online operation phase.

9. The adaptive quantitative water supply system based on anomaly detection according to claim 1, characterized in that, The defense strategy generation module includes: Using the abnormal intensity sequence as the control input, a hierarchical mapping is performed according to the abnormal intensity score corresponding to each time window in the abnormal intensity sequence, and the defense control parameters under the corresponding time window are determined respectively. The defense control parameters are combined in chronological order to form a defense strategy, which includes flow throttling parameters, flow limiting intensity parameters, pressure compensation parameters, and action duration parameters.

10. An adaptive quantitative water supply device based on anomaly detection, characterized in that, The device includes a plug body (11), a displacement sensor (3), a support frame (10), and a water outlet pipe (6). The water outlet pipe (6) is connected to the upper end of the plug body (11). A water supply channel is formed inside the plug body (11). A piston (8), a spring (9), and a cone (7) are arranged axially within the water supply channel. The cone (7) is positioned above the piston (8) and corresponds to the inlet position of the water outlet pipe (6). The piston (8) moves axially along the plug body (11) under the action of water supply pressure. The spring (9) is positioned between the piston (8) and the plug body (11). Between 11), the cone (7) changes its relative distance with the inlet of the outlet pipe (6) as the piston (8) moves axially. A sealing plate (5) is provided at the upper end of the plug body (11). The sealing plate (5) is connected to the plug body (11) through the sealing screw (4). The support frame (10) supports the plug body (11). A cavity structure is provided inside the plug body (11). The cavity structure is connected to the water supply channel. The displacement sensor (3) is installed inside the plug body (11) and connected to the metering and storage device (1) through the signal line (2).

Images