Abnormal detection method and device for dynamic management of liquid assets and related medium

Abstract

The invention discloses an anomaly detection method and device for dynamic management of liquid assets and a related medium, and the method comprises the steps: obtaining liquid level data and temperature and humidity sampling data of the liquid assets, and carrying out the data preprocessing of the liquid level data and the temperature and humidity sampling data, and obtaining time sequence state data; respectively inputting the time sequence state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly degree set; performing hierarchical fusion on the dual-mode anomaly degree set to obtain a fusion judgment result; a final judgment result is obtained through calculation according to the fusion judgment result; and generating alarm event information based on the final judgment result, and outputting abnormal detection data. According to the invention, by carrying out physical model and machine learning model parallel analysis on the liquid level and temperature and humidity time sequence data and carrying out hierarchical fusion judgment, normal natural loss and abnormal events can be distinguished in real time at a data source end, so that false alarm and missing detection are reduced, and invalid data report is reduced.

Description

An anomaly detection method, device, and related media for dynamic management of liquid assets. Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an anomaly detection method, device, and related media for dynamic management of liquid assets. Background Technology

[0002] In the storage management of high-value liquid assets, existing management methods typically employ manual periodic inventory checks or IoT-based fixed threshold alarms to monitor liquid level changes. However, liquid assets undergo natural attrition processes during normal storage due to environmental factors such as temperature and humidity, resulting in slow and continuous dynamic changes in liquid levels. In this context, fixed threshold alarms struggle to adapt to environmental fluctuations, making it difficult to effectively distinguish between normal natural attrition and abnormal events such as minor leaks or theft. This leads to frequent false alarms or missed early anomalies. Summary of the Invention

[0003] This invention provides an anomaly detection method, device, and related media for dynamic management of liquid assets, aiming to solve the technical problems in the prior art where liquid level monitoring based on fixed thresholds is difficult to adapt to changes in environmental conditions such as temperature and humidity, and cannot effectively distinguish abnormal events, resulting in frequent false alarms or missed anomalies.

[0004] In a first aspect, embodiments of the present invention provide an anomaly detection method for dynamic management of liquid assets, comprising: acquiring liquid level data and temperature and humidity sampling data of liquid assets respectively, and performing data preprocessing on the liquid level data and temperature and humidity sampling data to obtain time-series state data; inputting the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set; performing hierarchical fusion on the dual-mode anomaly set to obtain a fusion judgment result; generating liquid asset identifiers and timestamp identifiers according to the fusion judgment result, and integrating them to obtain a final judgment result; generating alarm event information based on the final judgment result, and outputting anomaly detection data.

[0005] Secondly, embodiments of the present invention provide an anomaly detection device for dynamic management of liquid assets, comprising: a data acquisition unit, used to acquire liquid level data and temperature and humidity sampling data of liquid assets respectively, and perform data preprocessing on the liquid level data and temperature and humidity sampling data to obtain time-series state data; a data analysis unit, used to input the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set; a data fusion unit, used to perform hierarchical fusion on the dual-mode anomaly set to obtain a fusion judgment result; a data judgment unit, used to generate liquid asset identifiers and timestamp identifiers according to the fusion judgment result, and integrate them to obtain a final judgment result; and a data output unit, used to generate alarm event information based on the final judgment result and output anomaly detection data.

[0006] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the anomaly detection method for dynamic management of liquid assets in the first aspect.

[0007] Fourthly, embodiments of the present invention provide a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the anomaly detection method for dynamic management of liquid assets of the first aspect.

[0008] This invention provides an anomaly detection method for dynamic management of liquid assets, including acquiring liquid level data and temperature and humidity sampling data of the liquid assets, preprocessing the liquid level data and temperature and humidity sampling data to obtain time-series state data; inputting the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set; performing hierarchical fusion of the dual-mode anomaly set to obtain a fusion judgment result; generating liquid asset identifiers and timestamp identifiers based on the fusion judgment result, and integrating them to obtain a final judgment result; generating alarm event information based on the final judgment result, and outputting anomaly detection data. This invention, by performing parallel analysis of liquid level and temperature and humidity time-series data using physical and machine learning models and performing hierarchical fusion judgment, can distinguish between normal natural losses and abnormal events in real time at the data source, reducing false alarms and missed detections and reducing invalid data reporting.

[0009] This invention also provides an anomaly detection device, computer equipment, and storage medium for dynamic management of liquid assets, which have the same beneficial effects as described above. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 is a flowchart illustrating an anomaly detection method for dynamic management of liquid assets provided in an embodiment of the present invention; Figure 2 is a schematic block diagram illustrating an anomaly detection device for dynamic management of liquid assets provided in an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0014] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0015] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0016] Please refer to Figure 1 below. Figure 1 is a flowchart of an anomaly detection method for dynamic management of liquid assets provided by an embodiment of the present invention, specifically including steps S101 to S105.

[0017] S101. Acquire liquid level data and temperature and humidity sampling data of liquid assets respectively, and preprocess the liquid level data and temperature and humidity sampling data to obtain time-series state data; S102. Input the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set; S103. Perform hierarchical fusion of the dual-mode anomaly set to obtain a fusion judgment result; S104. Generate liquid asset identifiers and timestamp identifiers according to the fusion judgment result, and integrate them to obtain the final judgment result; S105. Generate alarm event information based on the final judgment result and output anomaly detection data.

[0018] In step S101, the system's intelligent sensing node periodically collects liquid level data and temperature and humidity sampling data of liquid assets in a low-power, low-frequency sampling mode, and performs preprocessing and state estimation processing on the sampled data to output time-series state data that characterizes liquid level changes, ambient temperature and relative humidity over a period of time.

[0019] In one embodiment, step S101 includes: setting the liquid asset node to a low-frequency sampling mode to obtain sampling mode configuration data; collecting liquid level data and temperature and humidity sampling data of the liquid asset using a multi-dimensional sensing and data preprocessing unit based on the sampling mode configuration data, and processing it to obtain preprocessed sampling data; and using the preprocessed sampling data to calculate a high-confidence state mean through an extended Kalman filter to obtain time-series state data.

[0020] In this embodiment, the sampling can be performed by a liquid asset node deployed at the liquid asset storage site. This liquid asset node pre-enters a low-power operating state and configures its sampling strategy to a low-frequency sampling mode to obtain sampling mode configuration data. The low-frequency sampling mode can be set to sample once per minute, used to control the node's computational and communication load while ensuring continuous monitoring. After the sampling mode configuration data takes effect, the liquid asset node invokes a multi-dimensional sensing and data preprocessing unit to collect liquid level data and corresponding temperature and humidity sampling data for the ambient temperature and relative humidity in each sampling cycle. Preprocessing is then performed on the raw sampled values ​​to obtain preprocessed sampled data. The preprocessing may include timestamp alignment, missing value handling, outlier removal, range conversion, and digital filtering, ensuring that the liquid level data and temperature and humidity sampling data can be used for subsequent state estimation calculations under the same time reference.

[0021] Furthermore, at time t0, the liquid asset node triggers a state summary calculation: the multi-dimensional sensing and data preprocessing unit retrieves and organizes preprocessed sampling data from the past hour, constructs an observation sequence and a state update sequence, and inputs them into an extended Kalman filter (EKF) to estimate the high-confidence state mean, outputting time-series state data. This time-series state data may include statistical representations of the liquid level change trend and environmental conditions within the time window. For example, in this embodiment, the state mean calculated at time t0 may be a liquid level drop of 0.05 mm, an ambient temperature of 20.1°C, and a relative humidity of 0.75, thus providing a consistent, steady-state, and comparable time-series state input for the subsequent physical model engine and machine learning model engine. The time-series state data output after the above processing exhibits higher stability and consistency.

[0022] In step S102, the physical model engine calculates key quantities related to theoretical volatilization based on the current temperature and humidity and obtains the physical model anomaly degree; the machine learning model engine inputs the time-series state data into the deployed TCAE model to perform reconstruction calculation, obtains the machine learning model anomaly degree based on the reconstruction error, and can combine a lightweight temporal convolutional autoencoder to learn the spatiotemporal correlation of multidimensional data to obtain data-driven anomaly degree output.

[0023] In one embodiment, step S102 includes: obtaining the current temperature from the time-series state data to obtain temperature data; calculating the saturated vapor pressure of the temperature data using a preset equation to obtain saturated vapor pressure data; calculating the actual water vapor partial pressure in the air based on the saturated vapor pressure data to obtain water vapor partial pressure data; calculating the liquid level using the saturated vapor pressure data and the water vapor partial pressure data to obtain the liquid level drop; and calculating the physical model anomaly data using the liquid level drop.

[0024] In this embodiment, the physical model engine first reads the current temperature as temperature data T from the time-series state data, and simultaneously reads the relative humidity as RH, so as to maintain the same data benchmark as the subsequent vapor pressure calculation; for example, at time t0, the temperature data can be 20.1°C and the relative humidity can be 0.75.

[0025] After obtaining the temperature data T, the physics model engine calls a preset equation to calculate the saturated vapor pressure, obtaining the saturated vapor pressure data Psat. The preset equation uses the Antoine equation, taking empirical constants A, B, and C related to the liquid type and the temperature data T as input, and outputting the saturated vapor pressure at the current temperature. Based on the saturated vapor pressure data Psat and the relative humidity RH, the physics model engine calculates the actual water vapor partial pressure in the air, obtaining the water vapor partial pressure data Pair. The water vapor partial pressure data Pair is used to characterize the influence of the current environment on the driving force of volatilization and serves as a key input for subsequent liquid level calculations.

[0026] After obtaining Psat and Pair, the physics model engine uses the difference between them as a characterization of the evaporation driving force, and further combines it with the mass transfer formula to calculate the liquid level drop Δh. predicted The Δh predicted This is used to characterize the liquid level drop corresponding to the theoretical evaporation amount under current temperature and humidity conditions. Taking this embodiment as an example, the physical model engine can calculate that the liquid level drop corresponding to the theoretical evaporation amount is 0.048 mm.

[0027] The liquid level drop Δh was obtained. predicted Then, the physics model engine compares it with the actual liquid level change Δh obtained from the liquid level data within the same time window. measured Deviation calculations were performed, and the historical error standard deviation σ under normal operating conditions was considered. p Perform normalization processing and output physical model anomaly data. For example, in this embodiment, the physical model anomaly is 0.4. The relationship between the Antoine equation and the actual water vapor partial pressure calculation can be expressed by the following formula: in, The values ​​are saturated vapor pressure data in mmHg; A, B, and C are empirical constants related to the liquid type; T is temperature data in °C. Logarithmic operations to base 10; The values ​​represent water vapor partial pressure in mmHg; RH represents relative humidity, a dimensionless value between 0 and 1. The air temperature is typically taken as T.

[0028] In one embodiment, the step of calculating the liquid level drop using the saturated vapor pressure data and water vapor partial pressure data includes: obtaining the mass transfer coefficient, liquid surface area, and time increment corresponding to the mass transfer law to obtain mass transfer parameters; calculating the difference between the saturated vapor pressure data and the water vapor partial pressure data to obtain vapor pressure difference data; and calculating the liquid level drop based on the vapor pressure difference data and the mass transfer parameters.

[0029] In this embodiment, the physics model engine obtains the mass transfer coefficient k and the liquid surface area A corresponding to the mass transfer law. surface And the time increment Δt, to obtain the mass transfer parameters; among them, the mass transfer coefficient k is used to characterize the mass transfer capacity between the liquid surface and the surrounding gas, which can be determined by preset parameters on the node side or by historical calibration results, and the liquid surface area A surfaceThe time increment Δt is calculated from the geometric dimensions of the liquid asset and the current liquid level state. It characterizes the length of the time window corresponding to this theoretical calculation and maintains consistency with the statistical window of the time-series state data. After obtaining the mass transfer parameters, the physics model engine uses the saturated vapor pressure data P... sat With the water vapor partial pressure data P air The difference is calculated to obtain the vapor pressure difference data ΔP; the vapor pressure difference data is used to characterize the driving force between the saturated vapor pressure of the liquid surface and the actual partial pressure of the environment, and is updated synchronously with the temperature and humidity conditions, so that the theoretical calculation can change with the environment in real time.

[0030] Furthermore, based on the law of mass transfer, the physics model engine uses the vapor pressure difference data ΔP and mass transfer parameters to perform joint calculations to obtain the liquid level drop Δh. predicted This is used to characterize the theoretical drop in liquid level within the time increment Δt. The calculation relationship for the above liquid level drop can be expressed by the following formula (given only once here): in, The liquid level drop is measured in mm; k is the mass transfer coefficient, used to characterize the mass transfer intensity. The surface area of ​​the liquid; For liquid at temperature The saturated vapor pressure data is given below, in mmHg. This represents the liquid temperature, typically taken as the current temperature data. These are water vapor partial pressure data, in mmHg. This is the time increment, used to characterize the length of the time window for theoretical calculations. Through the above steps, the physical model engine outputs the theoretical liquid level drop that changes in real time with the environment, providing basic data for subsequent comparison of measured liquid level changes with theoretical changes.

[0031] In one embodiment, the step of calculating the physical model anomaly data using the liquid level drop includes: obtaining the actual liquid level change within a preset time increment based on the time-series state data to obtain the measured drop; determining the corresponding time based on the measured drop, and obtaining the corresponding theoretical drop in the liquid level drop based on the corresponding time; obtaining the historical error standard deviation of the preset physical model engine under normal operating conditions based on the theoretical drop; taking the absolute value of the difference between the measured drop and the theoretical drop, and normalizing it using the historical error standard deviation to obtain the physical model anomaly data.

[0032] In this embodiment, the physical model engine determines a statistical window corresponding to a preset time increment based on the time-series state data, and extracts the actual liquid level change from the liquid level data within the statistical window to obtain the measured drop Δh. measuredThe measured drop is used to characterize the actual drop in liquid level within the same time increment and is time-aligned with subsequent theoretical drops. The measured drop Δh is obtained from this data. measured Subsequently, the physics model engine accesses the liquid level drop sequence and extracts the theoretical drop Δh at the same time based on the time identifier t corresponding to the measured drop. predicted This allows for direct calculation of the difference between the measured and theoretical descent values ​​under the same time reference, thereby avoiding the accumulation of deviations caused by inconsistent sampling intervals or window boundaries.

[0033] Furthermore, the physics model engine completes the theoretical descent Δh predicted After positioning, the historical error standard deviation σ of the preset physical model engine under normal operating conditions is obtained. p The historical error standard deviation σ p Error scaling parameters, either pre-set at the node side or obtained from statistical analysis of historical normal operation data, are used to characterize the inherent error fluctuation range of the physical model under normal conditions and are used in subsequent normalization processing to improve the comparability of anomalies across different liquid assets and time windows. The physical model engine then applies the measured drop Δh... measured With the theoretical decrease Δh predicted Perform the difference calculation and take the absolute value to obtain the deviation magnitude, and then use the historical error standard deviation σ as the basis for the result. p The deviation magnitude is normalized to obtain the physical model anomaly data S. p (t), used to output the normalized deviation between theoretical and actual changes. The anomaly degree of the physical model can be calculated using the following formula: Among them, S p (t) represents the anomaly data of the physical model, used to characterize the normalized deviation at time t; The measured drop represents the actual change in liquid level obtained from the liquid level data within a preset time increment; The theoretical drop represents the drop in liquid level calculated from the liquid level and aligned with the measured drop at the same time. This indicates the absolute value operation; The historical error standard deviation represents the historical error standard deviation of the physical model under normal operating conditions. Through the above processing, the physical model engine outputs the physical model anomaly data S. p (t) can be directly used as the physical side anomaly input in the dual-mode anomaly set for subsequent hierarchical fusion determination.

[0034] In one embodiment, step S102 further includes: extracting cached data from the time-series state data and organizing it according to a preset sampling order to construct model input data; inputting the model input data into a deployed processing model for reconstruction and outputting reconstruction sequence data; calculating the reconstruction error using the model input data and the reconstruction sequence data and normalizing and taking the square root to obtain machine learning model anomaly data; inputting the machine learning model anomaly data into a deep learning-based behavior pattern analysis engine to call a lightweight temporal convolutional autoencoder to calculate analysis data; quantizing the analysis data to obtain a quantized model object; and updating the quantized model object to obtain machine learning model update data.

[0035] In this embodiment, the machine learning model engine first extracts multi-dimensional time-series data within a preset time window from the local cache of the time-series state data, and aligns and stitches the data such as liquid level, temperature, and relative humidity at each sampling time according to a preset sampling order to construct the model input data; for example, the model input data can be selected from the time-series data of the past hour, denoted as... This is used to characterize the continuous observation segment before time t0, thus ensuring that subsequent reconstruction calculations and anomaly calculations are based on the same time reference. After obtaining the model input data, the machine learning model engine inputs it into the deployed processing model TCAE to perform reconstruction inference, outputting reconstruction sequence data aligned with the same dimension and time step as the model input data. The machine learning model engine uses the model input data and the reconstruction sequence data to calculate the reconstruction error, and performs normalized square root processing on the reconstruction error to obtain the machine learning model anomaly data; in one example, the normalized model anomaly is 0.6, which characterizes the degree of deviation of the data-driven model from the current behavior within this time window. After obtaining the machine learning model anomaly data, it is input into the deep learning-based behavior pattern analysis engine to call the lightweight temporal convolutional autoencoder deployed inside the node. The lightweight temporal convolutional autoencoder is used to learn the spatiotemporal correlation patterns between multidimensional sensor data, and outputs the anomaly of the data-driven model based on the reconstruction error between the input data and the reconstruction data. This leads to the acquisition of analytical data. It can be calculated using the following formula: in, t represents the anomaly degree at time t; W represents the number of time steps, used to characterize the number of sampling steps within the time window; D represents the number of data dimensions, used to characterize the number of dimensions of the multidimensional sensor features. For the input data The value at the i-th time step and the j-th dimension; For the reconstructed data The value taken at the corresponding position; Used for mean normalization. Therefore, the analysis data may include the anomaly degree and its corresponding temporal reconstruction deviation characterization, which is used subsequently to participate in the construction of the dual-mode anomaly degree set together with the physical model output.

[0036] Furthermore, to adapt to the resource-constrained operating environment of microcontrollers (MCUs), the model inference process corresponding to the analysis data can be quantized within the behavior pattern analysis engine: dilated causal convolution is used on the lightweight temporal convolutional autoencoder to reduce computational overhead, and combined with pruning and 8-bit integer quantization, the model parameters and inference operators are converted into a quantized form suitable for execution on the node side, resulting in a quantized model object. This reduces node-side computing power and storage usage while maintaining temporal modeling capabilities. With the quantized model object available, the liquid asset node supports model update operations on the quantized model object: when a better model is obtained through cloud training, the node receives and deploys the update content via differential firmware update. After completing consistency verification, the corresponding parameters of the quantized model object are replaced, allowing the behavior pattern analysis engine to continue outputting analysis data based on the updated model in subsequent monitoring cycles. This yields machine learning model update data, used to characterize the model parameter state corresponding to this update.

[0037] In step S103, hierarchical fusion is performed on the dual-mode anomaly set: physical model threshold and machine learning model threshold are set, and high-confidence judgment and low-confidence judgment are performed in sequence; in the high-confidence judgment, when both types of anomalies exceed the corresponding threshold, an anomaly judgment is output; in the low-confidence judgment, when any anomaly exceeds the corresponding threshold, a suspected anomaly judgment is output; in other cases, a normal judgment is output, thereby obtaining the fusion judgment result.

[0038] In one embodiment, step S103 includes: extracting physical model anomaly data and machine learning model anomaly data from the dual-mode anomaly set, and integrating them to obtain anomaly data to be fused; setting physical model threshold parameters and machine learning model threshold parameters based on the anomaly data to be fused to obtain threshold configuration data; comparing the physical model anomaly data and the machine learning model anomaly data according to the threshold configuration data to obtain physical over-threshold identification data and machine over-threshold identification data; performing hierarchical fusion judgment based on the physical over-threshold identification data and the machine over-threshold identification data; generating anomaly identification data when both the physical over-threshold identification data and the machine over-threshold identification data indicate over-threshold; generating suspected anomaly identification data when either the physical over-threshold identification data or the machine over-threshold identification data indicates over-threshold; generating normal identification data when neither the physical over-threshold identification data nor the machine over-threshold identification data indicates over-threshold; and merging and outputting the anomaly identification data, suspected anomaly identification data, and normal identification data to obtain a fusion judgment result.

[0039] In this embodiment, the dual-mode anomaly set includes at least physical model anomaly data. Anomaly data with machine learning models The hierarchical fusion decision unit first performs field extraction and alignment processing on the dual-mode anomaly set: using the time identifier t as an index, it reads the corresponding values ​​at the same time. and The two are then integrated into anomaly data to be fused, which is used for subsequent threshold comparison and judgment calculation. To ensure consistency in the judgment, the hierarchical fusion decision unit can regard the output of the physical model as a physical benchmark that changes in real time with temperature and humidity, and use it to check and verify the output of the machine learning model, so that the fusion judgment result reflects both deviations from physical laws and deviations from historical behavior.

[0040] After obtaining the anomaly data to be fused, the hierarchical fusion decision unit sets the physical model threshold parameters based on the node preset strategy. Threshold parameters of machine learning models The threshold configuration data is obtained; this threshold configuration data can be given by node-fixed parameters or issued by the upper-level management terminal and take effect on the node side. Subsequently, the hierarchical fusion decision unit performs threshold comparisons respectively: and Compare the obtained physical threshold identification data, and and The comparison yields machine over-threshold identification data; among which, physical over-threshold identification data is used to indicate whether the physical model side exceeds the threshold, and machine over-threshold identification data is used to indicate whether the machine learning model side exceeds the threshold, providing Boolean input for hierarchical determination.

[0041] Furthermore, the hierarchical fusion decision unit performs hierarchical fusion judgment based on physical threshold-breaking identifier data and machine threshold-breaking identifier data: the first layer is a high-confidence judgment, which uses AND logic to jointly calculate the two types of threshold-breaking identifiers; the second layer is a low-confidence judgment, which uses OR logic to merge the two types of threshold-breaking identifiers; when both types of threshold-breaking identifiers indicate threshold breaking, abnormal identifier data is output; when only one type of threshold-breaking identifier indicates threshold breaking, suspected abnormal identifier data is output; when neither type of threshold-breaking identifier indicates threshold breaking, normal identifier data is output. The fusion judgment can be represented by the following logical expression: in, The result is the fusion determination result; t is the time marker; This is data on the anomaly of the physical model; This is data on anomalies in machine learning models; For physical model threshold parameters; For threshold parameters of machine learning models; Indicates AND logic; "OR" indicates OR logic; "otherwise" indicates other cases besides the above conditions. Based on this decision logic, the hierarchical fusion decision unit can sequentially complete high-confidence and low-confidence decisions within the same sampling period, for example, when... =0.4、 =0.6、 = When the value is 3.0, the first-level decision check (0.4>3.0) AND (0.6>3.0) returns false, and the second-level decision check (0.4>3.0) OR (0.6>3.0) returns false, thus outputting normal identification data.

[0042] After outputting the identification data, the hierarchical fusion decision unit merges the abnormal identification data, suspected abnormal identification data, and normal identification data to output a fusion judgment result, and writes the fusion judgment result into the node-side result cache for subsequent processing. Simultaneously, to coordinate with the operation of the event-driven communication unit, the hierarchical fusion decision unit can, according to... The value of generates the communication wake-up trigger condition: if and only if When an abnormality or suspected abnormality is detected, the communication module is triggered, which generates and reports a structured event message containing preliminary diagnostic information about the root cause. If the result is determined to be normal, the communication module remains in sleep mode and only the result of this fusion determination is recorded.

[0043] In step S104, a liquid asset identifier for characterizing the monitored object is generated based on the fusion determination result, and a timestamp identifier corresponding to this analysis is obtained. The identifier is integrated with the fusion determination result to obtain the final determination result, which is used to record the object ownership and time information of this monitoring.

[0044] In step S105, alarm event information is generated and abnormal detection data is output based on the final judgment result; when the final judgment result is abnormal or suspected abnormal, the communication module is woken up to generate a structured event message and report it; when the final judgment result is normal, the node remains silent, only records the local log and enters the next monitoring cycle.

[0045] In a specific implementation scenario, the anomaly detection method can be applied to the A1 cellar area of ​​a high-end baijiu (Chinese liquor) cellar, where the environment is relatively constant in temperature and humidity. The base liquor jar numbered T2701 is used as the monitoring object. An integrated hardware and software intelligent sensing node is installed on the jar. The node stores the geometric dimension parameters corresponding to T2701 for calculating the liquid surface area, as well as physical constants related to the base liquor type and the weight parameters of the AI ​​model TCAE. A threshold of 3.0 is configured, and the historical error standard deviation of the preset physical model engine under normal operating conditions is set to 0.005 mm.

[0046] Under normal evaporation conditions, the intelligent sensing node operates in a low-power, low-frequency sampling mode, with a sampling frequency set to once per minute. At time t0, the multi-dimensional sensing and data preprocessing unit retrieves and processes the sampling data from the past hour, and calculates a high-confidence state mean using extended Kalman filtering to form time-series state data. This time-series state data represents a liquid level drop of 0.05 mm, an ambient temperature of 20.1°C, and a relative humidity of 0.75. A dual-mode parallel analysis is then performed: the physical model engine, based on the current temperature and humidity, uses the Antoine equation and mass transfer formula to calculate the theoretical evaporation amount corresponding to a liquid level drop of 0.048 mm, and further outputs a physical model anomaly score of 0.4; the machine learning model engine then multiplies the time-series data from the past hour by... t0 The deployed TCAE model is input. After the model completes data reconstruction, the reconstruction error is calculated, and the normalized model anomaly score of 0.6 is output. After obtaining the dual-mode anomaly score, the hierarchical fusion decision unit sequentially performs high-confidence and low-confidence judgments: the first-level decision check (0.4>3.0) AND (0.6>3.0) results in a false judgment; the second-level decision check (0.4>3.0) OR (0.6>3.0) also results in a false judgment, thus generating a normal judgment result. Based on this final judgment result, the event-driven communication unit determines the state to be normal, the node remains silent, does not report data, only records the monitoring result in the local log, and enters the next monitoring cycle.

[0047] In summary, this application integrates data acquisition, dual-mode parallel analysis, and hierarchical fusion judgment at the liquid asset node level, enabling anomaly detection to be calculated and output on-site. This reduces network bandwidth consumption and cloud computing and storage burden caused by continuous uploading of raw data, and lowers communication latency introduced by remote processing, achieving real-time monitoring and rapid on-site response for liquid assets. Simultaneously, this method uses parallel acquisition of anomaly degrees from both physical and machine learning models, followed by hierarchical fusion to output the fusion judgment result. This ensures the judgment process is constrained by both physical laws and historical behavior patterns, effectively reducing false alarms and missed detections caused by the inability of a single model to distinguish between normal natural wear and tear and minor anomalies. Furthermore, it can trigger event-driven alarm reporting when anomalies or suspected anomalies occur, improving the sensitivity and timeliness of anomaly identification. Moreover, the integrated hardware and software node can operate stably on-site for extended periods, reducing deployment and maintenance costs, and is suitable for large-scale deployment and continuous management in warehousing scenarios.

[0048] Referring to Figure 2, which is a schematic block diagram of an anomaly detection device for dynamic management of liquid assets provided in an embodiment of the present invention, the anomaly detection device 200 for dynamic management of liquid assets includes: a data acquisition unit 201, used to acquire liquid level data and temperature and humidity sampling data of liquid assets respectively, and perform data preprocessing on the liquid level data and temperature and humidity sampling data to obtain time-series state data; a data analysis unit 202, used to input the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly degree set; a data fusion unit 203, used to perform hierarchical fusion of the dual-mode anomaly degree set to obtain a fusion judgment result; a data judgment unit 204, used to generate liquid asset identifiers and timestamp identifiers according to the fusion judgment result, and integrate them to obtain a final judgment result; and a data output unit 205, used to generate alarm event information based on the final judgment result and output anomaly detection data.

[0049] In this embodiment, the data acquisition unit 201 acquires liquid level data and temperature and humidity sampling data of the liquid asset, and performs data preprocessing on the liquid level data and temperature and humidity sampling data to obtain time-series state data; the data analysis unit 202 inputs the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set; the data fusion unit 203 performs hierarchical fusion on the dual-mode anomaly set to obtain a fusion judgment result; the data judgment unit 204 generates liquid asset identifiers and timestamp identifiers according to the fusion judgment result, and integrates them to obtain the final judgment result; the data output unit 205 generates alarm event information based on the final judgment result and outputs anomaly detection data.

[0050] In one embodiment, the data acquisition unit 201 is specifically used to: set the liquid asset node to a low-frequency sampling mode to obtain sampling mode configuration data; based on the sampling mode configuration data, use a multi-dimensional sensing and data preprocessing unit to collect liquid level data and temperature and humidity sampling data of the liquid asset, and process them to obtain preprocessed sampling data; use the preprocessed sampling data to calculate a high-confidence state mean through extended Kalman filtering to obtain time-series state data.

[0051] In one embodiment, the data analysis unit 202 is specifically used for: obtaining the current temperature from the time-series state data to obtain temperature data; calculating the saturated vapor pressure of the temperature data using a preset equation to obtain saturated vapor pressure data; calculating the actual water vapor partial pressure in the air based on the saturated vapor pressure data to obtain water vapor partial pressure data; performing liquid level calculation using the saturated vapor pressure data and water vapor partial pressure data to obtain liquid level drop; and calculating physical model anomaly data using the liquid level drop.

[0052] In one embodiment, the data analysis unit 202 is further specifically used to: obtain the mass transfer coefficient, liquid surface area, and time increment corresponding to the mass transfer law to obtain mass transfer parameters; calculate the difference between the saturated vapor pressure data and the water vapor partial pressure data to obtain vapor pressure difference data; and calculate the liquid level drop based on the vapor pressure difference data and the mass transfer parameters.

[0053] In one embodiment, the data analysis unit 202 is further configured to: obtain the actual liquid level change within a preset time increment based on the time-series state data, and obtain the measured drop; determine the corresponding time based on the measured drop, and obtain the theoretical drop corresponding to the liquid level drop based on the corresponding time; obtain the historical error standard deviation of the preset physical model engine under normal operating conditions based on the theoretical drop; take the absolute value of the difference between the measured drop and the theoretical drop, and normalize it using the historical error standard deviation to obtain physical model anomaly data.

[0054] In one embodiment, the data analysis unit 202 is further configured to: extract cached data from the time-series state data, and organize and construct model inputs according to a preset sampling order to obtain model input data; input the model input data into a deployed processing model for reconstruction, and output reconstruction sequence data; calculate the reconstruction error using the model input data and the reconstruction sequence data, and normalize and square root the error to obtain machine learning model anomaly data; input the machine learning model anomaly data into a deep learning-based behavior pattern analysis engine to call a lightweight temporal convolutional autoencoder to calculate analysis data; quantize the analysis data to obtain a quantized model object; and update the quantized model object to obtain machine learning model update data.

[0055] In one embodiment, the data fusion unit 203 is specifically configured to: extract physical model anomaly data and machine learning model anomaly data from the dual-mode anomaly set, and integrate them to obtain anomaly data to be fused; set physical model threshold parameters and machine learning model threshold parameters based on the anomaly data to be fused to obtain threshold configuration data; compare the physical model anomaly data and machine learning model anomaly data according to the threshold configuration data to obtain physical over-threshold identification data and machine over-threshold identification data; perform hierarchical fusion judgment based on the physical over-threshold identification data and machine over-threshold identification data; generate anomaly identification data when both the physical over-threshold identification data and the machine over-threshold identification data indicate over-threshold; generate suspected anomaly identification data when either the physical over-threshold identification data or the machine over-threshold identification data indicates over-threshold; generate normal identification data when neither the physical over-threshold identification data nor the machine over-threshold identification data indicates over-threshold; and merge and output the anomaly identification data, suspected anomaly identification data, and normal identification data to obtain a fusion judgment result.

[0056] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0057] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0058] This invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the computer device may also include various network interfaces, a power supply, a graphics card, etc., to utilize the graphics card's performance to operate the model, such as for inference and training.

[0059] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0060] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. An anomaly detection method for dynamic management of liquid assets, characterized in that, include: Liquid level data and temperature and humidity sampling data of liquid assets are acquired separately, and the liquid level data and temperature and humidity sampling data are preprocessed to obtain time-series status data. The time-series state data is input into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set. The dual-mode anomaly set is hierarchically fused to obtain the fusion judgment result; Based on the fusion determination result, liquid asset identifiers and timestamp identifiers are generated respectively, and then integrated to obtain the final determination result; Based on the final determination result, alarm event information is generated and anomaly detection data is output.

2. The anomaly detection method for dynamic management of liquid assets according to claim 1, characterized in that, The process of acquiring liquid level data and temperature and humidity sampling data of liquid assets, and preprocessing the liquid level data and temperature and humidity sampling data to obtain time-series state data includes: setting the liquid asset node to a low-frequency sampling mode to obtain sampling mode configuration data; collecting liquid level data and temperature and humidity sampling data of liquid assets using a multi-dimensional sensing and data preprocessing unit based on the sampling mode configuration data, and processing them to obtain preprocessed sampling data; and using the preprocessed sampling data to calculate a high-confidence state mean through extended Kalman filtering to obtain time-series state data.

3. The anomaly detection method for dynamic management of liquid assets according to claim 1, characterized in that, The step of inputting the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-model anomaly set includes: obtaining the current temperature from the time-series state data; calculating the saturated vapor pressure of the temperature data using a preset equation to obtain saturated vapor pressure data; calculating the actual water vapor partial pressure in the air based on the saturated vapor pressure data to obtain water vapor partial pressure data; calculating the liquid level using the saturated vapor pressure data and water vapor partial pressure data to obtain the liquid level drop; and calculating the physical model anomaly data using the liquid level drop.

4. The anomaly detection method for dynamic management of liquid assets according to claim 3, characterized in that, The step of calculating the liquid level drop using the saturated vapor pressure data and water vapor partial pressure data includes: obtaining the mass transfer coefficient, liquid surface area, and time increment corresponding to the mass transfer law to obtain mass transfer parameters; calculating the difference between the saturated vapor pressure data and the water vapor partial pressure data to obtain vapor pressure difference data; and calculating the liquid level drop based on the vapor pressure difference data and the mass transfer parameters.

5. The anomaly detection method for dynamic management of liquid assets according to claim 4, characterized in that, The step of calculating the physical model anomaly data using the liquid level drop includes: obtaining the actual liquid level change within a preset time increment based on the time-series state data to obtain the measured drop; determining the corresponding time based on the measured drop, and obtaining the corresponding theoretical drop in the liquid level drop based on the corresponding time; obtaining the historical error standard deviation of the preset physical model engine under normal operating conditions based on the theoretical drop; taking the absolute value of the difference between the measured drop and the theoretical drop, and normalizing it using the historical error standard deviation to obtain the physical model anomaly data.

6. The anomaly detection method for dynamic management of liquid assets according to claim 1, characterized in that, The step of inputting the time-series state data into a preset physical model engine and a preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set further includes: extracting cached data from the time-series state data and organizing it according to a preset sampling order to construct model input data; inputting the model input data into a deployed processing model for reconstruction and outputting reconstruction sequence data; calculating the reconstruction error using the model input data and the reconstruction sequence data and normalizing and square-rooting it to obtain machine learning model anomaly data; inputting the machine learning model anomaly data into a deep learning-based behavior pattern analysis engine to call a lightweight temporal convolutional autoencoder to calculate analysis data; quantizing the analysis data to obtain a quantized model object; and updating the quantized model object to obtain machine learning model update data.

7. The anomaly detection method for dynamic management of liquid assets according to claim 1, characterized in that, The hierarchical fusion of the dual-mode anomaly set to obtain a fusion judgment result includes: extracting physical model anomaly data and machine learning model anomaly data from the dual-mode anomaly set respectively, and integrating them to obtain anomaly data to be fused; setting physical model threshold parameters and machine learning model threshold parameters based on the anomaly data to be fused respectively to obtain threshold configuration data; comparing the physical model anomaly data and the machine learning model anomaly data according to the threshold configuration data to obtain physical over-threshold identification data and machine over-threshold identification data respectively; performing hierarchical fusion judgment based on the physical over-threshold identification data and the machine over-threshold identification data; generating anomaly identification data when both the physical over-threshold identification data and the machine over-threshold identification data indicate over-threshold; generating suspected anomaly identification data when either the physical over-threshold identification data or the machine over-threshold identification data indicates over-threshold; generating normal identification data when neither the physical over-threshold identification data nor the machine over-threshold identification data indicates over-threshold; and merging and outputting the anomaly identification data, suspected anomaly identification data, and normal identification data to obtain a fusion judgment result.

8. An anomaly detection device for dynamic management of liquid assets, characterized in that, include: The data acquisition unit is used to acquire liquid level data and temperature and humidity sampling data of liquid assets respectively, and to perform data preprocessing on the liquid level data and temperature and humidity sampling data to obtain time-series status data. The data analysis unit is used to input the time-series state data into the preset physical model engine and the preset machine learning model engine for parallel analysis to obtain a dual-mode anomaly set. The data fusion unit is used to perform hierarchical fusion of the dual-mode anomaly set to obtain a fusion determination result; The data determination unit is used to generate liquid asset identifiers and timestamp identifiers based on the fusion determination results, and integrate them to obtain the final determination result. The data output unit is used to generate alarm event information based on the final judgment result and output anomaly detection data.

9. A computer device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the anomaly detection method for dynamic management of liquid assets as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the anomaly detection method for dynamic management of liquid assets as described in any one of claims 1 to 7.

Images