Multi-sensor anomaly detection method and device, electronic equipment and storage medium
By deeply integrating cloud models and generalized moment estimation, a fuzzy weighted generalized moment estimation model is constructed, which solves the problem of poor robustness of anomaly detection in multi-sensor systems, realizes efficient anomaly detection of multi-source heterogeneous sensor data, and improves the robustness and reliability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-04-07
AI Technical Summary
In modern ship systems, multi-sensor systems are deployed in a way that makes existing anomaly detection methods less robust when faced with data uncertainty and ambiguity. In particular, when dealing with anomalies such as data fluctuations, system drift, and soft failures, they cannot effectively express and quantify the uncertainty in sensor sampling data, resulting in poor robustness of detection results.
A deep fusion approach combining cloud modeling and generalized moment estimation is adopted. By constructing a fuzzy weighted generalized moment estimation model, the uncertainty modeling capability of cloud modeling for multi-source heterogeneous sensor data and the structural parameter reasoning capability of generalized moment estimation are utilized to achieve consistent reasoning and optimization of target system parameters, thereby improving the robustness of anomaly detection.
It effectively improves the robustness of multi-sensor anomaly detection, significantly enhancing the accuracy and reliability of detection when facing soft failures and system drift, and strengthening the system's ability to identify and interpret anomalies.
Smart Images
Figure CN121808607A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of anomaly detection technology, specifically to a multi-sensor anomaly detection method, device, electronic device, and storage medium. Background Technology
[0002] Modern ship systems deploy numerous heterogeneous multi-source sensors to monitor the real-time operating status, environmental parameters, and structural behavior of various equipment. These multi-sensor systems typically exhibit the following characteristics: diverse data sources and strong heterogeneity; complex system environments and significant interference; high data uncertainty and diverse anomaly types, making them prone to problems such as measurement drift, communication delays, sudden anomalies, gradual degradation, and soft failures. Therefore, to ensure system safety and operational efficiency, it is crucial to fuse and analyze multi-source sensor data to achieve early identification and localization of potential anomalies.
[0003] Anomaly detection methods in related technologies often rely on threshold discrimination, statistical modeling, or supervised learning. They are often weak in modeling the uncertainty and fuzziness of multi-source heterogeneous sensor data. In particular, when faced with anomalies such as data fluctuations, system drift, and soft failures, they cannot effectively express and quantify the uncertainty in sensor sampling data, which can easily lead to poor robustness of detection results. Summary of the Invention
[0004] A multi-sensor anomaly detection method, apparatus, electronic device, and storage medium are provided to improve the robustness of detection results.
[0005] Firstly, a multi-sensor anomaly detection method is provided, comprising the following steps: Real-time acquisition of sampling data from each sensor in the target system; Based on cloud model theory, determine the current total confidence weight of all the sampled data; Based on the current total confidence weight and generalized moment estimate, a fuzzy weighted generalized moment estimation model is constructed. Based on the fuzzy weighted generalized moment estimation model, the optimal parameter estimates for each sensor are determined; The target system is subjected to anomaly detection using the optimal parameter estimation, and the anomaly detection result is determined.
[0006] In some embodiments, determining the current total confidence weight of all the sampled data based on cloud model theory includes: Based on the cloud model theory, the sampling data of each sensor are modeled to generate cloud models for each sensor. Based on the cloud model of each sensor, calculate the membership degree of each data point in each sampled data; All membership degrees are weighted and fused to determine the current total confidence weight of all sampled data.
[0007] In some embodiments, after determining the current total confidence weight for all the sampled data, the method further includes: Obtain the total confidence weight of all the sampled data at a preset historical sampling time, and determine the historical total confidence weight; The current total confidence score is dynamically updated based on the historical total confidence score weight and the current total confidence score weight.
[0008] In some embodiments, constructing a fuzzy weighted generalized moment estimation model based on the current total confidence weight and the generalized moment estimate includes: Based on the generalized moment estimation, the original residual function of the sampled data is constructed; Based on the current total confidence weight and the original residual function, a weighted residual function is constructed; Based on the weighted residual function, a fuzzy weighted generalized moment estimation model is constructed.
[0009] In some embodiments, determining the optimal parameter estimates for each sensor based on the fuzzy weighted generalized moment estimation model includes: The gradient descent method is used to iteratively optimize the fuzzy weighted generalized moment estimation model to determine the optimal parameter estimates for each sensor.
[0010] In some embodiments, the step of using the optimal parameter estimation to perform anomaly detection on the target system and determining the anomaly detection result includes: Based on the optimal parameter estimates of each sensor, the fuzzy weighted residual is determined; Based on the fuzzy weighted residual, the degree of anomaly in the sampling data of each sensor is determined; Based on a preset dynamic discrimination threshold and the degree of anomaly in the sampling data of each sensor, the anomaly detection result of the target system is determined.
[0011] In some embodiments, the step of determining the dynamic discrimination threshold includes: The mean and standard deviation of the anomaly degree of the sampling data of each sensor within a preset nearest window are statistically analyzed. The dynamic discrimination threshold is determined based on the mean and the standard deviation.
[0012] In some embodiments, the step of determining the dynamic discrimination threshold includes: Calculate the mean of the current total confidence weight of the target system within the preset sliding window; The dynamic discrimination threshold is determined based on the mean.
[0013] In some embodiments, the step of determining the dynamic discrimination threshold includes: Based on the degree of anomaly and membership of the sampling data from each sensor, determine the joint distribution of membership and degree of anomaly; The dynamic discrimination threshold is determined based on the joint distribution of membership degree and anomaly degree and the preset normal fluctuation range.
[0014] Secondly, a multi-sensor anomaly detection device is also provided, comprising: The data acquisition module is used to collect sampling data from each sensor in the target system in real time. The confidence weight calculation module is used to determine the current total confidence weight of all the sampled data based on cloud model theory. The model building module is used to construct a fuzzy weighted generalized moment estimation model based on the current total confidence weight and generalized moment estimation. The optimal parameter estimation module is used to determine the optimal parameter estimates for each sensor based on the fuzzy weighted generalized moment estimation model. An anomaly detection module is used to perform anomaly detection on the target system using the optimal parameter estimation and to determine the anomaly detection result.
[0015] Thirdly, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, which, when executed by the processor, implements the steps of the flow control method described in any of the above methods.
[0016] Fourthly, a computer-readable storage medium is also provided, on which a computer program is stored, the computer program being loaded by a processor to perform the steps of the flow control method described in any of the above methods.
[0017] Beneficial effects: By deeply integrating cloud models with generalized moment of moment (GMMo) estimation, the model's ability to model the uncertainty and fuzziness of multi-source heterogeneous sensor sampling data is fully utilized. Furthermore, the structural parameter reasoning capability of GMMo is leveraged to achieve consistent reasoning and optimization of target system parameters, thereby effectively improving the robustness of multi-sensor anomaly detection. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1This is a flowchart illustrating a multi-sensor anomaly detection method provided in some embodiments of this application; Figure 2 This is a schematic diagram of a multi-sensor anomaly detection device provided in some embodiments of this application; Figure 3 These are schematic diagrams of electronic device structures provided in some embodiments of this application. Detailed Implementation
[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0021] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.
[0022] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.
[0023] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not preclude applicability to or configuration to devices performing additional tasks or steps. Furthermore, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more conditions or values may in practice be based on additional conditions or values beyond those conditions.
[0024] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.
[0025] To ensure system safety and operational efficiency, it is essential to fuse and analyze data from multiple sensor sources to enable early identification and localization of potential anomalies.
[0026] Cloud models, being a representation of information uncertainty that integrates fuzziness and stochasticity, are suitable for unified modeling of volatility, fuzzy boundaries, and heterogeneity in multi-source data. However, cloud models lack systematic structural modeling and global parameter estimation capabilities, making them difficult to independently perform dynamic anomaly inference. Meanwhile, the Generalized Method of Moments (GMM) is a parameter estimation method widely used in structural modeling and system consistency inference, suitable for handling non-Gaussian, heteroscedastic, and endogeneity problems. However, traditional GMMs insufficiently model the confidence level of sampled data, making them susceptible to outliers and uncertain sampled data, resulting in poor estimation accuracy and anomaly detection performance.
[0027] This application proposes a multi-sensor anomaly detection method that effectively integrates the uncertainty modeling capability of cloud models with the parameter consistency reasoning capability of generalized method of moments (GMMS) to solve the anomaly detection problem under high uncertainty, high dimensionality, and weak anomaly conditions in multi-sensor systems, thereby improving the system's robustness, reliability, and interpretability in identifying anomalies.
[0028] On the one hand, this embodiment provides a multi-sensor anomaly detection method, applicable to intelligent detection scenarios requiring multi-source data-driven configurations, such as intelligent ship condition assessment, industrial equipment health monitoring, unmanned system condition recognition, intelligent manufacturing monitoring, power equipment monitoring, and medical sensor network monitoring. This application does not impose specific limitations on the application scenarios. Figure 1 As shown, it includes the following steps: Step S100: Collect sampling data from each sensor in the target system in real time.
[0029] Specifically, multi-source heterogeneous sensors deployed in the target system are used to collect sampling data in real time at key operational states of the target system. The target system refers to an intelligent monitoring system that requires the configuration of multi-source heterogeneous sensors to collect various types of sampling data, such as intelligent ship engine rooms, industrial equipment health monitoring, unmanned system status recognition systems, power equipment systems, medical sensor network systems, and intelligent manufacturing systems.
[0030] Multi-source heterogeneous sensors include vibration sensors, temperature sensors, pressure sensors, current / voltage sensors, speed / angle sensors, gas sensors, environmental monitoring sensors, vision sensors, and depth sensors. These sensors are used to collect sampling data related to the key operating states of the target system, such as vibration, temperature, pressure, electrical operating parameters, actuator motion behavior, oxygen content, temperature and humidity, images, and distance. For example, in an intelligent ship engine room system, vibration sensors monitor the mechanical vibration characteristics of the main propulsion system, motors, and pumps; temperature sensors collect the operating temperature of components such as bearings, cylinders, electronic control modules, and lubricating oil; pressure sensors monitor pressure fluctuations in the hydraulic, fuel, and cooling systems; current / voltage sensors collect electrical operating parameters of the main motor and power distribution system; speed / angle sensors acquire the motion behavior of shafts, propellers, or key actuators; and gas and environmental monitoring sensors monitor parameters such as oxygen content, temperature and humidity, and CO2 concentration within the engine room.
[0031] Furthermore, the collected sampling data undergoes preprocessing operations such as time synchronization, format standardization, missing value imputation, and outlier handling to eliminate data inconsistencies caused by communication delays, environmental noise, or equipment drift, ensuring data quality and consistency. This, in turn, ensures the stability and accuracy of subsequent modeling, providing a high-quality data foundation for anomaly detection.
[0032] Step S200: Based on cloud model theory, determine the current total confidence weight of all sampled data.
[0033] Specifically, based on cloud model theory, uncertainty modeling is performed on the sampling data of each sensor to obtain a cloud model. Key parameters of each sensor's sampling data are then calculated based on the cloud model to express the reliability of the sampling data. Next, the total confidence weight of all sensor sampling data at the current moment is calculated using these key parameters. These key sensor parameters include those calculated from recent sampling data, reflecting aspects such as sensor performance and noise levels. Examples include the expected value, entropy, and hyperentropy of the recent sampling data. The expected value reflects the degree to which the sampling data deviates from the baseline value under normal conditions; entropy reflects the fuzziness and uncertainty range of the sampling data; and hyperentropy reflects the stability of the sampling data's volatility.
[0034] Step S300: Based on the current total confidence weight and generalized moment estimation, construct a fuzzy weighted generalized moment estimation model.
[0035] Specifically, the current total confidence weight is introduced into the generalized method of moments (GSM) framework to construct a fuzzy weighted GSM model. By using the confidence weight as a weighting factor for the residuals of the sampled data, the residual function of traditional GSM is optimized. This allows the model to automatically reduce the contribution of low-confidence sampled data during parameter estimation, forming a fuzzy weighted GSM model that is adaptive to fuzziness and uncertainty. Here, low-confidence sampled data refers to outlier data or noise.
[0036] Step S400: Determine the optimal parameter estimates for each sensor based on the fuzzy weighted generalized moment estimation model.
[0037] Specifically, based on the fuzzy weighted generalized moment estimation model, the weighted generalized moment estimation method is used to estimate the parameters of each sensor in the target system. Then, the optimal parameter estimates of each sensor are determined through iterative optimization algorithms such as gradient descent, ensuring that the parameter estimates remain consistent and stable under abnormal data interference, thus providing a reliable model basis for anomaly scoring.
[0038] Step S500: Use the optimal parameter estimation to perform anomaly detection on the target system and determine the anomaly detection results.
[0039] Specifically, a multi-dimensional anomaly scoring strategy is generated from the residuals of the fuzzy weighted generalized moment estimation (GSM) model output by weighted generalized moment estimation (GSM). This strategy dynamically determines whether the current data is anomaly from different dimensions. As a preferred example, by setting dynamic thresholds for the target system under different key operating states, the system is judged to be in an abnormal state based on the anomaly score and the dynamic thresholds. This determines the degree and type of anomaly, supporting a tiered response strategy for the target system. The multi-dimensional anomaly scoring strategy refers to a strategy that comprehensively quantifies and scores the degree of anomaly of the target system from dimensions such as the residuals and confidence levels of the sampled data.
[0040] In this embodiment, by deeply integrating the cloud model with the generalized moment of moment (GMM) estimation, the model's ability to model the uncertainty and ambiguity of multi-source heterogeneous sensor sampling data is fully utilized. The structural parameter reasoning capability of the GMM is also used to achieve consistent reasoning and optimization of the target system parameters. This effectively overcomes the limitations of a single method and enables efficient processing of uncertainty, noise, and ambiguity issues in complex multi-source sensor data. It has significant advantages, especially when facing soft failures and system drift.
[0041] In a preferred embodiment, step S200, based on cloud model theory, determines the current total confidence weight of all sampled data, including: Step S210: Model the sampling data of each sensor based on cloud model theory to generate cloud models for each sensor.
[0042] Specifically, based on cloud model theory, a set of cloud model parameters is generated independently for each sensor based on the historical sampling data of each sensor, thereby establishing a dynamic model that can quantify the uncertainty and randomness of the sampling data of the corresponding sensor.
[0043] For example, for the j-th sensor, The cloud model under normal conditions can be expressed as a triple, namely: (1); in, This represents a one-dimensional normal cloud model along the data dimension of the j-th sensor. Let represent the expected value of the j-th sensor feature under normal conditions, and let represent the central tendency of the sampled data; The entropy represents the feature of the j-th sensor, reflecting the discreteness and ambiguity of the sampled data distribution; The hyperentropy of the j-th sensor feature represents the random fluctuation of entropy, reflecting the stability of uncertainty. It is easy to understand that in this embodiment, the data features of different sensors are used to characterize the data features of the target system in different dimensions.
[0044] As a preferred example, a dataset is constructed using the sampling data from all the collected sensors, denoted as: (2); in, T This is the total number of sampling times; It refers to the number of sensors; It is a moment t The sampling data from all sensors.
[0045] Solving for the expected value of each sensor ,entropy and hyperentropy The process of obtaining cloud model parameters is as follows: Assuming that historical data within a sliding window of length L can accurately characterize the recent normal behavior of the target system, the cloud model parameters calculated based on this can serve as a benchmark for judging whether future sampled data is abnormal. Therefore, this example calculates the cloud model parameters for each of the L consecutive sampled data points: (3); (4); (5); in, L Indicates the length of the sliding window; This represents the hyperentropy scaling factor, with a value ranging from 0.01 to 0.1, and a default value of 0.05. Indicates the start time of the sampling data calculation.
[0046] Step S220: Based on the cloud model of each sensor, calculate the membership degree of each data point in each sampled data.
[0047] Specifically, for any data point in the sampling data of each sensor at any given time, the sampling data is input into the cloud model corresponding to each feature dimension, and the cloud membership degree in each dimension is calculated through nonlinear transformation, which is the confidence degree of the sampling data. The membership degree value is between 0 and 1. The closer the value is to 1, the more the data point conforms to the characteristics of a normal state; conversely, it suggests that there may be anomalies or noise interference in the target system.
[0048] For example, for any sampling time The collected sampling data from all sensors can be represented as: (6); in, This represents the sampling data of the j-th sensor at time t.
[0049] The sampled data is then fed into the cloud model corresponding to each feature dimension, and the cloud membership degree on each dimension is calculated, which is the confidence degree of the sampled data.
[0050] For the j-th sensor, Calculate the cloud membership degree of the sensor at time t: (7); in, Indicates the first j Each sensor at time t Sampling data The smaller the membership degree, the more significantly the sampled data deviates from the expected value, which may be an anomaly or noise.
[0051] Step S230: Weight all membership degrees and merge them to determine the current total confidence weight of all sampled data.
[0052] Specifically, the membership degrees of each sensor at the same time are weighted and fused to generate the current total confidence weight. First, a fusion weight (such as an average weight or an importance-based adjusted weight) is assigned to each sensor. Then, the membership degrees of all sensors are aggregated using a weighted average algorithm, ultimately obtaining a comprehensive confidence value between 0 and 1. This value is dynamically updated over time. When data from some sensors continues to be abnormal, the system automatically reduces their contribution to ensure that the total weight accurately reflects the overall reliability of the data.
[0053] For example, by fusing the membership degrees of all sensors, the total confidence weight of all sampled data at time t can be calculated: (8); (9); in, It is a moment t Total confidence weight of all sampled data; It is the confidence fusion weight of the j-th sensor, with a default value of 1 / n. It can also be set according to prior importance or learned through entropy.
[0054] In this embodiment, by establishing an independent cloud model for each sensor, the uncertainty and random fluctuation characteristics of the sampled data in different dimensions are accurately reflected. By calculating the membership degree of each data point, complex observations are transformed into a unified reliable quantification index. Furthermore, by weighting and fusing the membership degrees of each data point to generate the current total confidence weight, a comprehensive assessment of the quality of multi-source heterogeneous data can be achieved. This process enables the target system to dynamically identify anomalies and mitigate the interference from anomalous sensor data.
[0055] In a preferred embodiment, considering that when the observation data of a certain sensor continuously deviates from the expected value over multiple consecutive time points, its confidence value will decrease accordingly, thereby affecting the weight of that data point in the subsequent generalized moment estimation, the confidence of the sensor will dynamically change over time. To enhance the system's robustness to outliers and noise, this embodiment, after determining the current total confidence weight of all sampled data in step S200, further includes step S600: dynamically adjusting the confidence of the sampled data of each sensor, specifically as follows: Step S610: Obtain the total confidence weight of all sampled data at the preset historical sampling time, and determine the historical total confidence weight.
[0056] Specifically, a historical data queue containing total confidence weights from multiple recent consecutive sampling times is obtained through a sliding window. The length of this queue is determined by a preset window size. Querying this historical data queue yields a set of historical total confidence weight sequences arranged in chronological order. These historical total confidence weight sequences reflect the overall data credibility level of the target system under recent historical operating conditions.
[0057] Step S620: Dynamically update the current total confidence based on the historical total confidence weight and the current total confidence weight.
[0058] Specifically, based on the historical total confidence weight sequence and the current total confidence weight calculated at the current moment, different proportional coefficients are assigned to the historical and current total confidence weights. A smoothing update algorithm is used to dynamically update the current total confidence weight, generating an updated and more robust total confidence. This ensures that the update of the current total confidence can respond promptly to instantaneous changes in data without abrupt changes due to drastic fluctuations at a single point, thus smoothly transitioning and maintaining the system's stability against long-term trends, enhancing the cloud model's adaptability to gradual data changes and temporary disturbances. During the current total confidence update process, if sensor data continuously deviates from the normal range, its confidence is automatically reduced, thereby weakening the impact of abnormal data on the cloud model and enhancing the target system's robustness to noise and soft failures.
[0059] As a preferred example, the process of dynamically updating the confidence level of the j-th sensor at time t using a smooth update algorithm can be represented as: (10); in, This is the smoothing coefficient, which controls the weight ratio between historical confidence levels and current confidence levels. Its value ranges from 0.1 to 0.3, with a default value of 0.2. At any moment t No. j The confidence level of each sensor is used as the current confidence level; It is the first jThe confidence level of each sensor at the previous time t-1. This example uses a smoothing coefficient to control the weight ratio between historical confidence levels and current confidence levels, smoothly fusing and dynamically updating historical information with current weights. This effectively suppresses drastic weight fluctuations caused by instantaneous noise or single outliers, ensuring that the cloud model can react quickly to data fluctuations while maintaining stability in the long term.
[0060] In this embodiment, by smoothly fusing and dynamically updating historical information with current weights, the confidence level of the sensor is adjusted in real time. This ensures that the target system's assessment of data credibility will not overreact to temporary data interference, and will gradually adapt to the long-term trend of sensor performance changes. Thus, while ensuring the sensitivity of anomaly detection, the stability and reliability of weight decision-making are greatly improved, and the adaptive ability of the entire cloud model to cope with complex operating environments is enhanced.
[0061] In a preferred embodiment, step S300 involves constructing a fuzzy weighted generalized moment estimation model based on the current total confidence weight and the generalized moment estimate, including: Step S310: Construct the original residual function of the sampled data based on generalized moment estimation.
[0062] Specifically, based on the generalized method of moments (GMMS), the original residual function of the sampled data (i.e., the sample) is constructed by defining the difference between the theoretical moment conditions and the sample moment conditions, so as to reflect the degree of matching between the actual sampled data and the theoretical model of the system. Nonlinear basis functions are used to map the original sensor data to a high-dimensional feature space to construct the original residual function, thereby capturing the complex data relationship between the target system parameters and the sampled data.
[0063] For example, let the nonlinear basis function be: (11); (12); (13); in, A kernel function is used; in this example, the Radial Basis Kernel Function is used as the nonlinear basis function.
[0064] Step S320: Construct a weighted residual function based on the current total confidence weight and the original residual function.
[0065] Specifically, the confidence weight of the sampling data from each sensor is used as a scaling factor to weight the original residuals. When the confidence of the sampling data from a certain sensor is high, its residual contribution to the model is enhanced; conversely, when the sampling data from a certain sensor may be anomalous or noisy, its low confidence weight will automatically weaken the influence of the residual, thereby achieving adaptive suppression of anomalous data.
[0066] Step S330: Construct a fuzzy weighted generalized moment estimation model based on the weighted residual function.
[0067] Specifically, a fuzzy weighted generalized method of moments (GSM) estimation model is constructed based on a weighted residual function. By substituting the weighted residual function into the objective function framework of GSM, an optimization problem incorporating data credibility information is formed. In minimizing the sum of squared weighted residuals, this model simultaneously considers data uncertainty and model structure consistency, making the parameter estimation process robust to outliers and providing a more reliable model foundation for subsequent anomaly detection.
[0068] As a preferred example, we construct a weighted generalized moment model, with a given p-dimensional parameter vector: (14); Here, the dimension p of the parameter vector represents the total number of free parameters that need to be estimated in the model. The dimension p of the parameter vector is determined by the structural complexity of the system model, the number of multimodal sensing features involved in the fusion, the number of learnable parameters in the confidence weight function, and the setting of the generalized moment condition.
[0069] Required conditions to be met: (15); Traditional generalized method of moments (GSM) estimates assume that each sample participates in the estimation with equal weight, making it impossible to distinguish between outliers and confidence values. To enhance robustness to outliers and sensor confidence fluctuations, this example uses a different approach, increasing the total confidence weight of the sampled data. In the estimation process, the weighted residual function is defined as follows: (16); in, It is the original residual function; The confidence weights are derived from the cloud model. The fuzzy weighted residual is calculated using a weighted residual function; when the sampled data When approaching an anomaly, As the value decreases, its impact on model estimation weakens.
[0070] After weighting the residuals of the sampled data, a fuzzy weighted generalized moment estimation model is constructed, namely: (17); in, It is a positive definite weighted matrix, which is the identity matrix by default. It can also be obtained by estimating the covariance of the sampled data, where T is the number of sampling times. The dimension represents the weighted residual vector or moment condition vector. This dimension reflects the number of statistical consistency constraints that the model must satisfy simultaneously. Physically, it corresponds to the number of error components involved in the estimation, the number of feature residuals, or the error dimension of multimodal sensors. The weighting matrix T is used to describe the covariance structure and relative weights of these residual components, thereby achieving consistent weighting and robust estimation of the model.
[0071] As can be seen, this example demonstrates how adding confidence constraints to the traditional generalized moment estimation automatically weakens the impact of outlier sampled data in model estimation.
[0072] In this embodiment, the confidence weights calculated by the cloud model are incorporated as weighting factors into the residual calculation to construct a fuzzy weighted generalized moment estimation model. This enables the model to distinguish and suppress the interference of outlier data, achieving an organic integration of uncertainty modeling and parameter estimation. This process ensures that the parameter estimation process no longer treats all data equally, but rather differentiates the processing based on the confidence level of each data point. This significantly improves the estimation accuracy and robustness of the model under complex conditions such as noise, drift, and soft failures.
[0073] In a preferred embodiment, step S400, based on the fuzzy weighted generalized moment estimation model, determines the optimal parameter estimates for each sensor, including: The gradient descent method is used to iteratively optimize the fuzzy weighted generalized moment estimation model to determine the optimal parameter estimates for each sensor.
[0074] Specifically, in the iterative optimization of the fuzzy weighted generalized moment estimation model using the gradient descent method, a set of random parameters of the target system are first initialized. In each iteration, the gradient direction of the fuzzy weighted generalized moment estimation model with respect to the current parameters is calculated. This gradient direction, after being modulated by the confidence weight, can accurately reflect the real demand of credible data for parameter correction. Then, all parameter values are updated along the inverse direction of the gradient with a preset learning rate step size. In multiple iterations, the fuzzy weighted generalized moment estimation model gradually suppresses the interference of abnormal data and continuously approaches the characteristics of the real system. Finally, when the model value converges to a stable minimum or reaches the maximum number of iterations, the iteration stops. The parameter vector obtained at this time is the result of the optimal parameter estimation of each sensor under the condition of considering data uncertainty.
[0075] As a preferred example, the iterative optimization process of the fuzzy weighted generalized moment estimation model using gradient descent is as follows: Assume the initial parameters are Set the weight matrix unit array I ,Right now Based on the sampling data and confidence level Calculate the weighted residual function in step S330. Gradient descent is used to refine the fuzzy weighted generalized moment estimation model in step S330. The optimization is iterative, with the ultimate goal of minimizing... To obtain the optimal parameter estimate ,Right now: (18); As another preferred example, the process of achieving optimal parameter estimation through a fuzzy weighted generalized moment estimation model is as follows: Assume the initial parameters are Set the weight matrix unit array I ,Right now Based on the sampling data and confidence level Calculate the weighted residual function in step S330. Gradient descent is used to refine the fuzzy weighted generalized moment estimation model in step S330. Optimization is performed to obtain intermediate parameter estimates. ;estimate The residual covariance matrix is S and order Then based on the sampling data and confidence level Calculate the weighted residual function in step S330. Gradient descent is used to estimate the fuzzy weighted generalized moment estimation model in step S330, resulting in... As the optimal parameter estimate .
[0076] It should be noted that the method of using the fuzzy weighted generalized moment estimation model to achieve optimal parameter estimation is equivalent to the second-order optimization of adaptive weights, which can significantly improve the estimation efficiency and robustness under conditions of heteroscedastic noise, multimodal anomalies and correlation errors. In contrast, the method of obtaining optimal parameter estimation by using only gradient descent for iterative optimization can obtain consistent estimates, but it does not have the optimal ability to handle complex noise and anomalies.
[0077] In this embodiment, the process of minimizing and iteratively optimizing the fuzzy weighted generalized moment estimation model using the gradient descent method can ensure that the parameter estimation results closely match the normal data pattern with high confidence, and automatically weaken the influence of low-quality abnormal data, thereby outputting stable and reliable optimal parameter estimates in complex multi-sensor environments.
[0078] In a preferred embodiment, step S500 involves using optimal parameter estimation to perform anomaly detection on the target system and determining the anomaly detection result, including: Step S510: Determine the fuzzy weighted residual based on the optimal parameter estimates of each sensor.
[0079] Specifically, the true value of the sampled data at the current moment is substituted into the fuzzy weighted generalized moment estimation model. The theoretical predicted value of the sampled data is generated using the optimal parameters. Then, based on the difference between the predicted value and the true value and the corresponding confidence weight, a fuzzy weighted residual that reflects the degree of data deviation and has been calibrated for confidence is obtained. This fuzzy weighted residual contains both model fitting error information and the quality assessment result of the data itself.
[0080] Step S520: Determine the degree of anomaly in the sampling data of each sensor based on the fuzzy weighted residual.
[0081] Specifically, the anomaly score (referred to as the anomaly score) of the sampled data at each time step is quantified by calculating the norm of the fuzzy weighted residuals. The magnitude of this anomaly score represents the degree to which the overall multi-sensor data deviates from the normal state. The higher the anomaly score, the greater the probability of an anomaly. Therefore, the anomaly scoring mechanism in this embodiment effectively integrates the residual information from all sensors and takes into account the confidence differences of each data point, thereby ensuring the comprehensiveness and robustness of the anomaly degree quantification standard.
[0082] As a preferred example, the optimal parameter estimate obtained in step S400 is used. Calculate the fuzzy weighted residual: (19); Define sampling data The degree of anomaly is its weighted residual norm, i.e.: (20); in, Represents the L2 norm; A score representing the degree of anomaly in the sampled data at time t.
[0083] It is evident that this anomaly index takes into account both the deviation between the sampled data and the model. And the reliability of the sampling data. This ensures that low-confidence data, even if it deviates from the model, will not be easily judged as anomaly, thus exhibiting good robustness.
[0084] Step S530: Based on the preset dynamic discrimination threshold and the degree of abnormality of the sampling data of each sensor, determine the abnormality detection result of the target system.
[0085] Specifically, the calculated anomaly level is compared with a preset dynamic discrimination threshold to determine the final anomaly detection result. The dynamic threshold is not a fixed value but is adaptively calculated based on the statistical characteristics of anomaly scores within a recent historical window (such as the mean and standard deviation of historical fuzzy weighted residuals). By comparing the current anomaly score with the threshold, the target system can not only determine whether an anomaly has occurred but also classify the anomaly into different levels such as mild, moderate, or severe based on the extent to which the score exceeds the threshold. It then outputs a comprehensive diagnostic report including anomaly type, level, and confidence level, and may even provide interpretable anomaly detection results and historical anomaly curves, thereby supporting the target system's tiered response strategy.
[0086] In this embodiment, by introducing fuzzy weighted residuals and a generalized moment constraint model, not only can the robustness of parameter estimation be improved, but also the fault tolerance to abnormal data can be enhanced. Through weighted residuals, outliers and noisy sampling data can be adaptively adjusted, thereby avoiding interference from abnormal data on the system model. This significantly improves the accuracy of multi-sensor anomaly detection, avoids false alarms or missed alarms caused by fixed thresholds, and enables precise alarm and classification of the system based on the degree of anomaly. Simultaneously, through interpretable anomaly levels and a dynamic threshold mechanism, the credibility and interpretability of multi-sensor anomaly detection results are enhanced, allowing the target system to intelligently adapt to more complex operating conditions, facilitating engineers to quickly locate problems and analyze the fault development process.
[0087] In a preferred embodiment, the step of determining the dynamic discrimination threshold in step S530 includes: Step S531: Calculate the mean and standard deviation of the anomaly degree of the sampling data of each sensor within the preset nearest window.
[0088] Specifically, the target system continuously records the anomaly scores of the sampling data from each sensor at multiple consecutive sampling times within a preset nearest window, forming an anomaly sequence. The length of this anomaly sequence is determined by the preset window size. Statistical characteristics of this historical sequence are calculated, such as the arithmetic mean of all anomaly scores within the preset nearest window, and the standard deviation, a measure of the degree of fluctuation of these scores around the mean. These two statistical measures together describe the normal fluctuation range of the target system under recent operating conditions.
[0089] Step S532: Determine the dynamic discrimination threshold based on the mean and standard deviation.
[0090] Specifically, a weighted sum of the mean and standard deviation is used to determine a dynamic discrimination threshold. This threshold adaptively follows the historical fluctuation characteristics of the system under normal conditions, thereby more accurately distinguishing between normal fluctuations and genuine abnormal events, and improving the adaptability and accuracy of anomaly detection. The coefficients of the mean and standard deviation can be flexibly adjusted according to different requirements for false alarm and false negative rates.
[0091] For example, given a window length of L, the mean and standard deviation of the outlier of the most recent L sampled data are calculated as follows: (twenty one); (twenty two); The dynamic discrimination threshold can then be expressed as: (twenty three); in, It is a coefficient for controlling sensitivity, with a value between 1 and 3, and a preferred value of 2.
[0092] In another preferred embodiment, the step of determining the dynamic discrimination threshold in step S530 includes: Step S533: Calculate the mean of the current total confidence weight of the target system within the preset sliding window.
[0093] Specifically, the target system continuously records the total confidence weight of the sampling data from each sensor at each consecutive sampling time within a preset recent window, forming a time series. Then, the arithmetic mean of all weight values in this time series is calculated to obtain the mean of the current total confidence weight, which represents the overall data confidence level of the system in the recent period, thus reflecting the stability of the data quality of the target system during normal operation.
[0094] Step S534: Determine the dynamic discrimination threshold based on the mean.
[0095] Specifically, an inverse adjustment relationship is established between the mean of the current total confidence weight and the dynamic discrimination threshold in the target system. When the mean is high, it indicates that the overall credibility of recent data is good and the target system is operating stably. At this time, the dynamic discrimination threshold will be lowered accordingly to improve detection sensitivity. Conversely, when the mean is low, it indicates that the data itself has greater uncertainty or interference. At this time, the dynamic discrimination threshold will be appropriately increased to enhance the system's anti-interference ability, avoid false alarms, and enable the target system to automatically reduce its tolerance adaptive adjustment ability when the total confidence is low (i.e., low confidence).
[0096] For example, calculate the current average confidence level of the target system: (twenty four); in, This represents the membership degree of each sensor in the current target system.
[0097] The dynamic threshold can then be set as follows: (25); in, It is a constant used to adjust parameters; It is a small constant that prevents division by zero.
[0098] In another preferred embodiment, the step of determining the dynamic discrimination threshold in step S530 includes: Step S535: Determine the joint distribution of membership degree and anomaly degree based on the anomaly degree and membership degree of the sampling data of each sensor.
[0099] Specifically, the target system collects the membership degree and anomaly score of each sensor data at each sampling time within a preset historical time period, and performs statistical analysis on the membership degree and anomaly score of each sensor data to determine the joint distribution of membership degree and anomaly degree, thereby establishing a probabilistic model that can describe what anomaly degree corresponds to what membership degree level.
[0100] Step S536: Determine the dynamic discrimination threshold based on the joint distribution of membership degree and anomaly degree and the preset normal fluctuation range.
[0101] Specifically, based on the joint distribution, the range of anomaly scores for the target system under normal operating conditions (i.e., when membership is high) is identified, and this range is defined as a preset normal fluctuation range. A dynamic discrimination threshold is defined as the upper boundary of this normal fluctuation range. When the anomaly score calculated from new sampled data exceeds this boundary, it is judged as an anomaly. This method allows the dynamic discrimination threshold to consider not only the absolute magnitude of the anomaly but also the credibility of the current data, achieving more refined and context-adaptive anomaly judgment. Therefore, this embodiment, through empirical distribution learning, enables the target system to maintain high detection sensitivity under high credibility, greatly improving the accuracy of anomaly detection and the rationality of decision-making in complex and uncertain environments.
[0102] In summary, compared with existing anomaly detection technologies, the multi-sensor anomaly detection method proposed in this application has at least the following beneficial effects, specifically including: 1. Improved the robustness and accuracy of anomaly detection.
[0103] This application introduces a cloud model to model the fuzziness and uncertainty of each sensor data, quantifies the reliability of the sampled data, and embeds it as a weighting factor into the generalized moment estimation process. This effectively suppresses the interference of outliers, sensor noise, and weak quality data on the estimation results, making the system more robust, especially when facing anomalies such as soft failure, gradual drift, and boundary ambiguity.
[0104] 2. A more expressive fuzzy moment conditional model was constructed.
[0105] Compared to the traditional generalized method of moments (GMMS) which only uses fixed moment conditions for residual constraints, the cloud model and data estimation joint mechanism proposed in this application integrates the membership function of the cloud model into the moment function construction, thereby achieving joint modeling of fuzziness and structural consistency. This gives anomaly identification multiple advantages, including dynamism, adaptability, and sensitivity to the features of sampled data.
[0106] 3. A dynamic modeling and updating mechanism for sensor data reliability has been implemented.
[0107] Based on the statistical characteristics of cloud models, this application can continuously monitor and adaptively evaluate the volatility, abnormal trends, and quality variations in sensor sampling data, thereby dynamically adjusting the weight structure of the detection model and having stronger online learning capabilities and the ability to adapt to environmental changes.
[0108] 4. It has good versatility and engineering deployment value.
[0109] The method described in this application does not rely on a specific data distribution and does not require prior labels. It is applicable to various types of multi-source heterogeneous sensor systems and can be widely used in a variety of engineering fields such as intelligent manufacturing, ship operation status monitoring, unmanned systems, power equipment, and medical sensor networks. It has good prospects for practical application and promotion.
[0110] On the other hand, such as Figure 2 As shown, this embodiment provides a multi-sensor anomaly detection device, including: The data acquisition module 201 is used to acquire sampling data from each sensor in the target system in real time; The confidence weight calculation module 202 is used to determine the current total confidence weight of all sampled data based on cloud model theory. Model building module 203 is used to build a fuzzy weighted generalized moment estimation model based on the current total confidence weight and generalized moment estimation; The optimal parameter estimation module 204 is used to determine the optimal parameter estimates for each sensor based on the fuzzy weighted generalized moment estimation model. The anomaly detection module 205 is used to perform anomaly detection on the target system using optimal parameter estimation and to determine the anomaly detection result.
[0111] This embodiment also provides an electronic device, such as... Figure 3 As shown, it includes a memory and a processor. In a specific example, the memory stores a computer program, which, when executed by the processor, implements the method of any of the above embodiments.
[0112] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to execute the arrangement in any of the methods described above.
[0113] In the embodiments of this application, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.
[0114] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0115] The foregoing has provided a detailed description of a multi-sensor anomaly detection method, apparatus, electronic device, and storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A multi-sensor anomaly detection method, characterized in that, Includes the following steps: Real-time acquisition of sampling data from each sensor in the target system; Based on cloud model theory, determine the current total confidence weight of all the sampled data; Based on the current total confidence weight and generalized moment estimate, a fuzzy weighted generalized moment estimation model is constructed. Based on the fuzzy weighted generalized moment estimation model, the optimal parameter estimates for each sensor are determined; The target system is subjected to anomaly detection using the optimal parameter estimation, and the anomaly detection result is determined.
2. The multi-sensor anomaly detection method according to claim 1, characterized in that, The determination of the current total confidence weight for all sampled data based on cloud model theory includes: Based on the cloud model theory, the sampling data of each sensor are modeled to generate cloud models for each sensor. Based on the cloud model of each sensor, calculate the membership degree of each data point in each sampled data; All membership degrees are weighted and fused to determine the current total confidence weight of all sampled data.
3. The multi-sensor anomaly detection method according to claim 2, characterized in that, After determining the current total confidence weight for all the sampled data, the method further includes: Obtain the total confidence weight of all the sampled data at a preset historical sampling time, and determine the historical total confidence weight; The current total confidence score is dynamically updated based on the historical total confidence score weight and the current total confidence score weight.
4. The multi-sensor anomaly detection method according to claim 1, characterized in that, The construction of a fuzzy weighted generalized moment estimation model based on the current total confidence weight and generalized moment estimation includes: Based on the generalized moment estimation, the original residual function of the sampled data is constructed; Based on the current total confidence weight and the original residual function, a weighted residual function is constructed; Based on the weighted residual function, a fuzzy weighted generalized moment estimation model is constructed.
5. The multi-sensor anomaly detection method according to claim 1, characterized in that, The determination of the optimal parameter estimates for each sensor based on the fuzzy weighted generalized moment estimation model includes: The gradient descent method is used to iteratively optimize the fuzzy weighted generalized moment estimation model to determine the optimal parameter estimates for each sensor.
6. The multi-sensor anomaly detection method according to claim 2, characterized in that, The step of using the optimal parameter estimation to perform anomaly detection on the target system and determining the anomaly detection result includes: Based on the optimal parameter estimates of each sensor, the fuzzy weighted residual is determined; Based on the fuzzy weighted residual, the degree of anomaly in the sampling data of each sensor is determined; Based on a preset dynamic discrimination threshold and the degree of anomaly in the sampling data of each sensor, the anomaly detection result of the target system is determined.
7. The multi-sensor anomaly detection method according to claim 6, characterized in that, The step of determining the dynamic discrimination threshold includes: The mean and standard deviation of the anomaly degree of the sampling data of each sensor within a preset nearest window are statistically analyzed. The dynamic discrimination threshold is determined based on the mean and the standard deviation.
8. The multi-sensor anomaly detection method according to claim 6, characterized in that, The step of determining the dynamic discrimination threshold includes: Calculate the mean of the current total confidence weight of the target system within the preset sliding window; The dynamic discrimination threshold is determined based on the mean.
9. The multi-sensor anomaly detection method according to claim 6, characterized in that, The step of determining the dynamic discrimination threshold includes: Based on the degree of anomaly and membership of the sampling data from each sensor, determine the joint distribution of membership and degree of anomaly; The dynamic discrimination threshold is determined based on the joint distribution of membership degree and anomaly degree and the preset normal fluctuation range.
10. A multi-sensor anomaly detection device, characterized in that, include: The data acquisition module is used to collect sampling data from each sensor in the target system in real time. The confidence weight calculation module is used to determine the current total confidence weight of all the sampled data based on cloud model theory. The model building module is used to construct a fuzzy weighted generalized moment estimation model based on the current total confidence weight and generalized moment estimation. The optimal parameter estimation module is used to determine the optimal parameter estimates for each sensor based on the fuzzy weighted generalized moment estimation model. An anomaly detection module is used to perform anomaly detection on the target system using the optimal parameter estimation and to determine the anomaly detection result.
11. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method as described in any one of claims 1 to 9.
12. A computer-readable storage medium, characterized in that, It stores a computer program, which is loaded by a processor to perform the steps of the method as described in any one of claims 1 to 9.