Bayesian inference based level measurement confidence assessment method

Abstract

The present application relates to the technical field of measurement credibility evaluation, and particularly relates to a level measurement credibility evaluation method based on Bayesian inference. The technical scheme comprises the following steps: constructing a multi-source observation data set of level measurement, wherein the multi-source observation data set comprises a raw inclination observation sequence from at least one inclination sensor built in the level, and an auxiliary observation sequence from at least one auxiliary sensor integrated or communicatively connected with the level. The present application effectively solves the core technical problems that the measurement data credibility of the existing level is low and the true value is difficult to identify in a complex dynamic environment by constructing an innovative credibility evaluation model based on Bayesian inference. The method combines multi-source sensor data fusion with probability inference, realizes real-time and online credibility quantitative evaluation of the measurement value, and significantly improves the reliability and practicability of level measurement.

Description

Technical Field

[0001] This invention relates to the field of measurement reliability assessment technology, and in particular to a method for assessing the reliability of level measurements based on Bayesian inference. Background Technology

[0002] A level, a precision instrument used to measure the levelness or tilt angle of an object's surface, is widely used in numerous fields such as construction, machinery installation, industrial manufacturing, and electronic equipment calibration. In these applications, the measurement accuracy of the level directly determines the quality and safety of subsequent operations. For example, in high-rise building construction, if the level's measurement error in wall verticality exceeds the allowable range, it may lead to a decrease in the building's structural stability. During the installation of precision machine tools, even minute deviations in tilt angle measurement can affect the machine tool's machining accuracy, potentially causing workpiece scrap. With the continuous improvement of industrial automation and intelligence, various industries have placed higher demands on the reliability and credibility of level measurement results. They not only need to obtain accurate tilt angle measurements but also need to assess the reliability of the measurement data in real time to avoid decision-making errors caused by unreliable data.

[0003] Existing solutions cannot dynamically model the measurement process using multi-source observation data, making it difficult to quantify and evaluate measurement reliability in real time and accurately. On the one hand, traditional levels often rely on a single tilt sensor, making them susceptible to environmental interference such as temperature drift, vibration, shock, and air pressure changes. Furthermore, sensor aging or temporary malfunctions can lead to data anomalies, and there is a lack of an effective mechanism to distinguish between normal noise and abnormal data. On the other hand, existing reliability assessments often employ post-verification or fixed-weight fusion strategies, which cannot dynamically adapt to changes in sensor characteristics and dynamic measurement scenarios (such as equipment movement or sudden interference). They can only provide a binary judgment of pass / fail, failing to offer continuous reliability probability values ​​in the 0-1 range, resulting in the inability to identify unreliable data in a timely manner, potentially leading to subsequent operational quality risks or decision-making errors. Therefore, this application proposes a level measurement reliability assessment method based on Bayesian inference. Summary of the Invention

[0004] The purpose of this invention is to address the problem in the background technology that existing solutions cannot dynamically model the measurement process by combining multi-source observation data, making it difficult to quantify and evaluate the reliability of measurements in real time and accurately. The invention proposes a level measurement reliability evaluation method based on Bayesian inference.

[0005] The technical solution of this invention: a method for evaluating the reliability of level measurements based on Bayesian inference, comprising the following steps:

[0006] S1. Construct a multi-source dataset containing the original observation sequences from the tilt sensor and the observation sequences from the auxiliary sensor;

[0007] S2. Establish a dynamic Bayesian network model to represent the true tilt angle. and binary credibility latent variables As a latent variable, where This indicates credibility. This indicates that the information is unreliable.

[0008] S3. Based on the multi-source dataset, construct a dynamic prior probability distribution for the dynamic Bayesian network model, including the true tilt angle state. Dynamic priors and credibility latent variables a priori ;

[0009] S4. Construct the observation likelihood function , among which when Time observations follow the principle of A Gaussian distribution centered at the center, when The time follows a uniform distribution;

[0010] S5. Recursively derive the posterior probability distribution using Bayesian filtering. ;

[0011] S6. Output the confidence probability value based on the posterior probability distribution. And the optimal tilt angle estimate.

[0012] Optionally, step S1 specifically includes constructing a multi-source observation dataset for level measurements. The multi-source dataset includes an original tilt observation sequence from at least one tilt sensor built into the level, and an auxiliary observation sequence from at least one auxiliary sensor integrated or communicatively connected to the level. The auxiliary sensor is used to collect environmental or state parameters related to measurement reliability.

[0013] The auxiliary sensors include at least two of the following: a triaxial accelerometer, a triaxial gyroscope, a triaxial magnetometer, a barometer, and a temperature sensor;

[0014] The auxiliary observation sequence includes at least two of the following types of data:

[0015] The derived tilt angle data, calculated from the data of the triaxial accelerometer and / or triaxial gyroscope, is used for cross-validation with the original tilt angle observation sequence. Before calculating the derived tilt angle data, the original data collected by the triaxial accelerometer and triaxial gyroscope are preprocessed. The preprocessing process includes removing impulse noise and eliminating zero drift error in the data.

[0016] The resultant acceleration or rate of change of acceleration of the device, calculated from the data of the triaxial accelerometer and / or triaxial gyroscope, is used to characterize whether the device is in a non-quasi-static abnormal motion state. When calculating the resultant acceleration of the device, the root mean square algorithm is used to synthesize the triaxial acceleration data. When calculating the rate of change of acceleration, the ratio of the acceleration difference between adjacent moments to the time interval is used. At the same time, a normal range threshold for the resultant acceleration and the rate of change of acceleration is set. If the threshold is exceeded, it is determined to be an abnormal motion state.

[0017] The temperature data collected by the temperature sensor chip is used to compensate for temperature drift or to determine whether the temperature exceeds the normal operating range. The temperature drift compensation adopts a piecewise linear compensation method, which divides the temperature range into multiple intervals. A linear relationship between temperature and drift is established in each interval, and the corresponding drift compensation amount is determined according to the real-time temperature. When determining whether the temperature exceeds the normal operating range, the normal operating temperature range is set to -10℃ to 60℃. If the temperature exceeds this range, a temperature abnormality prompt is triggered.

[0018] The air pressure data collected by the barometer is used to help determine whether the equipment height has changed drastically, thus introducing measurement interference. When determining whether the equipment height has changed drastically, the height change is calculated by using the air pressure data (using the standard atmospheric pressure and height correspondence formula), and a height change rate threshold (such as 1 m / s) is set. If the height exceeds the threshold, it is determined to be a drastic height change, thus introducing measurement interference.

[0019] Optionally, in step S2, a credibility assessment model based on Bayesian inference is established, which will evaluate the level at time [time]. True tilt state The observed values ​​in the multi-source dataset are treated as latent variables, and a binary latent variable representing the measurement confidence level is defined. ,in Indicates time The observation data is reliable. This indicates that the information is unreliable.

[0020] The credibility assessment model is a dynamic Bayesian network model, which includes a state layer and an observation layer. The state layer consists of the true tilt angle states. and credibility latent variables The observation layer consists of multi-source observation data; there are no connections between nodes within the layer, but connections are established between nodes in different layers through conditional probability; the true tilt state... The dynamic evolution process is modeled by first-order or higher-order Markov processes: The transmission probability is defined by the angle change obtained from the integration of gyroscope data or by a uniform / uniformly accelerated motion model. When the angle change is obtained from the integration of gyroscope data, the trapezoidal integral method is used to integrate the angular velocity data of the gyroscope, and the integration time interval is consistent with the data acquisition frequency of 10ms. When the uniform motion model is used, it is assumed that the actual tilt angle changes uniformly with time, and the rate of change is determined based on the linear fitting results of historical tilt angle data. When the uniformly accelerated motion model is used, the acceleration parameter is calculated by the difference in the angle change between adjacent time periods.

[0021] The credibility hidden variable The evolution is modeled by a two-state Markov chain: The parameters of its state transition matrix are adaptively adjusted based on historical reliability statistics and the abnormal features observed by the current auxiliary sensor. The historical reliability statistics are calculated by statistically analyzing the reliability results of the most recent 100 to 500 times to calculate the frequency of occurrence of reliable and unreliable states. When the auxiliary sensor observes abnormal features, such as temperature exceeding the normal range or equipment in an abnormal motion state, the transition probability from reliable to unreliable state in the state transition matrix is ​​adjusted according to the severity of the abnormal features. The more severe the abnormality, the greater the transition probability. Abnormal features include temperature exceeding the normal range and equipment in an abnormal motion state. The severity includes the magnitude of the temperature exceeding the range and the multiple by which the resultant acceleration exceeds the threshold.

[0022] Optionally, step S4 specifically includes constructing the observation likelihood function of the credibility assessment model. ,in represent The set of all observed data at time t, the likelihood function is constructed as follows: At that time, the observed values ​​are mainly determined by the true tilt angle. And determined by known observation noise, when At that time, the observations were dominated by an anomalous distribution with greater uncertainty;

[0023] The observation likelihood function The specific construction is as follows: when hour,

[0024]

[0025] in, Represents a Gaussian distribution. yes The original observations from the tilt sensor at any given time. It is its nominal noise variance, which is obtained by calculating the standard deviation of the data from 1000-5000 raw observations collected under standard calibration conditions of a level instrument; }(i=1...M) are M auxiliary observations, This auxiliary observation value is at the true dip angle of The expected theoretical value or calibration value is obtained by deriving the expected theoretical value through physical formulas, and the calibration value is obtained by collecting auxiliary observations at different true dip angles and establishing a mapping relationship table between the observations and the true dip angle. It is its corresponding noise variance, calculated in the same way as... Consistent, the standard deviation is calculated by collecting auxiliary observation data under a standard calibration environment;

[0026] when hour,

[0027]

[0028] in, Represents a uniform distribution. and The physical limits that may occur during tilt observation are defined and determined based on the design range of the level. and Defined the first The reasonable physical range in which each auxiliary observation may occur.

[0029] Optionally, in step S4, the abnormal distribution also introduces an adaptive weighting mechanism based on auxiliary sensor data;

[0030] When auxiliary sensors (such as accelerometers) detect strong vibrations or impacts, the corresponding auxiliary observation values ​​{ The uniform distribution range of} The likelihood function is dynamically broadened, and the broadening range is determined according to the intensity of the vibration or impact. The weight of the observed value in the likelihood function is reduced. The weight reduction adopts a linear decreasing method. The greater the intensity of the vibration or impact, the smaller the weight coefficient. This is achieved by multiplying the probability value corresponding to the observed value by the weight coefficient when calculating the likelihood function.

[0031] When the temperature sensor detects a temperature close to or exceeding the rated operating temperature range of the level sensor chip, the tilt sensor reading... Gaussian distribution variance It was replaced by a larger variance that is proportional to the temperature deviation value, even in In this case, the temperature deviation value is the absolute value of the difference between the real-time temperature and the midpoint of the rated operating temperature (25℃). The variance substitution formula is: , where k is the proportionality coefficient and ΔT is the temperature deviation value.

[0032] Optionally, step S5 specifically involves deriving the posterior probability distribution online and in real-time using a Bayesian filtering recursive framework based on the dynamic prior probability distribution and the observed likelihood function. ;

[0033] The recursive framework utilizing Bayesian filtering is specifically implemented using the Lao-Blackwell method based on particle filtering.

[0034] The state space is decomposed into a linear Gaussian subspace and a nonlinear non-Gaussian subspace, where the true tilt angle state is... The subspace it resides in is a linear Gaussian subspace, and the confidence level is a latent variable. The subspace in question is a nonlinear, non-Gaussian subspace; for the true tilt angle state When using Kalman filtering for optimal estimation, the state prediction equation of the Kalman filter is established based on the dynamic evolution model of the true tilt state, and the observation equation is based on the observation likelihood function. The Gaussian distribution relationship is established, and the state estimate and covariance matrix of the Kalman filter are updated in each recursive process; the binary confidence latent variable is... When using particle filtering for importance sampling and estimation, the number of particles is set to 500-2000. The importance sampling distribution is determined based on the prior probability distribution of the confidence latent variable and the observation likelihood function. By calculating the weight of each particle and performing normalization, the posterior probability distribution of the confidence latent variable is obtained. While ensuring estimation accuracy, the computational complexity is significantly reduced, meeting the requirements of real-time processing of the level instrument. The real-time processing time delay is controlled within 50ms, which is achieved by optimizing the calculation steps of Kalman filtering and particle filtering.

[0035] Optionally, the recursive process in step S5 also includes an online learning module for model parameters;

[0036] Based on historical observation data within a sliding time window, the noise variance parameter { in the observation likelihood function is calculated using maximum a posteriori probability estimation or expectation-maximization algorithms. , The state transition probability parameters of the Markov chain containing the latent variables of the confidence level are learned and adaptively updated online to dynamically track changes in sensor characteristics or environmental interference patterns. The sliding time window size is set to 50–200 data points, and the window sliding method is to remove the oldest data point and add the newest data point in each recursive process. When using maximum a posteriori probability estimation, the prior distribution adopts the conjugate prior. The posterior probability distribution is calculated by combining the prior information and the observed data in the window, and the mode of the posterior probability distribution is taken as the estimated value of the parameter. When using the expectation-maximization algorithm, the posterior expectation of the latent variables is treated as missing data, and the expectation step and the maximization step are performed alternately until the parameter estimate converges. The parameter update frequency is set to once every 5–10 recursive processes to avoid excessive parameter fluctuations caused by frequent updates.

[0037] Optionally, in step S6, the measurement confidence probability value at the current moment is calculated based on the posterior probability distribution. Based on a preset confidence threshold, it outputs the confidence assessment result of the current tilt angle measurement value, and also provides the optimal tilt angle state estimate after filtering correction. The reliability assessment result of the output current tilt angle measurement value specifically includes:

[0038] Output a continuous confidence probability value between 0 and 1. The probability value is rounded to three decimal places, making it easier for users to intuitively understand the reliability of the measurement.

[0039] When the probability value is lower than the first preset threshold, a visual warning (such as the indicator light turning red) or an auditory warning is triggered. The first preset threshold is set to 0.5-0.7. The visual warning uses a red LED light that stays on continuously, and the auditory warning uses a buzzer that emits an intermittent sound.

[0040] When the probability value remains below a lower second preset threshold for more than a set time, the current measurement value is determined to be invalid and is marked as an outlier in the data output or replaced by the best estimate from the previous moment. The second preset threshold is set to 0.2-0.4, and the set time is set to 1-5 seconds. When marking outliers, an "ERROR" sign is appended to the output data. When replacing with the best estimate from the previous moment, the best estimate from the previous moment is directly used as the output value for the current moment.

[0041] Optionally, the method further includes a measurement data fusion and output step based on the credibility assessment results:

[0042] Based on the current moment's credibility probability and the estimated value of the optimal tilt angle state Calculate a weighted final output value ;

[0043]

[0044] in, It is the optimal tilt angle state estimate at the current moment after filtering correction. It is the final output value from the previous moment, smoothly transitioning to historical reliable data when confidence is low, suppressing anomalous jumps; in the calculation... Then, the final output values ​​for 5 to 10 consecutive time points are processed by moving average filtering. The window size of the moving average filtering is consistent with the number of data points. The filtered output value is used as the final data presented to the user. At the same time, the confidence probability value, the optimal tilt angle state estimate, the final output value, and whether there are any warning messages are displayed in real time on the data output interface. The data refresh frequency is consistent with the data acquisition frequency.

[0045] Compared with the prior art, this application includes at least one of the following beneficial technical effects:

[0046] By fusing data from tilt sensors and multiple auxiliary sensors, and using Bayesian models for joint inference, the errors of a single sensor under abnormal environments (such as vibration and drastic temperature changes) can be effectively suppressed, providing better state estimates and reliability indicators, and ensuring the accuracy of the output results.

[0047] The prior distribution, likelihood function, and state transition parameters in the model can be dynamically adjusted and learned online based on real-time sensor data, enabling the system to adapt to different working environments (such as different motion states and temperature ranges) and maintain the stability of the evaluation performance.

[0048] It provides continuous probability values ​​as a reliability indicator and has a tiered warning mechanism (such as visual / auditory warnings, data tagging or replacement) so that users can clearly and intuitively understand the reliability of the current measurement results and make correct decisions.

[0049] By employing efficient algorithms such as the Rao-Blackwell particle filter and optimizing the computation process, a complex Bayesian inference process was realized with limited computing resources, meeting the low latency requirements for real-time data processing of the level.

[0050] By employing a credibility-based weighted data fusion output strategy, it is possible to smoothly revert to historical reliable data when unreliable measurements occur, effectively suppressing abnormal data jumps and ensuring the continuity and stability of the output data.

[0051] This invention effectively solves the core technical problem of low reliability and difficulty in identifying the true value of measurement data by existing levels in complex dynamic environments by constructing an innovative reliability assessment model based on Bayesian inference. This method combines multi-source sensor data fusion with probabilistic inference to realize real-time, online quantitative evaluation of the reliability of measurement values, significantly improving the reliability and practicality of level measurements. Detailed Implementation

[0052] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0053] Example 1

[0054] The present invention proposes a level measurement reliability assessment method based on Bayesian inference, and the following is a detailed description of each part of the method.

[0055] I. Construct a multi-source observation dataset for level measurements. The multi-source observation dataset includes raw tilt observation sequences from at least one tilt sensor built into the level, and auxiliary observation sequences from at least one auxiliary sensor integrated with or connected to the level. The auxiliary sensor is used to collect environmental or state parameters related to measurement reliability. The auxiliary sensor includes at least two of the following: a triaxial accelerometer, a triaxial gyroscope, a triaxial magnetometer, a barometer, and a temperature sensor.

[0056] The auxiliary observation sequences include at least two of the following types of data:

[0057] The derived tilt data, calculated from data from triaxial accelerometers and / or triaxial gyroscopes, is used for cross-validation with the original tilt observation sequence. Before calculating the derived tilt data, the original data acquired by the triaxial accelerometers and triaxial gyroscopes are preprocessed. The preprocessing process includes removing impulse noise and eliminating zero drift error. Impulse noise is removed using a sliding window midpoint filtering method with a window size of 3 to 5 data points. Zero drift error is eliminated by acquiring zero-position data for 5 to 10 minutes while the level is stationary, calculating the average value as the zero drift compensation value, and subtracting the real-time acquired data for compensation.

[0058] The resultant acceleration or rate of change of acceleration of the equipment, calculated from data from a triaxial accelerometer and / or a triaxial gyroscope, is used to characterize whether the equipment is in a non-quasi-static abnormal motion state. When calculating the resultant acceleration of the equipment, the root mean square algorithm is used to synthesize the triaxial acceleration data. When calculating the rate of change of acceleration, the ratio of the acceleration difference between adjacent time moments to the time interval is used. At the same time, normal range thresholds for resultant acceleration and rate of change of acceleration are set. If the threshold is exceeded, it is judged as an abnormal motion state.

[0059] Temperature data collected by the temperature sensor chip is used to compensate for temperature drift or to determine whether the temperature exceeds the normal operating range. Temperature drift compensation adopts a piecewise linear compensation method, which divides the temperature range into multiple intervals. A linear relationship between temperature and drift is established in each interval, and the corresponding drift compensation amount is determined based on the real-time temperature. When determining whether the temperature exceeds the normal operating range, the normal operating temperature range is set to -10℃ to 60℃. If the temperature exceeds this range, a temperature abnormality warning is triggered.

[0060] The barometric pressure data collected by the barometer is used to help determine whether the equipment height has changed drastically, thus introducing measurement interference. When determining whether the equipment height has changed drastically, the height change is calculated by the barometric pressure data. The formula corresponding to the standard atmospheric pressure and height is used, and a height change rate threshold of 1 m / s is set. If the height exceeds the threshold, it is determined to be a drastic height change, thus introducing measurement interference.

[0061] In this embodiment, by integrating the raw data from the tilt sensor with data from multiple auxiliary sensors, the limitation of traditional levels relying on a single sensor is overcome. Multi-dimensional acquisition and measurement of environmental and state parameters related to reliability provide rich data support for subsequent reliability assessment. Simultaneously, preprocessing of the raw data from the auxiliary sensors, such as impulse noise removal and zero-drift error elimination, effectively reduces interference components in the raw data, ensuring the accuracy of auxiliary observation data and preventing data distortion from affecting subsequent model calculation results. This provides a reliable data foundation for subsequent cross-validation and anomaly state judgment.

[0062] II. Establish a credibility assessment model based on Bayesian inference. This model will assess the level at time [time]. True tilt state Treating the observations in the multi-source observation dataset as latent variables, we treat the observations as manifest variables and define a binary latent variable to represent the measurement confidence. ,in Indicates time The observation data is reliable. This indicates unreliability; the reliability assessment model in step S2 is a dynamic Bayesian network model, which includes a state layer and an observation layer. The state layer consists of the true tilt angle states. and credibility latent variables Composition: The observation layer consists of multi-source observation data; there are no connections between nodes within the layer, but connections are established between nodes between layers through conditional probability; true dip state. The dynamic evolution process is modeled by first-order or higher-order Markov processes: The transmission probability is defined by the angle change obtained from the integration of gyroscope data or by a uniform / uniformly accelerated motion model. When the angle change is obtained from the integration of gyroscope data, the trapezoidal integral method is used to integrate the angular velocity data of the gyroscope, and the integration time interval is consistent with the data acquisition frequency of 10ms. When the uniform motion model is used, it is assumed that the actual tilt angle changes uniformly with time, and the rate of change is determined based on the linear fitting results of historical tilt angle data. When the uniformly accelerated motion model is used, the acceleration parameter is calculated by the difference in the angle change between adjacent time periods.

[0063] Credibility Latent Variables The evolution is modeled by a two-state Markov chain: The parameters of its state transition matrix are adaptively adjusted based on historical reliability statistics and the abnormal features observed by the current auxiliary sensor. The historical reliability statistics are calculated by statistically analyzing the reliability results of the most recent 100 to 500 times to calculate the frequency of occurrence of reliable and unreliable states. When the current auxiliary sensor observes abnormal features, such as temperature exceeding the normal range or equipment in an abnormal motion state, the transition probability from reliable to unreliable state in the state transition matrix is ​​adjusted according to the severity of the abnormal features. The more severe the abnormality, the greater the transition probability. Abnormal features include temperature exceeding the normal range and equipment in an abnormal motion state. The severity includes the magnitude of the temperature exceeding the range and the multiple by which the resultant acceleration exceeds the threshold. By setting both the true tilt angle state and confidence level as latent variables, and constructing the relationship between the state layer and the observation layer through a dynamic Bayesian network model, the actual scenario in level measurement where "the true state cannot be directly observed and must be inferred from the observation data" can be more closely aligned with reality. A Markov process is used to model the evolution of the true tilt angle state, and the propagation probability is defined by combining gyroscope data or motion models, which can accurately characterize the dynamic change of the true tilt angle over time, adapting to both static and dynamic measurement scenarios. A two-state Markov chain is used to model the confidence level latent variable, and the transition matrix is ​​adaptively adjusted based on historical statistical information and real-time anomaly characteristics, enabling the confidence level assessment to dynamically respond to environmental interference and sensor state changes, improving the flexibility and accuracy of the assessment.

[0064] III. Based on multi-source observation datasets, a dynamic prior probability distribution is constructed for the credibility assessment model, including the true dip angle state. Dynamic priors and credibility latent variables a priori .

[0065] 1. Among them, the true tilt angle state Dynamic priors The construction includes:

[0066] If a first-order Markov process is used for modeling, and the angle change is obtained by integrating gyroscope data to define the propagation probability: firstly, extract from the auxiliary observation sequence... Time to gyroscope angular velocity data at time 1 The change in angle Δθ during this time period is calculated using the trapezoidal integral method. + ) / 2*Δt, where for The gyroscope angular velocity data is collected at any given time, where Δt is the data acquisition time interval, set to 10ms; subsequently, the dynamic prior probability distribution is constructed as a Gaussian distribution. ,in To determine the noise variance of the angle change, 1000 sets of gyroscope angular velocity data were collected under standard calibration conditions (temperature 25℃±2℃, no vibration). The deviation between the integral obtained angle change and the standard angle change was calculated, and the variance of the deviation was taken as the noise variance. The initial value is then dynamically updated via the online learning module for model parameters.

[0067] If a uniform motion model is used to define the propagation probability: based on the true dip angle estimates (obtained from historical filtering results) of the first 50 time moments in the multi-source observation dataset (with a sliding window size of 50), the dip angle change rate is obtained through linear fitting. The fitting formula is:

[0068] ,in For time index, The window start time, For the first The true tilt angle estimate at time; the dynamic prior probability distribution is set as ,in The initial values ​​for the noise variance of the uniform motion model were obtained through a uniform motion simulation experiment under standard calibration conditions, with the level controlled at a known speed. Incline at a constant speed and calculate the rate of change of the fitted tilt angle. and The deviation is taken as the variance of the deviation. The initial value is then dynamically updated by the online learning module for model parameters.

[0069] 2. Latent variables of credibility Prior P( ) Build:

[0070] First, extract the credibility status of the most recent 200 moments (consistent with the sliding time window size in step S7) from the historical credibility assessment results. Count the number of occurrences of the credible state (C=1). The number of occurrences of the untrusted state C=0 Calculate the initial prior probability. , .

[0071] Dynamically adjust the prior probability based on the abnormal features observed by the current auxiliary sensor: if the temperature sensor detects... Temperature at any time Exceeding the normal operating range (-10℃~60℃), and the temperature deviation ΔT=| -25℃|≥10℃ (severe temperature anomaly), then P( =1) is adjusted to P0( =1)*0.5; if the accelerometer detects Time device combined acceleration If the speed exceeds the normal threshold by 2 m / s², and the excess is ≥1.5 times (severe motion abnormality), then P( =1) adjusted to ( =1)*0.6; If multiple abnormal features exist simultaneously, adjust them by multiplication. If severe temperature anomalies and severe motion anomalies exist simultaneously, P( =1)= ( =1)*0.5*0.6, ensuring that the prior probability can reflect the impact of real-time environmental interference on credibility.

[0072] In this embodiment, by combining gyroscope data or motion models to construct a dynamic prior of the true tilt angle state, the prior distribution is no longer a fixed static distribution, but can be dynamically adjusted based on historical states and real-time observations, which better reflects the evolution of the true tilt angle over time and reduces the uncertainty of subsequent posterior probability inference. The prior of the latent confidence variable is dynamically adjusted by combining historical statistical information and real-time anomaly characteristics, avoiding the defect of traditional fixed priors that cannot respond to real-time interference. This allows the prior probability to reflect the impact of the environment or sensor state on confidence in advance, providing a more accurate initial basis for subsequent likelihood function calculation and posterior inference, and improving the overall accuracy of confidence assessment.

[0073] IV. Constructing the Observation Likelihood Function of the Credibility Assessment Model ,in represent The set of all observed data at time t, the likelihood function is constructed as follows: At that time, the observed values ​​are mainly determined by the true tilt angle. And determined by known observation noise, when At that time, the observations are dominated by an outlier distribution with greater uncertainty; the observation likelihood function The specific construction is as follows: when hour,

[0074]

[0075] in, Represents a Gaussian distribution. yes The original observations from the tilt sensor at any given time. This is its nominal noise variance, which is obtained by collecting 1000-5000 raw observations under standard calibration conditions for a level instrument (temperature 25℃±2℃, vibration-free, horizontal placement) and calculating the standard deviation of the data; }(i=1...M) are M auxiliary observations, This auxiliary observation value is at the true dip angle of The expected theoretical value or calibration value is obtained by deriving physical formulas, such as based on the physical relationship between the accelerometer and the tilt angle. The calibration value is obtained by collecting auxiliary observations at different real tilt angles and establishing a mapping table between the observations and the real tilt angle. It is its corresponding noise variance, calculated in the same way as... Consistent, the standard deviation is calculated by collecting auxiliary observation data under a standard calibration environment;

[0076] when hour,

[0077]

[0078] in, Represents a uniform distribution. and The physical limits that may occur during tilt observation are defined and determined according to the design range of the level, such as -90° to 90°. and Defined the first The reasonable physical range of possible auxiliary observation values ​​is determined based on the technical parameters of the auxiliary sensors and the actual application scenario, ensuring that the reasonable physical range covers all possible observation values ​​under normal application. The technical parameters include the accelerometer range of ±16g and the temperature sensor range of -40℃ to 125℃.

[0079] It should be noted that the anomaly distribution also introduces an adaptive weighting mechanism based on auxiliary sensor data; when the auxiliary sensor (such as an accelerometer) detects strong vibrations or impacts, the corresponding auxiliary observation value { The uniform distribution range of} The range is dynamically broadened, and the extent of the broadening is determined based on the intensity of the vibration or impact. For example, if the vibration acceleration value exceeds the normal threshold by 1 time, the range is broadened by 10%; if it exceeds 2 times, the range is broadened by 20%. Furthermore, the weight of the observed value in the likelihood function is reduced. The weight reduction adopts a linear decreasing method. The greater the vibration or impact intensity, the smaller the weight coefficient. The weight coefficient ranges from 0.1 to 1.0. This is achieved by multiplying the probability value corresponding to the observed value by the weight coefficient when calculating the likelihood function.

[0080] When the temperature sensor detects a temperature close to or exceeding the rated operating temperature range of the level sensor chip, the tilt sensor reading... Gaussian distribution variance It was replaced by a larger variance that is proportional to the temperature deviation value, even in In this case, the temperature deviation value is the absolute value of the difference between the real-time temperature and the midpoint of the rated operating temperature (25℃). The variance substitution formula is: , where k is the proportionality coefficient (obtained by fitting experimental data, such as 0.02 / ℃), and ΔT is the temperature deviation value.

[0081] In this embodiment, likelihood functions are constructed for the two states of the confidence latent variable, when When using a Gaussian distribution, it conforms to the characteristics of normal measurement scenarios where "observed values ​​fluctuate around the true value, and noise follows a Gaussian distribution," ensuring the accuracy of likelihood calculation for normal data; when When the likelihood is 0, a uniform distribution is adopted to adapt to the characteristics of "strong randomness and high uncertainty of observation values" in abnormal scenarios, and to avoid the misleading effect of abnormal data on likelihood calculation. An adaptive weighting mechanism based on auxiliary sensor data is introduced to dynamically adjust the distribution range and weight according to abnormal conditions such as vibration and temperature, so that the likelihood function can accurately characterize the characteristics of observation data under different disturbances, further improving the relevance and accuracy of likelihood calculation, and providing reliable likelihood support for posterior probability inference.

[0082] V. Based on the dynamic prior probability distribution and the observation likelihood function, the posterior probability distribution is derived online and in real time using a Bayesian filtering recursive framework. The recursive framework of Bayesian filtering is utilized, specifically employing the Lao-Blackwell transformation method based on particle filtering. The state space is decomposed into a linear Gaussian subspace and a nonlinear non-Gaussian subspace, where the true tilt angle state... The subspace it resides in is a linear Gaussian subspace, and the confidence level is a latent variable. The subspace in question is a nonlinear, non-Gaussian subspace; for the true tilt angle state When using Kalman filtering for optimal estimation, the state prediction equation of the Kalman filter is established based on the dynamic evolution model of the true tilt state, and the observation equation is based on the observation likelihood function. The Gaussian distribution relationship is established, and the state estimate and covariance matrix of the Kalman filter are updated in each recursive process; the binary confidence latent variable is... When using particle filtering for importance sampling and estimation, the number of particles is set to 500-2000. The importance sampling distribution is determined based on the prior probability distribution of the confidence latent variable and the observation likelihood function. By calculating the weight of each particle and performing normalization, the posterior probability distribution of the confidence latent variable is obtained. While ensuring estimation accuracy, the computational complexity is significantly reduced, meeting the requirements of real-time processing of the level instrument. The real-time processing time delay is controlled within 50ms. This is achieved by optimizing the calculation steps of Kalman filtering and particle filtering, using fast matrix operations to reduce the number of particle resampling times.

[0083] It should be noted that the recursive process also includes an online learning module for model parameters;

[0084] Based on historical observation data within a sliding time window, the noise variance parameter { in the observation likelihood function is calculated using maximum a posteriori probability estimation or expectation-maximization algorithms. , The state transition probability parameters of the Markov chain containing the confidence latent variable are learned and adaptively updated online to dynamically track changes in sensor characteristics or environmental interference patterns. The sliding time window size is set to 50–200 data points, and the window sliding method is to remove the oldest data point and add the newest data point in each recursive process. When using maximum a posteriori probability estimation, the prior distribution adopts the conjugate prior, and the prior distribution of the noise variance parameter adopts the inverse gamma distribution. The posterior probability distribution is calculated by combining the prior information and the observed data within the window, and the mode of the posterior probability distribution is taken as the estimated value of the parameter. When using the expectation-maximization algorithm, the posterior expectation of the latent variables is used as the missing data. The latent variables include the true tilt angle state and the confidence latent variable. The expectation step and the maximization step are performed alternately. The expectation step calculates the expectation of the missing data, and the maximization step maximizes the likelihood function based on the complete data until the parameter estimate converges. The parameter update frequency is set to once every 5–10 recursive processes to avoid excessive parameter fluctuations caused by frequent updates.

[0085] A Lao-Blackwell transformation method based on particle filtering is adopted to decompose the state space into linear Gaussian and nonlinear non-Gaussian subspaces. Kalman filtering and particle filtering are used for processing respectively. This method leverages the advantages of Kalman filtering in terms of high estimation accuracy and computational efficiency for linear Gaussian systems, while particle filtering solves the estimation problem of nonlinear non-Gaussian latent variables of confidence. It reduces computational complexity while ensuring estimation accuracy, meeting the requirements of real-time processing of level instruments (delay ≤50ms). An online learning module for model parameters is introduced, which dynamically updates the noise variance and state transition probability based on historical data within a sliding window. This allows the model parameters to track long-term dynamic processes such as sensor aging and changes in environmental interference patterns, avoiding model performance degradation caused by fixed parameters and improving the long-term stability and adaptability of the method.

[0086] VI. Calculate the measurement confidence probability value at the current moment based on the posterior probability distribution. Based on a preset confidence threshold, it outputs the confidence assessment result of the current tilt angle measurement value, and also provides the optimal tilt angle state estimate after filtering correction. Output the confidence assessment result of the current tilt angle measurement, specifically including: outputting a continuous confidence probability value between 0 and 1. The probability value is rounded to three decimal places for easy user understanding of measurement reliability. When the probability value falls below a first preset threshold, a visual warning (e.g., an indicator light turns red) or an auditory warning is triggered. The first preset threshold is set between 0.5 and 0.7, and can be adjusted according to the reliability requirements of the actual application. For example, it is set to 0.7 for high-precision measurement scenarios and 0.5 for general measurement scenarios. The visual warning uses a constantly lit red LED, and the auditory warning uses a buzzer to emit intermittent sounds at a frequency of 1kHz, sounding for 0.5 seconds and then stopping for 0.5 seconds. If the probability value remains below a lower second preset threshold for more than a set duration, a determination is made. The current measurement value is invalid and is marked as an outlier in the data output or replaced by the best estimate from the previous moment. The second preset threshold is set to 0.2-0.4, which is less than the first preset threshold. For example, if the first preset threshold is 0.6, the second preset threshold is set to 0.3. The set duration is set to 1-5 seconds, which is adjusted according to the dynamic nature of the measurement scenario. For example, it is set to 1 second for dynamic measurement scenarios and 5 seconds for static measurement scenarios. When marking outliers, an "ERROR" label is appended to the output data. When replacing with the best estimate from the previous moment, the best estimate from the previous moment is directly used as the output value for the current moment. It outputs a continuous confidence probability value in the 0-1 range. Compared with the traditional binary judgment of "qualified / unqualified", it can more precisely quantify and measure confidence, meeting the differentiated confidence requirements of different scenarios (such as high-precision manufacturing and ordinary construction). It sets two thresholds to trigger different warnings and data processing strategies. When the probability is lower than the first threshold, a warning is triggered to promptly remind the user that the current data confidence is low. When the probability continues to be lower than the second threshold, abnormal or alternative data is marked to avoid the use of unreliable data for subsequent decisions, effectively reducing the operational risks caused by unreliable data. The warning method combines visual and auditory warnings to ensure that users can perceive warning information in a timely manner in different environments, improving the practicality and safety of the method.

[0087] Example 2

[0088] This embodiment, based on embodiment 1, further includes a step of measurement data fusion and output based on the credibility assessment results:

[0089] Based on the current moment's credibility probability and the optimal tilt angle state estimate Calculate a weighted final output value ;

[0090]

[0091] in, It is the optimal tilt angle state estimate at the current moment after filtering correction. This is the final output value from the previous moment, thus enabling the current optimal estimate to be output first when confidence is high, and a smooth transition to historical reliable data when confidence is low, suppressing anomalous jumps; to further improve the stability of the data output, after calculation... Then, the final output values ​​for 5 to 10 consecutive time points are processed by moving average filtering. The window size of the moving average filtering is consistent with the number of data points. The filtered output value is used as the final data presented to the user. At the same time, the confidence probability value, the optimal tilt angle state estimate, the final output value, and whether there are any warning messages are displayed in real time on the data output interface. The data refresh frequency is consistent with the data acquisition frequency (e.g., 100Hz), which facilitates the user to monitor the measurement status in real time.

[0092] It is worth noting that this embodiment uses a weighted fusion formula to calculate the final output value. When the confidence level is high, the current best estimate is used first, while when the confidence level is low, the data smoothly transitions to historical reliable data. This ensures measurement accuracy in high-confidence scenarios and suppresses abnormal jumps in low-confidence scenarios, thus improving the stability of the output data. A moving average filter is applied to the output values ​​at multiple consecutive time points to further reduce the impact of random noise on the output data and ensure data smoothness. Key parameters such as confidence level, best estimate, and warning information are displayed in real time on the output interface, and the data refresh frequency is consistent with the acquisition frequency. This allows users to monitor the measurement status in real time, promptly grasp the data confidence level and equipment operating status, and improve the ease of use and transparency of the method.

[0093] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for evaluating the reliability of level measurements based on Bayesian inference, characterized in that, Includes the following steps: S1. Construct a multi-source dataset containing the original observation sequences from the tilt sensor and the observation sequences from the auxiliary sensor; S2. Establish a dynamic Bayesian network model to represent the true tilt angle. and binary credibility latent variables As a latent variable, where This indicates credibility. This indicates that the information is unreliable. S3. Based on the multi-source dataset, construct a dynamic prior probability distribution for the dynamic Bayesian network model, including the true tilt angle state. Dynamic priors and credibility latent variables a priori ; S4. Construct the observation likelihood function , among which when The time observations follow the principle of A Gaussian distribution centered at the center, when The time follows a uniform distribution; S5. Recursively derive the posterior probability distribution using Bayesian filtering. ; S6. Output the confidence probability value based on the posterior probability distribution. and the optimal tilt angle estimate; In step S2, a credibility assessment model based on Bayesian inference is established. This model will assess the level at time [time]. True tilt state The observed values ​​in the multi-source dataset are treated as latent variables, and a binary latent variable representing the measurement confidence level is defined. ,in Indicates time The observation data is reliable. This indicates that the information is unreliable. The credibility assessment model is a dynamic Bayesian network model, which includes a state layer and an observation layer. The state layer consists of the true tilt angle states. and credibility latent variables The observation layer consists of multi-source observation data; there are no connections between nodes within the layer, but connections are established between nodes in different layers through conditional probability; the true tilt state... The dynamic evolution process is modeled by first-order or higher-order Markov processes: The transmission probability is defined by the angle change obtained from the integration of gyroscope data or by a uniform / uniformly accelerated motion model. When the angle change is obtained from the integration of gyroscope data, the trapezoidal integral method is used to integrate the angular velocity data of the gyroscope, and the integration time interval is consistent with the data acquisition frequency of 10ms. When the uniform motion model is used, it is assumed that the actual tilt angle changes uniformly with time, and the rate of change is determined based on the linear fitting results of historical tilt angle data. When the uniformly accelerated motion model is used, the acceleration parameter is calculated by the difference in the angle change between adjacent time periods. The credibility hidden variable The evolution is modeled by a two-state Markov chain: The parameters of its state transition matrix are adaptively adjusted based on historical credibility statistics and the abnormal features observed by the current auxiliary sensors. The historical credibility statistics are calculated by using the credibility results of the most recent 100 to 500 times to calculate the frequency of occurrence of credible and untrustworthy states.

2. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 1, characterized in that, In step S1, the specific steps include constructing a multi-source observation dataset for level measurements. The multi-source dataset includes an original tilt observation sequence from at least one tilt sensor built into the level, and an auxiliary observation sequence from at least one auxiliary sensor integrated or communicatively connected to the level. The auxiliary sensor is used to collect environmental or state parameters related to the measurement reliability. The auxiliary sensors include at least two of the following: a triaxial accelerometer, a triaxial gyroscope, a triaxial magnetometer, a barometer, and a temperature sensor; The auxiliary observation sequence includes at least two of the following types of data: The derived tilt angle data, calculated from the data of the triaxial accelerometer and / or triaxial gyroscope, is used for cross-validation with the original tilt angle observation sequence. Before calculating the derived tilt angle data, the original data collected by the triaxial accelerometer and triaxial gyroscope are preprocessed. The preprocessing process includes removing impulse noise and eliminating zero drift error in the data. The resultant acceleration or rate of change of acceleration of the device, calculated from the data of the triaxial accelerometer and / or triaxial gyroscope, is used to characterize whether the device is in a non-quasi-static abnormal motion state. When calculating the resultant acceleration of the device, the root mean square algorithm is used to synthesize the triaxial acceleration data. When calculating the rate of change of acceleration, the ratio of the acceleration difference between adjacent moments to the time interval is used. At the same time, a normal range threshold for the resultant acceleration and the rate of change of acceleration is set. If the threshold is exceeded, it is determined to be an abnormal motion state. The temperature data collected by the temperature sensor chip is used to compensate for temperature drift or to determine whether the temperature exceeds the normal operating range. The temperature drift compensation adopts a piecewise linear compensation method, which divides the temperature range into multiple intervals. A linear relationship between temperature and drift is established in each interval, and the corresponding drift compensation amount is determined according to the real-time temperature. When determining whether the temperature exceeds the normal operating range, the normal operating temperature range is set to -10℃ to 60℃. If the temperature exceeds this range, a temperature abnormality prompt is triggered. The air pressure data collected by the barometer is used to help determine whether the equipment height has changed drastically, thereby introducing measurement interference. When determining whether the equipment height has changed drastically, the height change is calculated through the air pressure data, and a height change rate threshold is set. If the height change exceeds the threshold, it is determined to be a drastic height change, thus introducing measurement interference.

3. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 1, characterized in that, Step S4 specifically includes constructing the observation likelihood function of the credibility assessment model. ,in represent The set of all observed data at time t, the likelihood function is constructed as follows: At that time, the observed values ​​are mainly determined by the true tilt angle. And determined by known observation noise, when At that time, the observations were dominated by an anomalous distribution with greater uncertainty; The observation likelihood function The specific construction is as follows: when hour, in, Represents a Gaussian distribution. yes The original observations from the tilt sensor at any given time. It is its nominal noise variance, which is obtained by calculating the standard deviation of the data from 1000-5000 raw observations collected under standard calibration conditions of a level instrument; }(i=1...M) are M auxiliary observations, This auxiliary observation value is at the true dip angle of The expected theoretical value or calibration value is obtained by deriving the expected theoretical value through physical formulas, and the calibration value is obtained by collecting auxiliary observations at different true dip angles and establishing a mapping relationship table between the observations and the true dip angle. It is its corresponding noise variance, calculated in the same way as... Consistent, the standard deviation is calculated by collecting auxiliary observation data under a standard calibration environment; when hour, in, Represents a uniform distribution. and The physical limits that may occur during tilt observation are defined and determined based on the design range of the level. and Defined the first The reasonable physical range in which each auxiliary observation may occur.

4. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 3, characterized in that, In step S4, the abnormal distribution also introduces an adaptive weighting mechanism based on auxiliary sensor data; When the auxiliary sensor detects strong vibration or impact, the corresponding auxiliary observation value { The uniform distribution range of} The likelihood function is dynamically broadened, and the broadening range is determined according to the intensity of the vibration or impact. The weight of the observed value in the likelihood function is reduced. The weight reduction adopts a linear decreasing method. The greater the intensity of the vibration or impact, the smaller the weight coefficient. This is achieved by multiplying the probability value corresponding to the observed value by the weight coefficient when calculating the likelihood function. When the temperature sensor detects a temperature close to or exceeding the rated operating temperature range of the level sensor chip, the tilt sensor reading... Gaussian distribution variance It was replaced by a larger variance that is proportional to the temperature deviation value, even in In this case, the temperature deviation value is the absolute value of the difference between the real-time temperature and the midpoint of the rated operating temperature (25℃). The variance substitution formula is: , where k is the proportionality coefficient and ΔT is the temperature deviation value.

5. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 4, characterized in that, Step S5 specifically involves deriving the posterior probability distribution online and in real-time using a Bayesian filtering recursive framework, based on the dynamic prior probability distribution and the observed likelihood function. ; The recursive framework utilizing Bayesian filtering is specifically implemented using the Lao-Blackwell method based on particle filtering. The state space is decomposed into a linear Gaussian subspace and a nonlinear non-Gaussian subspace, where the true tilt angle state is... The subspace it resides in is a linear Gaussian subspace, and the confidence level is a latent variable. The subspace in question is a nonlinear, non-Gaussian subspace; for the true tilt angle state When using Kalman filtering for optimal estimation, the state prediction equation of the Kalman filter is established based on the dynamic evolution model of the true tilt state, and the observation equation is based on the observation likelihood function. The Gaussian distribution relationship is established, and the state estimate and covariance matrix of the Kalman filter are updated in each recursive process; the binary confidence latent variable is... When using particle filtering for importance sampling and estimation, the number of particles is set to 500-2000. The importance sampling distribution is determined based on the prior probability distribution of the confidence latent variable and the observation likelihood function. By calculating the weight of each particle and performing normalization, the posterior probability distribution of the confidence latent variable is obtained.

6. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 5, characterized in that, The recursive process in step S5 also includes an online learning module for model parameters; Based on historical observation data within a sliding time window, the noise variance parameter { in the observation likelihood function is calculated using maximum a posteriori probability estimation or expectation-maximization algorithms. , The state transition probability parameters of the Markov chain, which is a latent variable of the confidence level, are learned and updated online to dynamically track changes in sensor characteristics or environmental interference patterns. The sliding time window is set to 50-200 data points, and the window sliding method is to remove the oldest data point and add the newest data point in each recursive process. When using maximum a posteriori probability estimation, the prior distribution adopts the conjugate prior. The posterior probability distribution is calculated by combining the prior information and the observed data within the window, and the mode of the posterior probability distribution is taken as the estimated value of the parameter. When using the expectation-maximization algorithm, the posterior expectation of the latent variable is treated as missing data, and the expectation step and the maximization step are performed alternately until the parameter estimate converges.

7. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 6, characterized in that, In step S6, the measurement confidence probability value at the current time is calculated based on the posterior probability distribution. Based on a preset confidence threshold, it outputs the confidence assessment result of the current tilt angle measurement value, and also provides the optimal tilt angle state estimate after filtering correction. ; The reliability assessment result of the output current tilt angle measurement value includes: Output a continuous confidence probability value between 0 and 1. This probability value is rounded to three decimal places. When the probability value is lower than the first preset threshold, a visual warning or an auditory warning is triggered. The first preset threshold is set to 0.5-0.

7. The visual warning uses a red LED light that is constantly on, and the auditory warning uses a buzzer to emit intermittent sounds.

8. The method for evaluating the reliability of level measurements based on Bayesian inference according to claim 7, characterized in that, The method also includes a measurement data fusion and output step based on the credibility assessment results: Based on the current moment's credibility probability value and the estimated value of the optimal tilt angle state Calculate a weighted final output value ; in, It is the optimal tilt angle state estimate at the current moment after filtering correction. It is the final output value from the previous moment, smoothly transitioning to historical reliable data when confidence is low, suppressing anomalous jumps; in the calculation... Then, the final output values ​​for 5 to 10 consecutive time points are subjected to moving average filtering, with the window size of the moving average filter being the same as the number of data points.

Images