Calibration method for inclinometer probe sensor based on intelligent sensing system

Abstract

This invention discloses a method for calibrating inclinometer probe sensors based on an intelligent sensing system, belonging to the field of sensor calibration technology. The method includes placing the inclinometer probe to be calibrated at different spatial tilt angles on a preset multi-attitude calibration platform; synchronously acquiring the sensor's raw data sequences and the platform's baseline attitude ground truth data under each attitude using an intelligent sensing system; performing time-frequency domain feature extraction on the raw data sequences to generate a hybrid feature vector containing signal stability and dynamic response characteristics, and aligning it with the baseline ground truth data; constructing a nonlinear mapping relationship model between the sensor output and the actual attitude based on the correlated data; using this model to perform inverse mapping calculation on the sensor output under any attitude, deriving the deviation characteristics and generating a set of calibration parameters; writing the calibration parameters into a non-volatile memory and verifying the calibration accuracy. This method can improve calibration accuracy and adapt to multi-attitude calibration requirements.

Description

Technical Field

[0001] This invention belongs to the field of sensor calibration technology, specifically a calibration method for inclinometer probe sensors based on intelligent sensing systems. Background Technology

[0002] The calibration of the built-in sensors of the inclinometer probe is a key step in ensuring the accuracy of inclinometer data. In the existing technology, the inclinometer probe to be calibrated is usually placed at different inclination angles on the calibration platform, and the raw data output by the sensor and the true value data of the reference attitude output by the platform are collected. Then, through simple time domain or frequency domain single feature extraction, the mapping relationship between the sensor output and the true attitude is constructed, and calibration parameters are generated and the calibration is completed.

[0003] In existing technical solutions, feature extraction methods are relatively simple, only capturing some signal features of the raw data. This fails to fully reflect the stability and dynamic response of the sensor output signal, resulting in insufficient correlation between the extracted features and the true reference attitude data. Furthermore, the constructed mapping relationships are mostly linear models, which are difficult to adapt to the complex nonlinear relationship between sensor output and the true attitude. This makes it impossible to accurately capture the sensor's deviation characteristics under different attitudes, and the generated calibration parameters lack specificity, making it difficult for the calibrated sensor output accuracy to meet practical application requirements.

[0004] How to fully extract the characteristics of the sensor output signal and achieve accurate correlation with the reference true value data, how to construct a mapping model that can adapt to nonlinear relationships to accurately derive the sensor deviation characteristics, and thus generate reliable calibration parameters, have become urgent problems to be solved in the current calibration process of inclinometer probe sensors. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a calibration method for inclinometer probe sensors based on an intelligent sensing system, including: On a preset multi-attitude calibration platform, inclinometer probes to be calibrated are placed at different spatial tilt angles. The intelligent sensing system synchronously collects the raw data sequences output by the built-in sensors of the inclinometer probes under each attitude and the reference attitude true value data output by the multi-attitude calibration platform. The original data sequence is subjected to time-frequency domain feature fusion extraction processing to generate a fusion feature vector containing signal stability features and dynamic response features, and the fusion feature vector is associated and aligned with the reference attitude ground truth data at the corresponding time. Based on the correlated and aligned hybrid feature vectors and the baseline attitude ground data, a nonlinear mapping relationship model between sensor output and true attitude is constructed. Using the aforementioned nonlinear mapping model, the sensor output of the inclinometer probe to be calibrated under any attitude is inversely mapped and calculated to derive the sensor deviation characteristics, and a set of calibration parameters is generated based on the sensor deviation characteristics. The calibration parameter set is written into the non-volatile memory of the inclinometer probe to be calibrated, and the accuracy of the sensor output after calibration is verified.

[0006] Furthermore, on a preset multi-attitude calibration platform, the inclinometer probes to be calibrated are placed at different spatial tilt angles. The intelligent sensing system synchronously collects the raw data sequences output by the built-in sensors of the inclinometer probes at each attitude, as well as the baseline attitude ground truth data output by the multi-attitude calibration platform, including: The inclinometer probe to be calibrated is fixedly installed on the stage of the multi-attitude calibration platform, and a synchronous clock signal is established between the intelligent sensing system, the inclinometer probe, and the multi-attitude calibration platform. A set of calibration attitude angle sequences covering the measurement range is preset, and the multi-attitude calibration platform is controlled to sequentially adjust the stage to each target attitude in the calibration attitude angle sequence. After the multi-attitude calibration platform stabilizes at each target attitude, the data acquisition unit of the intelligent sensing system continuously acquires the voltage signals or digital signals output by the accelerometer and gyroscope built into the inclinometer probe to form the original data sequence under the current attitude. The angle values ​​output by the high-precision angle encoder of the multi-attitude calibration platform during the acquisition period are recorded synchronously and used as the reference attitude true value data under the current attitude. The raw data sequence collected under each target pose is packaged with the corresponding baseline pose ground truth data to form a calibration data record.

[0007] Further, time-frequency domain feature fusion extraction processing is performed on the original data sequence to generate a hybrid feature vector containing signal stability features and dynamic response features, including: The original data sequence is preprocessed, including outlier removal, moving average filtering, and zero-bias coarsening compensation. Temporal features are extracted from the preprocessed data sequence. The mean, variance, peak-to-peak value, zero-crossing rate, and the value of a specific lag point of the autocorrelation function of the original data sequence are calculated to form a temporal feature sub-vector. The preprocessed data sequence is transformed in the frequency domain to obtain its spectrum. The DC component, the amplitude of the main frequency component, the centroid of the spectrum, the spectrum width, and the energy proportion in the specified frequency band are extracted from the spectrum to form a frequency domain feature vector. Multiple subsequences are extracted from the original data sequence, and the difference between the features of each subsequence and the features of the whole sequence is calculated to evaluate the stability of the signal and form a stability index. The time-domain feature vector, the frequency-domain feature vector, and the stability index are concatenated to generate the hybrid feature vector.

[0008] Further, associating and aligning the hybrid feature vector with the ground truth data of the reference pose at the corresponding time point includes: A timestamp is added to the hybrid feature vector, the timestamp being derived from the synchronization clock signal of the intelligent sensing system; Extract the timestamp information corresponding to the baseline attitude truth data, and interpolate the baseline attitude truth data to a time point that is completely aligned with the timestamp of the hybrid feature vector; Establish data pairs that strictly correspond to timestamps between the hybrid feature vector and the interpolated baseline pose ground truth data; The hybrid feature vectors under the corresponding timestamps are packaged and stored with the baseline pose ground truth data to form a calibration sample dataset for modeling.

[0009] Furthermore, the construction of a nonlinear mapping model between the sensor output and the true attitude based on the correlated and aligned hybrid feature vector and the reference attitude ground truth data includes: The mixed feature vectors in the calibration sample dataset are used as the model input, and the corresponding baseline pose ground value data are used as the model's expected output. A feedforward neural network with multiple hidden layers is constructed as the basic structure of the nonlinear mapping relationship model; The feedforward neural network is trained using the calibration sample dataset, and the network connection weights are adjusted during the training process using the backpropagation algorithm. During training, a portion of the labeled sample data is used as a validation set. The prediction error of the model on the validation set is monitored. Training is stopped when the prediction error no longer decreases to prevent overfitting. The structure and weight parameters of the trained and validated feedforward neural network are fixed as the final nonlinear mapping relationship model.

[0010] Furthermore, using the aforementioned nonlinear mapping model, the sensor output of the inclinometer probe to be calibrated under any attitude is inversely mapped and calculated to derive the sensor deviation characteristics, including: Starting from the input layer of the nonlinear mapping model, calculate the contribution of each neuron in the final pose output; Based on the contribution, several input mixture features that have the greatest impact on the final output are traced back, and the input mixture features correspond to key patterns in the original sensor data sequence; Design a set of virtual ideal sensor outputs, which should be able to map to a set of known, unbiased ideal attitude truth values ​​under the nonlinear mapping relationship model; The raw sensor data collected by the inclinometer probe under actual working posture is processed by time-frequency domain feature extraction to obtain the actual mixed feature vector, which is then input into the nonlinear mapping relationship model to obtain the posture value predicted by the model. The difference sequence between the attitude values ​​predicted by the model and the reference attitude values ​​obtained by other independent measurement methods reflects the comprehensive deviation of the sensor system under the actual working conditions corresponding to the prediction by the nonlinear mapping model, which is the sensor deviation characteristic.

[0011] Further, a set of calibration parameters is generated based on the sensor deviation characteristics, including: Analyze the variation of the sensor's deviation characteristics with attitude, time, or temperature; Based on the aforementioned change pattern, a deviation compensation model is established. The deviation compensation model takes the sensor's original output, temperature, and historical deviation data as inputs and the predicted deviation amount as output. The deviation compensation model is subjected to parameter identification to obtain the coefficient matrix and offset vector in the model; The coefficient matrix, offset vector, structural code of the deviation compensation model, and applicable conditions are collectively encapsulated into the calibration parameter set.

[0012] Further, the calibration parameter set is written into the non-volatile memory of the inclinometer probe to be calibrated, and the accuracy of the sensor output after calibration is verified, including: A data connection is established with the inclinometer probe to be calibrated through the communication interface of the intelligent sensing system. The encapsulated set of calibration parameters is transmitted and written to a specified address in the non-volatile memory of the inclinometer probe according to a preset communication protocol and data format. After the writing is completed, the multi-attitude calibration platform is controlled to place the inclinometer probe in a set of known verification attitudes. This set of verification attitudes should be different from the calibration attitude angle sequence used during modeling. Collect and record the raw data output by the sensor of the calibrated inclinometer probe under the verification attitude; The deviation compensation model in the written calibration parameter set is called to perform real-time online compensation on the original data under the verification attitude, so as to obtain the compensated attitude solution value. Compare the compensated attitude solution value with the known true value of the verification attitude, calculate the maximum error, average error and root mean square error, and verify whether the calibration accuracy meets the preset technical indicators.

[0013] Furthermore, the step of invoking the deviation compensation model in the pre-written calibration parameter set to perform real-time online compensation on the raw data under the verification posture includes: Read the raw data output by the sensor under the current verification posture, as well as the current temperature data output by the temperature sensor; Read the calibration parameter set from the non-volatile memory and load the structural code, coefficient matrix, and offset vector of the deviation compensation model; The current raw data and current temperature data are input into the loaded deviation compensation model, and the model outputs the predicted sensor deviation at the current moment according to its internal calculation rules. Subtract the predicted sensor bias from the current raw data to obtain preliminary compensation data; The preliminary compensation data is input into the standard attitude calculation algorithm of the inclinometer probe to calculate the compensated attitude calculation value.

[0014] Furthermore, after verifying whether the calibration accuracy meets the preset technical specifications, the process also includes: If the calibration accuracy meets the preset technical specifications, a unique calibration certificate is generated for the inclinometer probe. The calibration certificate records the version number of the calibration parameter set, the calibration time, the calibration environmental conditions and the accuracy verification results, and writes the certificate number into the inclinometer probe's memory. If the calibration accuracy does not meet the preset technical specifications, analyze the source of the error and determine whether the calibration parameter set is inapplicable or whether the sensor performance has drifted or malfunctioned. If the calibration parameter set is not applicable, return to adjust the structure of the deviation compensation model or re-perform the calibration data acquisition and modeling process; If there is a sensor performance problem, it is determined that the inclinometer probe needs to be repaired or replaced, the calibration process is terminated and a fault alarm is issued.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The raw data sequence output by the inclinometer probe's built-in sensor undergoes time-frequency domain feature fusion extraction processing to generate a hybrid feature vector containing signal stability features and dynamic response features. This hybrid feature vector is then aligned with the corresponding ground truth data of the baseline attitude. Time-frequency domain feature fusion extraction can simultaneously capture the time-domain stability features and frequency-domain dynamic response features of the raw data. Compared to conventional single-domain feature extraction methods, it can more comprehensively and completely characterize the inherent characteristics of the sensor output signal, avoiding feature loss problems caused by single-domain feature extraction. The alignment of the hybrid feature vector with the ground truth data eliminates the time deviation between features and ground truth, making the correspondence between the two more accurate and avoiding deviations in subsequent model construction caused by feature-ground truth mismatch.

[0016] Based on the correlated and aligned hybrid feature vectors and the baseline attitude ground truth data, a nonlinear mapping model between sensor output and true attitude is constructed. Using this nonlinear mapping model, the sensor output of the inclinometer probe to be calibrated is inversely mapped under any attitude, deriving the sensor deviation characteristics, and generating a set of calibration parameters based on these characteristics. The nonlinear mapping model can accurately fit the nonlinear correlation between sensor output and true attitude. Compared to conventional linear mapping models, it can more accurately reflect the actual working characteristics of the sensor and avoid deviations caused by the inability of linear models to adapt to nonlinear relationships. By deriving the sensor deviation characteristics through inverse mapping calculation, the deviation patterns of the sensor under different attitudes can be comprehensively captured. The generated set of calibration parameters can specifically compensate for sensor deviations. Compared to conventional calibration parameter generation methods, it can effectively improve the consistency and accuracy of the calibrated sensor output and adapt to the calibration requirements under different attitudes. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the inclinometer sensor calibration method based on an intelligent sensing system described in this invention. Figure 2 A flowchart for data acquisition on a multi-pose calibration platform; Figure 3 A flowchart illustrating the association and alignment of feature vectors with ground truth data; Figure 4 This is a graph showing the variation of sensor bias characteristics with attitude angle. Figure 5 This is a comparison chart of sensor errors before and after calibration. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 On a pre-defined multi-attitude calibration platform, inclinometer probes to be calibrated are placed at different spatial tilt angles. An intelligent sensing system synchronously collects raw data sequences from the probe's built-in sensors at various attitudes, along with the baseline attitude ground truth data output by the multi-attitude calibration platform. Time-frequency domain feature extraction is performed on the raw data sequences to generate a hybrid feature vector containing signal stability and dynamic response features. This hybrid feature vector is then correlated and aligned with the corresponding baseline attitude ground truth data. Based on the correlated and aligned hybrid feature vector and the baseline attitude ground truth data, a nonlinear mapping model between the sensor output and the actual attitude is constructed. Using this nonlinear mapping model, inverse mapping calculations are performed on the sensor outputs of the inclinometer probe under arbitrary attitudes to derive sensor deviation characteristics. A calibration parameter set is then generated based on these characteristics. The calibration parameter set is written into the non-volatile memory of the inclinometer probe to be calibrated, and the accuracy of the calibrated sensor output is verified.

[0020] In one embodiment of the present invention, the implementation of the inclinometer sensor calibration method based on an intelligent sensing system involves data acquisition and synchronization steps implemented through a structured process, as described in the following document. Figure 2 The inclinometer probe to be calibrated is securely mounted at the center of the stage of the multi-attitude calibration platform using a mechanical clamp. The main control unit of the intelligent sensing system is connected to both the data interface of the inclinometer probe and the control interface of the multi-attitude calibration platform via a cable, and sends a unified synchronization clock signal to both to establish a precise time reference. In some embodiments, the operator presets a set of calibration attitude angle sequences covering the entire range of the inclinometer probe through the human-machine interface of the intelligent sensing system, such as a pitch angle sequence from -90 degrees to +90 degrees at 10-degree intervals. The main control unit of the intelligent sensing system then sends instructions to the motion controller of the multi-attitude calibration platform, controlling the servo motor of the multi-attitude calibration platform to drive the stage to rotate sequentially to each target attitude in the calibration attitude angle sequence.

[0021] In practical implementation, once the multi-attitude calibration platform moves according to instructions and stabilizes in a target attitude, the data acquisition unit of the intelligent sensing system is triggered. It continuously acquires analog voltage signals output from the triaxial accelerometer and gyroscope built into the inclinometer probe via a high-precision analog-to-digital converter. The sampling frequency is set to 1000 Hz, and acquisition continues for 2 seconds, forming a raw data sequence containing 6000 data points in the current attitude. It can be understood that the synchronous recording process and the data acquisition process are strictly based on the same synchronous clock signal. The intelligent sensing system synchronously records the real-time angle values ​​output by the high-precision optical angle encoder built into the multi-attitude calibration platform during the 2-second acquisition period. These angle values ​​are directly used as the unquestionable reference attitude truth data in the current attitude. In some embodiments, the data processing unit of the intelligent sensing system packages the raw data sequence acquired in each target attitude along with the corresponding reference attitude truth data, attitude identifier, and timestamp into a structured calibration data record and stores it in a local database.

[0022] Optionally, the accuracy of the synchronization clock signal is crucial for data alignment. The intelligent sensing system can use a network based on the IEEE 1588 precise time protocol or a dedicated hardware trigger line to distribute the clock signal, ensuring that the deviation between the sampling time inside the inclinometer probe and the reading time of the angle encoder on the multi-attitude calibration platform is controlled within microseconds. In specific implementations, for each preset calibration attitude, the intelligent sensing system can be configured with a stable dwell time, for example, a 200-millisecond delay after the platform stabilizes before starting data acquisition, to eliminate transient interference caused by mechanical vibration and ensure that the acquired raw data sequence can accurately reflect the sensor output under that static attitude. It is understood that the design of the calibration attitude angle sequence needs to cover all quadrants of the inclinometer probe's expected working space. A typical sequence may include combinations of pitch, roll, and azimuth angles around different axes to ensure that the acquired calibration data record has complete attitude space representation capabilities. Optionally, after each data acquisition, the intelligent sensing system will immediately display the waveform of the currently stored raw data sequence and the corresponding reference attitude true value on the interactive interface, allowing operators to conduct a preliminary and intuitive inspection of the data quality on-site, thereby identifying obvious installation looseness or signal abnormality problems in the early stages of the process.

[0023] In one embodiment of the present invention, time-frequency domain feature extraction processing is performed on the original data sequence and its association and alignment with the reference attitude ground truth data are completed. The process begins with preprocessing the acquired original data sequence. The preprocessing operations include outlier removal based on the Laida criterion, moving average filtering with a window length of 5, and coarse zero-bias compensation for typical zero-bias values ​​provided in the sensor datasheet. In a specific implementation, time-domain feature extraction is performed on the preprocessed data sequence. The arithmetic mean, statistical variance, peak-to-peak value of the difference between the maximum and minimum values, zero-crossing rate of the signal crossing the zero level, and the specific values ​​of the autocorrelation function at hysteresis points 1, 5, and 10 are calculated. These nine calculated values ​​together constitute a nine-dimensional time-domain feature sub-vector. A fast Fourier transform is performed on the preprocessed data sequence to obtain its amplitude spectrum. The amplitude of the DC component at zero frequency, the amplitude of the dominant frequency component corresponding to the largest amplitude value in the amplitude spectrum, the spectral centroid, the spectral width, and the percentage of energy in the 1-10Hz frequency band constitute a five-dimensional frequency-domain feature sub-vector.

[0024] In some embodiments, signal stability is assessed by calculating the subsequence feature difference. M non-overlapping subsequences are uniformly extracted from the original data sequence of length N. The mean feature of each subsequence is calculated, and then the standard deviation between the mean of these M subsequences and the mean of the entire sequence is calculated. This standard deviation is used as a stability index to assess signal stability. In a specific implementation, a nine-dimensional time-frequency domain feature vector, a five-dimensional frequency domain feature vector, and a scalar stability index are concatenated sequentially to generate a fifteen-dimensional hybrid feature vector. (See also...) Figure 3 Each generated hybrid feature vector is appended with a precise timestamp, which is derived from the global synchronization clock signal of the intelligent sensing system and records the absolute time of the starting sampling point of the original data sequence.

[0025] It is understandable that the correlation alignment operation requires a unified time base. The timestamp information corresponding to the baseline attitude ground truth data stream is extracted. The baseline attitude ground truth data is typically recorded at a frequency higher than the sensor sampling rate. Cubic spline interpolation is used to interpolate the baseline attitude ground truth data to discrete time points that are perfectly aligned with the timestamps of all mixed feature vectors. In some embodiments, a strict one-to-one mapping relationship is established between the mixed feature vectors and the interpolated baseline attitude ground truth data. For a mixed feature vector with timestamp , its corresponding baseline attitude ground truth data is the attitude angle ground truth obtained by interpolation at the same timestamp. In a specific implementation, the mixed feature vectors and baseline attitude ground truth data at the corresponding timestamps are packaged and stored to form a structured calibration sample dataset containing tens of thousands of "feature-ground truth" data pairs for subsequent modeling work. Optionally, the calculation of the spectral centroid can be performed using the following formula:

[0026] in: This represents the calculated spectral centroid. This represents the frequency value of the i-th frequency point in the spectrum. Indicates the frequency in the spectrum The corresponding amplitude value, This represents the total number of frequency points in the spectrum.

[0027] Optionally, during the generation of the hybrid feature vector, the value of the autocorrelation function at a specific lag point in the time-domain feature vector can be directly calculated by calling digital signal processing library functions, while the spectral width in the frequency-domain feature vector is defined as the frequency range where the power drops to 3dB from the peak value in the amplitude spectrum. In specific implementation, the entire process of hybrid feature vector generation and data association alignment is encoded into an executable script. After collecting a batch of calibration data records, the intelligent sensing system automatically calls this script to process all data records in batches and output the final calibration sample dataset file. It can be understood that the accuracy of timestamp alignment directly determines the accuracy of the correspondence between the "feature-true value" data pairs. Therefore, the stability of the synchronization clock signal and the selection of the interpolation algorithm are crucial. The intelligent sensing system uses a highly stable temperature-controlled crystal oscillator to generate the clock source and uses the least squares fitting method to perform high-precision timestamp alignment calculations to ensure that the sensor signal mode represented by the hybrid feature vector strictly corresponds to the reference attitude true value in the time dimension.

[0028] In one embodiment of the present invention, a nonlinear mapping model between sensor output and true attitude is constructed. This process uses a mixed feature vector from a calibration sample dataset as model input and the corresponding ground truth attitude data as the model's expected output. For example, a 15-dimensional mixed feature vector is used as the input vector, and the corresponding ground truth attitude angle or a set of Euler angles is used as the output target. In a specific implementation, a feedforward neural network with multiple hidden layers is constructed as the basic structure of the nonlinear mapping model. A typical network structure can be configured as follows: the input layer has 15 neurons, corresponding to the dimension of the mixed feature vector; the first hidden layer contains 30 neurons; the second hidden layer contains 20 neurons; the third hidden layer contains 10 neurons; the number of neurons in the output layer is consistent with the dimension of the ground truth attitude value. If the ground truth attitude value is a single tilt angle, then the output layer has one neuron. In a specific implementation, the feedforward neural network is trained using a complete calibration sample dataset. During training, an error backpropagation algorithm combined with a gradient descent strategy is used to iteratively adjust the connection weights and bias parameters of each layer in the network to minimize the difference between the network's predicted output and the ground truth attitude data.

[0029] In some embodiments, the activation state of each neuron in the hidden layer of a feedforward neural network can be calculated using an activation function, the expression of which is:

[0030] in: This represents the output activation value of the j-th hidden layer neuron. This represents the weight connecting the i-th input feature to the j-th neuron. This represents the value of the i-th input feature. This represents the bias of the j-th neuron. This represents the Sigmoid activation function. During training, a portion of the labeled sample data is used as a validation set. For example, 20% of the sample data is randomly selected from the total dataset to form the validation set. After each training iteration, the prediction error of the model on the validation set is monitored. When the prediction error no longer decreases over several consecutive training cycles, an early stopping mechanism is automatically triggered to terminate training. This measure aims to prevent the model from overfitting to the training data. It can be understood that the structure and weight parameters of the trained and validated feedforward neural network are fixed as the final determined nonlinear mapping model. The model's structure definition, weight matrices of each layer, and bias vectors are serialized and saved as an independent model file for subsequent bias characteristic derivation processes.

[0031] Optionally, the specific algorithm implementation for model training can employ a mature deep learning framework. In the training loop, mean squared error is defined as the loss function, and the Adam optimizer is used to update the network parameters. The initial learning rate is set to 0.001, and the batch size is set to 64. In some embodiments, the training process is monitored by plotting the training set loss curve and the validation set loss curve. Operators can visually observe the loss decline trend and whether overfitting occurs, i.e., the validation set loss increases instead of decreasing in the later stages of training. In specific implementation, the final determined nonlinear mapping model possesses the ability to nonlinearly map from mixed feature vectors to the true pose value. The model file is loaded into the memory of the intelligent sensing system, becoming an executable function module. This module receives new mixed feature vector inputs and directly outputs the predicted pose angle values. It can be understood that the role of the validation set is to provide real-time evaluation of the model's generalization ability. Its samples do not participate in weight updates; they are only used to evaluate the model's predictive performance on unseen data during training, thereby guiding the timely termination of the training process and ensuring that the final nonlinear mapping model neither overfits to the training data nor fails to maintain good pose prediction ability on new sensor data. Optionally, the depth and width of each layer of the feedforward neural network can be adjusted according to the size of the calibration dataset and the sensor noise characteristics. For more complex multi-axis attitude synchronization calibration, the number of neurons in the output layer is increased accordingly to predict multiple attitude angles simultaneously, while the dimension of the input layer remains consistent with the total dimension of the hybrid feature vector extracted from the multi-axis sensor data.

[0032] In one embodiment of the invention, a nonlinear mapping model is used to derive sensor bias characteristics and generate a calibration parameter set. The process begins at the input layer of the nonlinear mapping model, calculating the contribution of each neuron to the final pose output. One implementation involves using gradient backpropagation to calculate the partial derivatives of the output layer neurons' outputs with respect to the activation values ​​of each hidden layer neuron, and then backpropagating layer by layer to the mixed feature vector of the input layer, thereby obtaining the contribution value of each input feature to the final output pose value. In some embodiments, based on the calculated contribution values, several input mixed features with the greatest impact on the final output are traced back. For example, the top five feature dimensions with a total contribution exceeding 80% are identified from the fifteen-dimensional mixed feature vector. These input mixed features correspond to key patterns in the original sensor data sequence, such as energy in a specific frequency band or specific time-domain statistics. In a specific implementation, a set of virtual ideal sensor outputs is designed. Under the nonlinear mapping model, these ideal sensor outputs should be able to map to a set of known, unbiased ideal pose truth values. This set of ideal sensor outputs can be generated through theoretical calculations or by simulating the output of a high-precision reference sensor in an interference-free environment. The corresponding ideal pose truth values ​​are provided by a high-precision calibration platform.

[0033] It is understandable that the derivation of the deviation characteristics requires comparison between actual data and model predictions. The raw sensor data collected by the inclinometer probe under its actual working attitude is processed through time-frequency domain feature extraction to obtain the actual mixed feature vector, which is then input into a trained nonlinear mapping model. The model then outputs a predicted attitude value. In practice, the attitude value predicted by the model is compared with the reference attitude value obtained at the same time point through other independent measurement methods (such as a laser tracker or a higher-precision inertial measurement unit). The difference sequence reflects the comprehensive deviation of the sensor system under the actual working conditions corresponding to the prediction of the nonlinear mapping model. This difference sequence is the derived sensor deviation characteristic. The variation of the sensor deviation characteristics with attitude, time, or temperature is analyzed. By plotting the relationship curves between the deviation value and attitude angle, working time, or sensor temperature, it is observed whether there is a linear, polynomial, or periodic trend. Table 1 shows a simplified example of sensor deviation characteristics varying with attitude angle.

[0034] Table 1: Example Table of Relationship between Sensor Deviation Characteristics and Attitude Angle

[0035] Based on the observed patterns of change, a deviation compensation model is established. This model takes the sensor's raw output, temperature, and historical deviation data as input, and the predicted deviation as output. For example, a model can be established based on the current sensor output value... and ambient temperature To predict the bias using a multinomial model for the independent variable. The model expression is as follows:

[0036] in: This represents the predicted sensor bias. This represents the sensor's current raw output value. This indicates the current temperature of the sensor. This represents the model coefficients to be identified. Optionally, parameter identification can be performed on the deviation compensation model using a series of sensor deviation characteristic data (i.e., deviation values ​​under different conditions) derived from the derivation. ) and the corresponding sensor raw output and temperature The data is solved using optimization algorithms such as the least squares method to obtain the coefficient matrix and offset vector in the deviation compensation model. In some embodiments, the obtained coefficient matrix, offset vector, structural definition code of the deviation compensation model, and information such as the applicable attitude range, temperature range, and sensor model are collectively encapsulated into a structured calibration parameter set file. In specific implementations, the calibration parameter set file is stored in JSON or binary format, and its content includes a model type identifier, coefficient array, effective range field, and version number, facilitating subsequent writing to the inclinometer probe's memory and being read and called by its firmware program. It can be understood that the structure of the deviation compensation model can be selected according to the complexity of the deviation characteristic variation law. For simple linear relationships, a first-order linear model can be used, while for complex nonlinear relationships, a higher-order polynomial or neural network model can be used. The model's structural code defines how to use the coefficients and input data to calculate the final predicted deviation, and is the core of the entire calibration parameter set that can be executed.

[0037] See Figure 4 This is a graph analyzing the sensor deviation characteristics as a function of attitude angle. It shows the pattern of deviation value changing with the true value of attitude angle during the calibration of the inclinometer sensor, and is a core visualization result in the derivation of deviation characteristics. When the attitude angle is close to 0°, the deviation value fluctuates slightly around 0, generally approaching an ideal state. As the attitude angle increases towards ±60°, the deviation value shows a significant non-linear increasing trend, especially at the positive angle (>0°), where the deviation value rapidly deteriorates towards the negative direction, reaching approximately -0.4° at 60°. The curve shows slight fluctuations, reflecting sensor noise and environmental interference. This curve clearly reveals that the sensor deviation is not a constant value, but rather exhibits a quadratic non-linear change with the attitude angle. The deviation increases significantly at large angles, indicating that the compensation effect under extreme attitudes needs to be focused on during calibration, or piecewise modeling should be considered in these areas.

[0038] In one embodiment of the present invention, the process of writing the calibration parameter set into the inclinometer to be calibrated and verifying its accuracy begins by establishing a data connection with the inclinometer through the communication interface of the intelligent sensing system, for example, through a physical connection via RS-485 or CAN bus, and exchanging a predefined handshake protocol to establish a communication session. In a specific implementation, the encapsulated calibration parameter set is transmitted and written to a designated address in the inclinometer's non-volatile memory according to a preset communication protocol and data format. One implementation involves the intelligent sensing system splitting the calibration parameter set file into several data frames conforming to the Modbus-RTU protocol format and sending them sequentially to the slave address of the inclinometer. The inclinometer's firmware receives these data frames and writes them to a specific sector of its on-chip FLASH memory. In practice, after the write operation is completed, the multi-attitude calibration platform is controlled to place the inclinometer probe in a set of known verification attitudes. This set of verification attitudes should be different from the calibration attitude angle sequence used during modeling. For example, if integer angles from -90 degrees to +90 degrees were used during modeling, non-integer angles such as -75 degrees, -15 degrees, and 45 degrees would be used during verification. The raw data output by the sensors of the calibrated inclinometer probe in the verification attitudes is collected and recorded. The collection conditions are consistent with those used during calibration data collection, including the same sampling frequency and collection duration.

[0039] In some embodiments, the deviation compensation model in the pre-written calibration parameter set is invoked to perform real-time online compensation on the raw data under the verification attitude, obtaining the compensated attitude solution value. The process involves reading the raw data output by the sensor under the current verification attitude, as well as the current temperature data output by the temperature sensor integrated inside the inclinometer tube. In a specific implementation, the calibration parameter set is read from a specified address in the non-volatile memory, and the structure code, coefficient matrix, and offset vector of the deviation compensation model are loaded into the running memory of the inclinometer tube. The loading process involves parsing the parameter file header information and initializing the corresponding calculation function pointers and data structures according to the model type identifier. It can be understood that the current raw data and current temperature data are passed as input parameters to the loaded deviation compensation model. The deviation compensation model performs calculations according to its internal calculation rules and outputs the predicted sensor deviation at the current moment. A simplified compensation calculation can be expressed as:

[0040] in: This represents the sensor data after preliminary compensation. This represents the sensor's current raw output data. This represents the raw sensor readings related to the current attitude. Indicates the current temperature. This represents the coefficient matrix in the calibration parameter set. This represents the calculation function of the deviation compensation model. Preliminary compensation data is obtained by subtracting the predicted sensor deviation from the current raw data. This preliminary compensation data is then input into the standard attitude calculation algorithm of the inclinometer probe. This algorithm is typically based on the fusion calculation of accelerometers and gyroscopes, ultimately calculating the compensated attitude solution value.

[0041] Optionally, the process of verifying whether the calibration accuracy meets the preset technical specifications involves comparing the compensated attitude calculation value with the known true value of the verification attitude, and calculating the maximum error, average error, and root mean square error. For example, at five verification attitude points, the calculated average error is 0.05 degrees, the root mean square error is 0.08 degrees, and the maximum error is 0.12 degrees. In specific implementation, if all calculated error indicators are less than the thresholds specified in the preset technical specification document, the calibration accuracy is determined to meet the preset technical specifications. The intelligent sensing system then generates a unique calibration certificate for the inclinometer probe. The calibration certificate records the version number of the calibration parameter set, the calibration time, the calibration environmental conditions, and the accuracy verification results, and writes the certificate number to another independent sector of the inclinometer probe's memory. If the calibration accuracy does not meet the preset technical specifications, the source of the error is analyzed. The intelligent sensing system can determine whether the calibration parameter set is inapplicable or whether the sensor performance has drifted or malfunctioned by comparing the error patterns under different verification attitudes. In some embodiments, if the calibration parameter set is unsuitable, such as when the error exhibits a clear systematic pattern and is strongly correlated with temperature or attitude, the calibration process is reversed, adjusting the structure of the deviation compensation model or re-performing the calibration data acquisition and modeling process. In specific implementations, for sensor performance issues, such as random and large errors, or significant anomalies in the sensor output signal, the inclinometer probe is determined to require repair or replacement. The intelligent sensing system terminates the calibration process and displays a fault alarm message containing a fault code on the human-machine interface. Optionally, the calibration certificate is generated using digital signature technology to ensure its authenticity. The certificate number is bound to the unique serial number of the inclinometer probe and uploaded to a cloud database for archiving and full lifecycle management and traceability.

[0042] See Figure 5This is a comparative analysis chart of sensor errors before and after calibration, visually demonstrating the improvement effect of the deviation compensation model on the measurement error of the inclinometer sensor. Before calibration, the sensor error fluctuated drastically, ranging from -0.5° to 1.0°, and showed a significant increasing trend in error as the absolute value of the angle increased. After calibration, the error was effectively suppressed, with the overall fluctuation range significantly narrowed to between -0.2° and 0.2°, and remained stable across the entire angle range. The ideal error represents the theoretical zero-error baseline, used for visual comparison. The error curve after compensation closely follows the ideal error line, proving that the established deviation compensation model and calibration parameter set are effective, significantly improving the measurement accuracy and stability of the sensor. The compensated error was effectively controlled across the entire angle range, indicating that the calibration parameter set has a wide applicability and reliable performance. By comparing the curves, the maximum error, average error, and root mean square error before and after calibration can be calculated.

[0043] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A calibration method for inclinometer probe sensors based on an intelligent sensing system, characterized in that, The method includes: On a preset multi-attitude calibration platform, inclinometer probes to be calibrated are placed at different spatial tilt angles. The intelligent sensing system synchronously collects the raw data sequences output by the built-in sensors of the inclinometer probes under each attitude and the reference attitude true value data output by the multi-attitude calibration platform. The original data sequence is subjected to time-frequency domain feature fusion extraction processing to generate a fusion feature vector containing signal stability features and dynamic response features, and the fusion feature vector is associated and aligned with the reference attitude ground truth data at the corresponding time. Based on the correlated and aligned hybrid feature vectors and the baseline attitude ground data, a nonlinear mapping relationship model between sensor output and true attitude is constructed. Using the aforementioned nonlinear mapping model, the sensor output of the inclinometer probe to be calibrated under any attitude is inversely mapped and calculated to derive the sensor deviation characteristics, and a set of calibration parameters is generated based on the sensor deviation characteristics. The calibration parameter set is written into the non-volatile memory of the inclinometer probe to be calibrated, and the accuracy of the sensor output after calibration is verified.

2. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 1, characterized in that, The process involves placing the inclinometer probes to be calibrated at different spatial tilt angles on a preset multi-attitude calibration platform. The intelligent sensing system synchronously collects the raw data sequences output by the built-in sensors of the inclinometer probes at each attitude, as well as the baseline attitude ground truth data output by the multi-attitude calibration platform. This includes: The inclinometer probe to be calibrated is fixedly installed on the stage of the multi-attitude calibration platform, and a synchronous clock signal is established between the intelligent sensing system, the inclinometer probe, and the multi-attitude calibration platform. A set of calibration attitude angle sequences covering the measurement range is preset, and the multi-attitude calibration platform is controlled to sequentially adjust the stage to each target attitude in the calibration attitude angle sequence. After the multi-attitude calibration platform stabilizes at each target attitude, the data acquisition unit of the intelligent sensing system continuously acquires the voltage signals or digital signals output by the accelerometer and gyroscope built into the inclinometer probe to form the original data sequence under the current attitude. The angle values ​​output by the high-precision angle encoder of the multi-attitude calibration platform during the acquisition period are recorded synchronously and used as the reference attitude true value data under the current attitude. The raw data sequence collected under each target pose is packaged with the corresponding baseline pose ground truth data to form a calibration data record.

3. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 2, characterized in that, The original data sequence is subjected to time-frequency domain feature fusion extraction processing to generate a hybrid feature vector containing signal stability features and dynamic response features, including: The original data sequence is preprocessed, including outlier removal, moving average filtering, and zero-bias coarsening compensation. Temporal features are extracted from the preprocessed data sequence. The mean, variance, peak-to-peak value, zero-crossing rate, and the value of a specific lag point of the autocorrelation function of the original data sequence are calculated to form a temporal feature sub-vector. The preprocessed data sequence is transformed in the frequency domain to obtain its spectrum. The DC component, the amplitude of the main frequency component, the centroid of the spectrum, the spectrum width, and the energy proportion in the specified frequency band are extracted from the spectrum to form a frequency domain feature vector. Multiple subsequences are extracted from the original data sequence, and the difference between the features of each subsequence and the features of the whole sequence is calculated to evaluate the stability of the signal and form a stability index. The time-domain feature vector, the frequency-domain feature vector, and the stability index are concatenated to generate the hybrid feature vector.

4. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 3, characterized in that, Associating and aligning the hybrid feature vector with the ground truth data of the reference pose at the corresponding time point includes: A timestamp is added to the hybrid feature vector, the timestamp being derived from the synchronization clock signal of the intelligent sensing system; Extract the timestamp information corresponding to the baseline attitude truth data, and interpolate the baseline attitude truth data to a time point that is completely aligned with the timestamp of the hybrid feature vector; Establish data pairs that strictly correspond to timestamps between the hybrid feature vector and the interpolated baseline pose ground truth data; The hybrid feature vectors under the corresponding timestamps are packaged and stored with the baseline pose ground truth data to form a calibration sample dataset for modeling.

5. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 4, characterized in that, The nonlinear mapping model between sensor output and true attitude is constructed based on the correlated and aligned hybrid feature vector and the reference attitude ground truth data, including: The mixed feature vectors in the calibration sample dataset are used as the model input, and the corresponding baseline pose ground value data are used as the model's expected output. A feedforward neural network with multiple hidden layers is constructed as the basic structure of the nonlinear mapping relationship model; The feedforward neural network is trained using the calibration sample dataset, and the network connection weights are adjusted during the training process using the backpropagation algorithm. During training, a portion of the labeled sample data is used as a validation set. The prediction error of the model on the validation set is monitored. Training is stopped when the prediction error no longer decreases to prevent overfitting. The structure and weight parameters of the trained and validated feedforward neural network are fixed as the final nonlinear mapping relationship model.

6. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 5, characterized in that, Using the aforementioned nonlinear mapping model, the sensor output of the inclinometer probe to be calibrated under any attitude is inversely mapped and calculated to derive the sensor deviation characteristics, including: Starting from the input layer of the nonlinear mapping model, calculate the contribution of each neuron in the final pose output; Based on the contribution, several input mixture features that have the greatest impact on the final output are traced back, and the input mixture features correspond to key patterns in the original sensor data sequence; Design a set of virtual ideal sensor outputs, which should be able to map to a set of known, unbiased ideal attitude truth values ​​under the nonlinear mapping relationship model; The raw sensor data collected by the inclinometer probe under actual working posture is processed by time-frequency domain feature extraction to obtain the actual mixed feature vector, which is then input into the nonlinear mapping relationship model to obtain the posture value predicted by the model. The difference sequence between the attitude values ​​predicted by the model and the reference attitude values ​​obtained by other independent measurement methods reflects the comprehensive deviation of the sensor system under the actual working conditions corresponding to the prediction by the nonlinear mapping model, which is the sensor deviation characteristic.

7. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 6, characterized in that, A set of calibration parameters is generated based on the sensor deviation characteristics, including: Analyze the variation of the sensor's deviation characteristics with attitude, time, or temperature; Based on the aforementioned change pattern, a deviation compensation model is established. The deviation compensation model takes the sensor's original output, temperature, and historical deviation data as inputs and the predicted deviation amount as output. The deviation compensation model is subjected to parameter identification to obtain the coefficient matrix and offset vector in the model; The coefficient matrix, offset vector, structural code of the deviation compensation model, and applicable conditions are collectively encapsulated into the calibration parameter set.

8. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 7, characterized in that, The calibration parameter set is written into the non-volatile memory of the inclinometer probe to be calibrated, and the accuracy of the sensor output after calibration is verified, including: A data connection is established with the inclinometer probe to be calibrated through the communication interface of the intelligent sensing system. The encapsulated set of calibration parameters is transmitted and written to a specified address in the non-volatile memory of the inclinometer probe according to a preset communication protocol and data format. After the writing is completed, the multi-attitude calibration platform is controlled to place the inclinometer probe in a set of known verification attitudes. This set of verification attitudes should be different from the calibration attitude angle sequence used during modeling. Collect and record the raw data output by the sensor of the calibrated inclinometer probe under the verification attitude; The deviation compensation model in the written calibration parameter set is called to perform real-time online compensation on the original data under the verification attitude, so as to obtain the compensated attitude solution value. Compare the compensated attitude solution value with the known true value of the verification attitude, calculate the maximum error, average error and root mean square error, and verify whether the calibration accuracy meets the preset technical indicators.

9. The method for calibrating the inclinometer sensor based on an intelligent sensing system according to claim 8, characterized in that, The step of calling the deviation compensation model in the pre-written calibration parameter set to perform real-time online compensation on the raw data under the verification posture includes: Read the raw data output by the sensor under the current verification posture, as well as the current temperature data output by the temperature sensor; Read the calibration parameter set from the non-volatile memory and load the structural code, coefficient matrix, and offset vector of the deviation compensation model; The current raw data and current temperature data are input into the loaded deviation compensation model, and the model outputs the predicted sensor deviation at the current moment according to its internal calculation rules. Subtract the predicted sensor bias from the current raw data to obtain preliminary compensation data; The preliminary compensation data is input into the standard attitude calculation algorithm of the inclinometer probe to calculate the compensated attitude calculation value.

10. The method for calibrating a clinometer sensor based on an intelligent sensing system according to claim 9, characterized in that, After verifying whether the calibration accuracy meets the preset technical specifications, the process also includes: If the calibration accuracy meets the preset technical specifications, a unique calibration certificate is generated for the inclinometer probe. The calibration certificate records the version number of the calibration parameter set, the calibration time, the calibration environmental conditions and the accuracy verification results, and writes the certificate number into the inclinometer probe's memory. If the calibration accuracy does not meet the preset technical specifications, analyze the source of the error and determine whether the calibration parameter set is inapplicable or whether the sensor performance has drifted or malfunctioned. If the calibration parameter set is not applicable, return to adjust the structure of the deviation compensation model or re-perform the calibration data acquisition and modeling process; If there is a sensor performance problem, it is determined that the inclinometer probe needs to be repaired or replaced, the calibration process is terminated and a fault alarm is issued.

Images