Inclinometer data processing and analyzing method based on magnet ring-encoder collaborative calibration

Abstract

The invention discloses an inclinometer data processing and analyzing method based on magnet ring-encoder collaborative calibration, and relates to the technical field of geotechnical engineering safety monitoring. A magnetic ring is arranged in the inclinometer tube, a reference depth sequence is established, and an encoder distance measurement model is initialized; in the probe lowering process, the measured depth and the reference depth are compared in real time by triggering the magnetic ring, the distance measurement coefficient is dynamically corrected, a depth partition distance measurement factor is generated, and intelligent diagnosis of the health state of the magnetic ring is synchronously achieved; a machine learning error prediction model is introduced, future distance measurement errors are predicted according to the characteristics of environment temperature and accumulative operation duration, feedforward compensation is carried out on distance measurement factors, anti-drift distance measurement factors are generated to control a probe to carry out forward and reverse inclination angle collection, and forward and reverse inclination angle data are fused to obtain a distance measurement result. And calculating the horizontal displacement of each depth point by combining the anti-drifting ranging factor and adopting a segmented accumulation algorithm, and outputting a monitoring result. According to the invention, the precision, long-term stability and system reliability of depth positioning and displacement calculation are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering safety monitoring technology, and more specifically, to a method for processing and analyzing inclinometer data based on magnetic ring-encoder collaborative calibration. Background Technology

[0002] Deep horizontal displacement monitoring is a crucial link in the safety monitoring and geological disaster early warning of major geotechnical engineering projects such as deep foundation pit excavation, high slope treatment, and soft soil foundation treatment. Its measurement accuracy directly affects the assessment of the project's safety status and the effectiveness of risk prevention and control. Currently, commonly used inclinometer monitoring methods mainly include two categories: mobile inclinometers and fixed inclinometers, both of which have certain technical limitations. While mobile inclinometers offer high measurement accuracy and low equipment cost, they rely on manual operation for probe lifting, data acquisition, and processing, resulting in low efficiency and difficulty in achieving real-time monitoring. This poses a dual challenge of safety and timeliness, especially when operating in harsh environments or remote areas. Fixed inclinometers, although achieving automated monitoring, require multiple sensors deployed at different depths within the inclinometer tube. The greater the hole depth, the higher the equipment cost. Furthermore, because the sensors are fixedly installed, bidirectional observation is impossible, making it difficult to effectively eliminate zero-drift errors and ensuring long-term monitoring accuracy.

[0003] In recent years, winch-type automated inclinometers have attempted to combine the advantages of mobile and fixed systems to achieve automated lifting and data acquisition of a single probe. However, these devices primarily rely on encoders for step control in depth positioning. Affected by factors such as wire rope tension and steering wheel slippage, the cumulative error increases significantly with hole depth, typically reaching 1–2 cm, severely impacting the accuracy of displacement calculations. Furthermore, existing calibration methods largely depend on two reference points—the top and bottom of the borehole—and cannot dynamically identify and correct local errors at different depths within the borehole, leading to significant accuracy degradation over long-term monitoring. Therefore, effectively improving the positioning accuracy and data reliability of the inclinometer across the entire borehole depth range while maintaining its automation advantages has become a bottleneck in current technological development. To address these issues, a novel data processing method is urgently needed. This method involves introducing a multi-node magnetic ring reference and encoder collaborative calibration mechanism to construct a segmented depth correction model, and combining this with bidirectional observation data fusion technology to achieve high-precision, high-stability automated deep displacement monitoring. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a data processing and analysis method for inclinometers based on magnetic ring-encoder collaborative calibration, which significantly improves the accuracy of single measurements and enhances the long-term stability, self-diagnostic capabilities, and intelligent early warning level of the inclinometer system.

[0005] This invention provides a method for processing and analyzing inclinometer data based on magnetic ring-encoder collaborative calibration, comprising the following steps: Multiple magnetic rings are arranged at a preset interval inside the inclinometer tube of the monitoring hole, and the center distance between adjacent magnetic rings is accurately measured as a reference depth sequence. At the same time, the encoder ranging model of the winch device is initialized. The winch motor is started to lower the probe. When the magnetic sensing device of the probe triggers the magnetic ring, the encoder pulse value is recorded and the measured depth is calculated. The measured depth is compared with the reference depth sequence. Based on the deviation, a weighted update strategy is used to dynamically correct the ranging coefficient of the current depth segment, generate the depth partition ranging factor, and intelligently diagnose the health status of the magnetic ring based on the magnetic field signal characteristics at the time of triggering and historical deviation data. Based on the depth partitioning ranging factor, a machine learning error prediction model is introduced. The ambient temperature, the cumulative running time of the wire rope, and the historical ranging coefficient deviation sequence are used as input features to predict the depth measurement error of the encoder ranging system in the future monitoring cycle. Feedforward compensation is performed on the depth partitioning ranging factor to generate an optimized anti-drift ranging factor. After calibration and compensation are completed, the control probe is raised and lowered according to the optimized step distance to collect forward and reverse tilt angle data respectively. The forward and reverse tilt angle data at the same depth are paired and fused. Using the anti-drift ranging factor and the fused tilt angle data, the horizontal displacement of each depth point is calculated using a piecewise accumulation algorithm to generate monitoring results.

[0006] In this scheme, multiple magnetic rings are arranged at preset intervals inside the inclinometer tube of the monitoring hole, and the center distance between adjacent magnetic rings is accurately measured as a reference depth sequence. Simultaneously, the encoder ranging model of the winch device is initialized, including: After installing a clinometer tube with guide grooves in the monitoring hole, multiple positioning magnetic rings are arranged at a preset interval of 5 to 10 meters along the tube depth direction. The magnetic rings are smart magnetic rings with unique identification codes. After construction is completed, the center distance between adjacent magnetic rings is obtained as the reference value for segmentation. The center distance of all adjacent magnetic rings is associated with the corresponding smart magnetic ring identification code to generate a reference depth sequence with identification. Securely install the winch-type automatic inclinometer at the pipe opening, ensuring that the encoder and the winch steering wheel are arranged coaxially, initialize the encoder ranging model, establish an initial pulse number-depth conversion model, and introduce a model pre-optimization mechanism based on historical data. Simultaneously, the winch is driven to idle for short-stroke self-calibration operation. By analyzing the stability and repeatability of the encoder pulses, the confidence of the initial model is evaluated and fine-tuned to generate a dynamically optimized initial ranging model.

[0007] In this scheme, the measured depth is compared with the reference depth sequence, and a weighted update strategy is used to dynamically correct the ranging coefficient of the current depth segment based on the deviation, generating a depth partition ranging factor, including: The calculated measured depth is compared with the reference depth of the corresponding magnetic ring in the reference depth sequence to obtain the depth deviation. The time-domain and frequency-domain characteristics of the magnetic ring trigger signal are obtained. The time-domain characteristics include the signal peak amplitude, the signal half-peak width, and the symmetry of the rising edge slope and falling edge slope of the signal waveform. The frequency-domain characteristics include the energy of the main frequency component and the harmonic distortion. Using the time-domain and frequency-domain features, an initial quality score of the magnetic field signal is obtained based on a pre-trained signal health inference model. The initial quality score is then fused with the current monitoring context information to obtain the magnetic field signal quality score. The theoretical deformation effects of depth deviation, magnetic field signal quality score, and ambient temperature on the wire rope are used as inputs to the fuzzy logic system for fuzzy processing. The fuzzy input and historical calibration data sequence are then imported into a lightweight DNN module. The lightweight DNN module is used to mine the nonlinear relationship between multi-sensor data and ranging error, and output dynamic weight coefficients. Based on the dynamic weighting coefficient The ranging coefficient for the current depth segment is corrected using the following formula: ,in This represents the distance measurement coefficient after correction for the nth depth segment. This represents the ranging coefficient before correction for the nth depth segment. This represents the depth deviation measured at the nth magnetic ring. This represents the reference length value for the nth depth segment. This represents the adaptive learning rate factor. Represents the environmental compensation function, using real-time temperature. and relative humidity As input, the influence of environmental factors on the elastic modulus of the wire rope and the frictional characteristics of the winch system is compensated; The correction process is carried out segment by segment as the probe is lowered. When the probe is lowered to the bottom of the hole and all magnetic rings are triggered and calibrated, a depth zone ranging factor lookup table covering the entire hole depth is generated.

[0008] In this solution, based on the characteristics of the magnetic field signal at the time of triggering and historical deviation data, the health status of the magnetic ring is intelligently diagnosed, including: When the probe triggers the magnetic ring, a multi-dimensional magnetic ring feature signal sequence, including magnetic field signal waveform characteristics, spatial consistency characteristics, and depth deviation timing characteristics, is simultaneously acquired. The multi-dimensional magnetic ring feature signal sequence is then augmented to generate a version with specific differences. The data-enhanced signal is used to construct positive sample pairs with temporal perturbations. The positive sample pairs are then applied to stacked isomorphic variational autoencoders with shared input parameters. The encoder maps the input to the latent space to obtain the latent variable distribution parameters. The corresponding latent vectors are obtained by reparameter sampling. The latent vectors are then reconstructed by the decoder. Contrastive learning is introduced into the stacked isomorphic variational autoencoder, and the obtained latent vectors are used as positive sample pairs to maximize the similarity of the positive sample pairs in the latent space, while making the positive samples distinguish the latent vectors of other magnetic ring samples in the same batch. End-to-end training of stacked isomorphic variational autoencoders is performed using the inherent loss of variational autoencoders and contrastive learning loss to obtain low-dimensional feature representations of the health status of magnetic rings. Using labeled magnetic ring sample data, a preliminary signal is generated for a classifier based on latent spatial features. Uncertainty sampling is used to select magnetic ring samples whose low-dimensional feature representation distribution is closest to a uniform distribution with respect to health status. Representative sampling is used to select magnetic ring samples with the greatest difference from labeled magnetic ring samples. The training set is expanded using dynamically selected samples, and the health status diagnosis model is dynamically optimized. During the online diagnosis phase, a pre-trained stacked isomorphic variational autoencoder is used to obtain a low-dimensional feature representation of each magnetic ring. This representation is then substituted into the health diagnosis model to obtain a probability distribution based on Softmax. The category with the highest probability is used as the preliminary diagnosis result, and the corresponding probability value is used as the preliminary health score to obtain the corresponding health level. The uncertainty of the prediction is obtained by the entropy value of the probability distribution, and the representativeness is obtained by calculating the average cosine similarity between the feature vector of the current magnetic ring and the feature vector of the labeled magnetic ring sample. Based on the uncertainty and representativeness, confidence and risk labels are obtained. The health level, confidence level, and risk marker are used as outputs to obtain the diagnostic results of the magnetic ring's health status.

[0009] In this scheme, a machine learning error prediction model is introduced based on the depth partitioning ranging factor to predict the depth measurement error of the encoder ranging system in future monitoring cycles. Feedforward compensation is then applied to the depth partitioning ranging factor to generate an optimized anti-drift ranging factor, including: The system acquires the ambient temperature, cumulative running time of the wire rope, and historical ranging coefficient deviation sequence within the historical monitoring period, constructs a multi-dimensional time-series feature dataset related to the error, builds an error prediction model based on spatiotemporal graph neural network, uses different depth partitions as nodes, constructs an edge structure based on the physical distance and mechanical linkage relationship between nodes, and constructs a graph structure of ranging coefficient deviation based on the nodes and edge structure. The ambient temperature, the cumulative running time of the wire rope, and the graph structure are input together into the gated loop unit to obtain the graph structure change pattern of the ranging coefficient deviation throughout the entire depth range caused by changes in working conditions and load. By jointly training the spatiotemporal network, the nonlinear laws of the error in both depth and time dimensions are obtained. The ambient temperature, cumulative running time of the wire rope, and historical ranging coefficient deviation sequence within the current monitoring period are imported into the trained error prediction model. Obtain the predicted deviation of the ranging coefficients in each depth zone within a future monitoring period, and synthesize the depth zone ranging factor with the predicted deviation to generate an optimized anti-drift ranging factor for this monitoring.

[0010] In this scheme, after calibration and compensation are completed, the control probe is raised and lowered according to the optimized step size, and forward and reverse tilt angle data are collected respectively, including: Before the probe is lifted to collect forward data, the anti-drift ranging factor is loaded into the motion control unit. Based on the anti-drift ranging factor corresponding to the current depth, the pulse threshold of the drive motor is dynamically calculated and adjusted, and equal-step lifting is performed based on the pulse threshold. When the probe reaches the target depth, the built-in miniature inertial measurement unit monitors the probe's triaxial angular velocity and acceleration in real time. At the same time, the winch servo motor applies adaptive constant tension to suppress the swing of the wire rope. When the angular velocity and acceleration are both less than the preset threshold, the tilt sensor is triggered to collect data. The probe is raised to the tube opening and enters the rotation mechanism. A high-resolution optical encoder is used to provide real-time feedback on the rotation angle for position control. After rotation, the dual-axis tilt sensor inside the probe measures the static attitude angle after rotation and compares it with the attitude angle before rotation. The comparison result is used to perform a second verification of the rotation result. After completing the rotation calibration, the probe is lowered with a compensated step size, and reverse data acquisition is performed using the same steady-state sensing and triggering mechanism as the forward data acquisition.

[0011] In this scheme, positive and negative dip angle data at the same depth are paired and fused. Using the anti-drift ranging factor and the fused dip angle data, a piecewise accumulation algorithm is used to calculate the horizontal displacement of each depth point, generating monitoring results, including: Dynamic time warping is used to nonlinearly align the forward and reverse depth-dip angle curves to obtain data point pairs at the same depth position. Consistency verification and adaptive fusion are performed on the paired dip angle data pairs to obtain the true dip angle. The true physical step distance of each depth segment after calibration with the anti-drift ranging factor is obtained. The segmented horizontal displacement increment is calculated using the true tilt angle and the true physical step distance. The horizontal displacement increment of each depth segment is accumulated segment by segment to obtain the cumulative horizontal displacement of any depth point relative to the reference point, and the monitoring result is generated.

[0012] In this scheme, trend prediction is made based on the monitoring results of horizontal displacement, including: A depth-horizontal displacement spatiotemporal matrix is ​​constructed from monitoring data of several consecutive periods. Empirical mode decomposition is used to decompose the spatiotemporal matrix to obtain deformation modes with different spatiotemporal characteristics. Modal feature vectors of different deformation modes are obtained. Multi-scale weighted similarity measurement is used. The similarity between the modal feature vectors and the feature vectors of each typical destruction mode in the feature library is calculated to identify the top N candidate modes with the highest similarity and their scores. Different levels of deformation warnings are generated based on modal similarity and modal evolution trends, and the deformation warnings are output and sent.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates the multi-node physical reference provided by the magnetic ring with the depth of encoder ranging to construct a ranging factor that dynamically corrects with depth. This suppresses the cumulative depth error caused by factors such as wire rope tension and slippage, achieving high-precision depth positioning across the entire borehole range. Through bidirectional observation and data fusion, systematic errors introduced by probe zero drift and attitude offset are effectively eliminated, making the tilt angle measurement results more accurate and reliable. The machine learning error prediction model in this solution can proactively compensate for the time drift of the ranging system based on environmental conditions and system operating history, ensuring high accuracy during long-term continuous operation and effectively combating mechanical wear and material fatigue. Simultaneously, the magnetic ring health self-diagnosis function continuously assesses the reliability of the reference system, promptly detecting and warning of magnetic ring performance degradation or displacement risks, ensuring the long-term stability of the measurement reference. Furthermore, this solution reduces monitoring costs and personnel safety risks, making it particularly suitable for harsh environments and remote areas. The intelligent early warning mechanism based on deformation modal similarity matching breaks through the traditional lagging early warning mode based solely on displacement thresholds. It can identify the development trend of specific failure modes earlier based on the morphological evolution characteristics of the displacement field, providing more forward-looking early warning information for engineering safety risks and assisting managers in making accurate decisions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0015] Figure 1 A flowchart of a method for processing and analyzing inclinometer data based on magnetic ring-encoder co-calibration is shown; Figure 2 A flowchart illustrating the health status of an intelligent diagnostic magnetic ring is shown. Figure 3 The flowchart for generating the optimized anti-drift ranging factor is shown; Figure 4 A flowchart is shown to generate monitoring results by calculating the horizontal displacement at each depth point. Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0018] like Figure 1 As shown, this embodiment provides a method for inclinometer data processing and analysis based on magnetic ring-encoder collaborative calibration, including: Multiple magnetic rings are arranged at a preset interval inside the inclinometer tube of the monitoring hole, and the center distance between adjacent magnetic rings is accurately measured as a reference depth sequence. At the same time, the encoder ranging model of the winch device is initialized. The winch motor is started to lower the probe. When the magnetic sensing device of the probe triggers the magnetic ring, the encoder pulse value is recorded and the measured depth is calculated. The measured depth is compared with the reference depth sequence. Based on the deviation, a weighted update strategy is used to dynamically correct the ranging coefficient of the current depth segment, generate the depth partition ranging factor, and intelligently diagnose the health status of the magnetic ring based on the magnetic field signal characteristics at the time of triggering and historical deviation data. Based on the depth partitioning ranging factor, a machine learning error prediction model is introduced. The ambient temperature, the cumulative running time of the wire rope, and the historical ranging coefficient deviation sequence are used as input features to predict the depth measurement error of the encoder ranging system in the future monitoring cycle. Feedforward compensation is performed on the depth partitioning ranging factor to generate an optimized anti-drift ranging factor. After calibration and compensation are completed, the control probe is raised and lowered according to the optimized step distance to collect forward and reverse tilt angle data respectively. The forward and reverse tilt angle data at the same depth are paired and fused. Using the anti-drift ranging factor and the fused tilt angle data, the horizontal displacement of each depth point is calculated using a piecewise accumulation algorithm to generate monitoring results.

[0019] It should be noted that after installing the inclinometer tube with guide grooves inside the monitoring hole, multiple positioning magnetic rings are arranged at preset intervals of 5-10 meters along the tube depth direction. These magnetic rings are intelligent magnetic rings with unique identification codes; the identification fingerprint signal can be identified and decoded by the magnetic sensing device inside the probe. After construction, a high-precision laser rangefinder is used to measure the center distance between adjacent magnetic rings bidirectionally, and the average value is taken as the reference value for segmentation. The center distances of all adjacent magnetic rings are associated with the corresponding intelligent magnetic ring identification codes to generate a reference depth sequence with identification. The winch-type automatic inclinometer is then securely installed at the tube opening. Ensure the encoder and winch steering wheel are coaxially arranged, initialize the encoder ranging model, input static parameters such as steering wheel radius and wire rope diameter to establish an initial pulse number-depth conversion model, introduce a model pre-optimization mechanism based on historical data, call the initial ranging parameter combination verified under similar working conditions (such as equal hole depth and similar temperature and humidity), instead of relying solely on theoretical calculation values; at the same time, drive the winch to idle for short-stroke self-calibration operation, and evaluate and fine-tune the confidence of the initial model by analyzing the stability and repeatability of encoder pulses, and generate a dynamically optimized initial ranging model.

[0020] It should be noted that when the winch motor is started, the probe is lowered at a constant low speed. When the Hall sensor array built into the probe detects the magnetic ring and triggers, multi-dimensional trigger data is collected, including: the precise pulse count value of the encoder at the moment of triggering, the peak value of the magnetic field strength sensed by the sensor, and the gradient characteristics of the magnetic field strength change. The calculated measured depth is compared with the reference depth of the corresponding magnetic ring in the reference depth sequence to obtain the depth deviation. The time-domain and frequency-domain characteristics of the magnetic ring trigger signal are obtained. The time-domain characteristics include the peak amplitude of the signal, the half-peak width of the signal, and the symmetry of the rising and falling slopes of the signal waveform. The frequency-domain characteristics include the energy of the main frequency component and the harmonic distortion. Using the aforementioned time-domain and frequency-domain features, an initial quality score for the magnetic field signal is obtained based on a pre-trained signal health inference model (such as a support vector machine or gradient boosting decision tree). This initial quality score is then fused with the current monitoring context information to obtain the final magnetic field signal quality score. Correction factors include at least historical consistency and motion state correlation. Historical consistency compares the characteristics of the current trigger signal with the average level of trigger signal characteristics of the same magnetic ring in previous monitoring cycles; if the deviation is within a reasonable range, the confidence level is increased. Motion state correlation checks whether the winch system is in a uniform and stable operating state at the signal trigger time. The depth deviation, magnetic field signal quality score, and the theoretical deformation influence of ambient temperature on the wire rope are used as inputs to a fuzzy logic system for fuzzy processing. The fuzzified input and historical calibration data sequence are then imported into a lightweight DNN module. This lightweight DNN module is used to mine the nonlinear relationship between multi-sensor data and ranging error, outputting dynamic weight coefficients. When the DNN module determines that the data reliability is high... The value approaches 1, allowing for the absorption of deviation values; when interference is detected, Approaching 0 makes the correction behavior more conservative.

[0021] Based on the dynamic weighting coefficient The ranging coefficient for the current depth segment is corrected using the following formula: ,in This represents the distance measurement coefficient after correction for the nth depth segment. This represents the ranging coefficient before correction for the nth depth segment. This represents the depth deviation measured at the nth magnetic ring. This represents the reference length value for the nth depth segment. This represents the adaptive learning rate factor. Represents the environmental compensation function, using real-time temperature. and relative humidity As input, the influence of environmental factors on the elastic modulus of the wire rope and the friction characteristics of the winch system is compensated. The correction process is carried out segment by segment as the probe is lowered. When the probe is lowered to the bottom of the hole and all magnetic rings are triggered and calibrated, a depth partitioning distance factor lookup table covering the entire hole depth is generated to improve the depth measurement accuracy and suppress cumulative errors.

[0022] It should be noted that real-time monitoring and intelligent diagnosis of the magnetic ring reference component's own condition ensures the long-term reliability of the reference system. For example... Figure 2As shown, when the probe triggers the magnetic ring, a multi-dimensional magnetic ring feature signal sequence (depth deviation value generated by the magnetic ring in multiple consecutive monitoring cycles) is synchronously acquired, including magnetic field signal waveform characteristics (peak intensity, half-wave width, waveform symmetry, and rising slope of the magnetic field strength-time curve), spatial consistency characteristics (spatial symmetry of magnetic field distribution), and depth deviation timing characteristics. The multi-dimensional magnetic ring feature signal sequence is then augmented to generate a version with defined differences. The data augmentation perturbation strategy includes random translation within a preset range on the sequence time axis, adding random noise with a signal-to-noise ratio higher than a specific threshold, or randomly scaling the signal amplitude within a specific proportional range.

[0023] The augmented signal is used to construct positive sample pairs with temporal perturbations. These pairs represent the same health status information of magnetic ring instances, including temporal and amplitude differences introduced by the inherent characteristics of the measurement system and environmental background. The positive sample pairs are then fed into a stacked isomorphic variational autoencoder with shared input parameters. The encoder maps the input to the latent space, obtaining the latent variable distribution parameters (mean and variance). Reparameter sampling is used to obtain the corresponding latent vectors, which are then reconstructed by the decoder. Contrastive learning is introduced into the stacked isomorphic variational autoencoder, using the obtained latent vectors as positive sample pairs to maximize their similarity in the latent space and distinguish them from the latent vectors of other magnetic ring samples within the same batch. End-to-end training of the stacked isomorphic variational autoencoder is performed using the inherent loss of the variational autoencoder and the contrastive learning loss to obtain a low-dimensional feature representation of the magnetic ring health status. The inherent loss of the variational autoencoder is used to further refine this representation. Including reconstruction loss and KL divergence loss, it ensures that the latent space can effectively represent the input data and conform to the prior distribution, expressed as: , In the formula Let be the mean square error function. For decoder, , To input the original magnetic field signal After data augmentation, the two derived versions are obtained as latent space feature vectors through the encoder and reparameterization, respectively. These are the weighting coefficients. Let KL divergence function be used. The posterior distribution of the encoder output. Let be the prior distribution of the latent variables; Among them, contrastive learning loss Represented as: , In the formula The cosine similarity function is used. For temperature hyperparameters, For batch size, For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. In the batch, except and its positive samples The latent vectors of other samples are negative samples.

[0024] In the latent space of the pre-trained stacked isomorphic variational autoencoder, feature vectors of the same magnetic ring triggered multiple times in a healthy state are aggregated, while feature vectors of different fault modes are distributed in different and distinguishable regions of the latent space.

[0025] Using labeled magnetic ring sample data (a small number of known healthy and known failed data samples), a preliminary signal is generated for a classifier based on latent spatial features. In subsequent long-term monitoring, uncertainty sampling is used to select magnetic ring samples whose low-dimensional feature representation distribution is closest to a uniform distribution of health status. Representative sampling is used to select magnetic ring samples with the greatest difference from the labeled magnetic ring samples. The training set is expanded using dynamically selected samples, and the health diagnosis model is dynamically optimized. This ensures that the diagnosis model can achieve maximum performance improvement and knowledge coverage with minimal labeling cost, enabling the model to continuously adapt to new magnetic ring failure modes and avoid model obsolescence or misjudgment due to insufficient initial training data.

[0026] During the online diagnostic phase, a pre-trained stacked isomorphic variational autoencoder is used to obtain a low-dimensional feature representation for each magnetic ring. This representation is then substituted into a health diagnosis model to obtain a Softmax-based probability distribution. The category with the highest probability is used as the preliminary diagnostic result, and the corresponding probability value is used as the preliminary health score to obtain the corresponding health level. The uncertainty of the prediction is obtained based on the entropy value of the probability distribution; the more uniform the predicted probability distribution, the higher the entropy value and the higher the uncertainty. The average cosine similarity between the feature vector of the current magnetic ring and the feature vector of the labeled magnetic ring samples is calculated to obtain representativeness. Higher representativeness indicates a novel sample, which in turn indicates an unknown fault mode. Based on the uncertainty and representativeness, confidence and risk markers are obtained. The health level, confidence marker, and risk marker are used as outputs to obtain the diagnostic result of the magnetic ring's health status. When the diagnostic result is high-confidence health, an early warning is triggered; when it is low-confidence or an unknown fault, a higher-level alarm is triggered, prompting management personnel to intervene and verify the situation.

[0027] It should be noted that a machine learning error prediction model is introduced based on the depth partitioning ranging factor to predict the depth measurement error of the encoder ranging system in future monitoring cycles. Feedforward compensation is then applied to the depth partitioning ranging factor to generate an optimized anti-drift ranging factor, thereby actively suppressing systematic errors and significantly improving the stability of long-term monitoring. For example... Figure 3As shown, the average ambient temperature, cumulative running time of the wire rope, and historical ranging coefficient deviation sequences of each depth zone after collaborative calibration in the historical monitoring period are obtained, and a multi-dimensional time series feature dataset related to the error is constructed.

[0028] An error prediction model based on a spatiotemporal graph neural network is constructed. Different depth zones are used as nodes, and edge structures are built according to the physical distance and mechanical linkage between nodes (the tension of the upstream wire rope will cumulatively affect the downstream ranging). Based on the nodes and edge structures, a graph structure of ranging coefficient deviation is constructed. The spatial correlation and propagation law of errors between different depth zones are learned through graph convolutional networks. Ambient temperature, cumulative running time of the wire rope, and the graph structure are input into a gated loop unit to obtain the graph structure change pattern of ranging coefficient deviation throughout the entire depth range caused by changes in working conditions and load. The nonlinear law of error in both depth and time dimensions is obtained through joint training of spatiotemporal networks. The ambient temperature, cumulative running time of the wire rope, and historical ranging coefficient deviation sequences in the current monitoring cycle are imported into the trained error prediction model. The predicted deviation of the ranging coefficient in each depth zone in the next monitoring cycle is obtained. Depth partitioning ranging factor Deviation from the prediction Synthesize and generate the optimized anti-drift ranging factor for this monitoring. The synthesis formula is: If the model predicts a positive bias (the measured value is too large), the ranging factor is pre-adjusted to compensate. At the end of each monitoring cycle, the actual bias obtained from the collaborative calibration is compared with the model's predicted bias. When the prediction error continues to exceed the threshold, the model is incrementally learned using the latest data to enable the model to track performance degradation or changes in operating conditions, ensuring the accuracy of long-term predictions.

[0029] It should be noted that after calibration and compensation, the probe is raised and lowered according to the optimized step size, collecting forward and reverse tilt angle data respectively to ensure the probe's attitude stability at each depth point and the spatiotemporal consistency of forward and reverse data acquisition. Before the probe is raised for forward data acquisition, the anti-drift ranging factor is loaded into the motion control unit. Based on the anti-drift ranging factor corresponding to the current depth, the pulse threshold of the drive motor is dynamically calculated and adjusted. Based on the pulse threshold, equal-step raising is performed to ensure that the physical position of each observation point does not drift due to errors. When the probe reaches the target depth, it enters the intelligent dwell acquisition stage. The built-in micro inertial measurement unit monitors the probe's three-axis angular velocity and acceleration in real time. At the same time, an adaptive constant tension is applied using a winch servo motor to suppress the swing of the wire rope. When both the angular velocity and acceleration are less than the preset threshold, it is determined that the probe has reached a mechanically stable state, and the tilt sensor is triggered to acquire data, avoiding measurement noise introduced by probe swing. The probe is raised to the tube opening and enters the rotation mechanism. A high-resolution optical encoder provides real-time feedback of the rotation angle for position control, precisely controlling the rotation angle. After rotation, the dual-axis tilt sensor inside the probe measures the static attitude angle after rotation and compares it with the attitude angle before rotation. The comparison result is used to perform a second verification of the rotation result. After the rotation verification is completed, the probe is lowered with a compensated step size and reverse data acquisition is performed using the same steady-state sensing and triggering mechanism as the forward data acquisition.

[0030] It should be noted that, as Figure 4 As shown, tilt and depth data are converted into horizontal displacement to generate predictive monitoring results. Dynamic time warping is used to nonlinearly align the forward and reverse depth-tilt curves, obtaining data point pairs at the same depth position to achieve optimal data pairing on a substep scale. Consistency verification and adaptive fusion are performed on the paired tilt data pairs to obtain the true tilt angle. In the consistency verification, the values ​​of the forward and reverse tilt angle data are calculated; if they exceed a preset threshold, the data point is marked as low confidence. The fused true tilt angle is calculated using (forward tilt angle data - reverse tilt angle data) / 2, eliminating zero-drift errors that are in phase with the probe attitude.

[0031] By combining the depth accuracy of the anti-drift ranging factor with the directional accuracy of the true tilt angle data, the true physical step distance of each depth segment after calibration with the anti-drift ranging factor is obtained. The true physical step distance = nominal step distance of the depth segment × anti-drift ranging factor. The anti-drift ranging factor corrects the rope length counted by the encoder to the actual depth the probe moves within the hole. Using the true tilt angle and the true physical step distance, the segmented horizontal displacement increment is calculated. The calibrated hypotenuse length is projected onto the horizontal direction, and the horizontal displacement increments of each depth segment are accumulated segment by segment to obtain the cumulative horizontal displacement of any depth point relative to the reference point, generating the monitoring result.

[0032] It should be noted that a depth-horizontal displacement spatiotemporal matrix is ​​constructed from monitoring data of several consecutive periods, where row vectors represent depth coordinates and column vectors represent event vectors. The matrix describes the displacement changes with time and space. Two-dimensional empirical mode decomposition is used to decompose the spatiotemporal matrix to obtain eigenmode fractions and residual terms characterizing deformation modes with different spatiotemporal features. The different eigenmode fractions include high-frequency, small-amplitude periodic fluctuations caused by temperature changes, mid-frequency deformation components related to groundwater levels, local abrupt signals of potential slip surface formation, and long-term monotonic trends caused by soil creep. A convolutional autoencoder is used to compress and learn the two-dimensional morphology of the modalities. Modal feature vectors of different deformation modes are obtained through feature encoding. A multi-scale weighted similarity metric is used to calculate the similarity between the modal feature vectors and the feature vectors of each typical failure mode in the feature library. The multi-scale weighted similarity metric includes global similarity and local feature similarity. The global similarity is calculated by the cosine similarity of the two modal feature vectors, and the local feature similarity is calculated by the morphological similarity in key depth intervals such as potential slip surfaces and key time intervals such as deformation acceleration segments. Different weights are assigned to global and local similarities and they are weighted and fused to generate a comprehensive similarity score. The top N candidate modes with the highest similarity and their scores are identified. Deformation warnings of different levels are issued based on modal similarity and the evolution trend of the modes, and the deformation warnings are output and sent out. The selected deformation early warning levels include modal manifestation early warning, modal acceleration early warning, and critical state early warning. A modal manifestation early warning is triggered if the comprehensive similarity score of the current modal feature vector is consistently higher than a preset threshold for the first time. A modal acceleration early warning is triggered if the comprehensive similarity score of the modal continues to rise and the energy contribution rate represented by the modal components significantly increases. A critical state early warning is triggered if both the modal similarity and energy contribution rate reach higher thresholds, and the displacement time history curve of the modality shows acceleration characteristics. By integrating similarity and modal energy evolution into a multi-level early warning mechanism, the foresight and accuracy of geological disaster early warning are greatly improved.

[0033] A second embodiment of the present invention provides a computer-readable storage medium, which includes a program for a method of inclinometer data processing and analysis based on magnetic ring-encoder co-calibration. When the program is executed by a processor, it implements the steps of a method of inclinometer data processing and analysis based on magnetic ring-encoder co-calibration.

[0034] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0035] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing and analyzing inclinometer data based on magnetic ring-encoder collaborative calibration, characterized in that, Includes the following steps: Multiple magnetic rings are arranged at a preset interval inside the inclinometer tube of the monitoring hole, and the center distance between adjacent magnetic rings is accurately measured as a reference depth sequence. At the same time, the encoder ranging model of the winch device is initialized. The winch motor is started to lower the probe. When the magnetic sensing device of the probe triggers the magnetic ring, the encoder pulse value is recorded and the measured depth is calculated. The measured depth is compared with the reference depth sequence. Based on the deviation, a weighted update strategy is used to dynamically correct the ranging coefficient of the current depth segment, generate the depth partition ranging factor, and intelligently diagnose the health status of the magnetic ring based on the magnetic field signal characteristics at the time of triggering and historical deviation data. Based on the depth partitioning ranging factor, a machine learning error prediction model is introduced. The ambient temperature, the cumulative running time of the wire rope, and the historical ranging coefficient deviation sequence are used as input features to predict the depth measurement error of the encoder ranging system in the future monitoring cycle. Feedforward compensation is performed on the depth partitioning ranging factor to generate an optimized anti-drift ranging factor. After calibration and compensation are completed, the control probe is raised and lowered according to the optimized step distance to collect forward and reverse tilt angle data respectively. The forward and reverse tilt angle data at the same depth are paired and fused. Using the anti-drift ranging factor and the fused tilt angle data, the horizontal displacement of each depth point is calculated using a piecewise accumulation algorithm to generate monitoring results.

2. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 1, characterized in that, Multiple magnetic rings are arranged at preset intervals inside the inclinometer tube of the monitoring hole, and the center distance between adjacent magnetic rings is accurately measured as a reference depth sequence. Simultaneously, the encoder ranging model of the winch device is initialized, including: After installing a clinometer tube with guide grooves in the monitoring hole, multiple positioning magnetic rings are arranged at a preset interval of 5 to 10 meters along the tube depth direction. The magnetic rings are smart magnetic rings with unique identification codes. After construction is completed, the center distance between adjacent magnetic rings is obtained as the reference value for segmentation. The center distance of all adjacent magnetic rings is associated with the corresponding smart magnetic ring identification code to generate a reference depth sequence with identification. Securely install the winch-type automatic inclinometer at the pipe opening, ensuring that the encoder and the winch steering wheel are arranged coaxially, initialize the encoder ranging model, establish an initial pulse number-depth conversion model, and introduce a model pre-optimization mechanism based on historical data. Simultaneously, the winch is driven to idle for short-stroke self-calibration operation. By analyzing the stability and repeatability of the encoder pulses, the confidence of the initial model is evaluated and fine-tuned to generate a dynamically optimized initial ranging model.

3. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 1, characterized in that, The measured depth is compared with the reference depth sequence, and the ranging coefficients of the current depth segment are dynamically corrected based on the deviation using a weighted update strategy to generate depth partition ranging factors, including: The calculated measured depth is compared with the reference depth of the corresponding magnetic ring in the reference depth sequence to obtain the depth deviation. The time-domain and frequency-domain characteristics of the magnetic ring trigger signal are obtained. The time-domain characteristics include the signal peak amplitude, the signal half-peak width, and the symmetry of the rising edge slope and falling edge slope of the signal waveform. The frequency-domain characteristics include the energy of the main frequency component and the harmonic distortion. Using the time-domain and frequency-domain features, an initial quality score of the magnetic field signal is obtained based on a pre-trained signal health inference model. The initial quality score is then fused with the current monitoring context information to obtain the magnetic field signal quality score. The theoretical deformation effects of depth deviation, magnetic field signal quality score, and ambient temperature on the wire rope are used as inputs to the fuzzy logic system for fuzzy processing. The fuzzy input and historical calibration data sequence are then imported into a lightweight DNN module. The lightweight DNN module is used to mine the nonlinear relationship between multi-sensor data and ranging error, and output dynamic weight coefficients. Based on the dynamic weighting coefficient The ranging coefficient for the current depth segment is corrected using the following formula: ,in This represents the distance measurement coefficient after correction for the nth depth segment. This represents the ranging coefficient before correction for the nth depth segment. This represents the depth deviation measured at the nth magnetic ring. This represents the reference length value for the nth depth segment. This represents the adaptive learning rate factor. Represents the environmental compensation function, using real-time temperature. and relative humidity As input, the influence of environmental factors on the elastic modulus of the wire rope and the frictional characteristics of the winch system is compensated; The correction process is carried out segment by segment as the probe is lowered. When the probe is lowered to the bottom of the hole and all magnetic rings are triggered and calibrated, a depth zone ranging factor lookup table covering the entire hole depth is generated.

4. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 1, characterized in that, Based on the characteristics of the magnetic field signal at the time of triggering and historical deviation data, the health status of the magnetic ring is intelligently diagnosed, including: When the probe triggers the magnetic ring, a multi-dimensional magnetic ring feature signal sequence, including magnetic field signal waveform characteristics, spatial consistency characteristics, and depth deviation timing characteristics, is simultaneously acquired. The multi-dimensional magnetic ring feature signal sequence is then augmented to generate a version with specific differences. The data-enhanced signal is used to construct positive sample pairs with temporal perturbations. The positive sample pairs are then applied to stacked isomorphic variational autoencoders with shared input parameters. The encoder maps the input to the latent space to obtain the latent variable distribution parameters. The corresponding latent vectors are obtained by reparameter sampling. The latent vectors are then reconstructed by the decoder. Contrastive learning is introduced into the stacked isomorphic variational autoencoder, and the obtained latent vectors are used as positive sample pairs to maximize the similarity of the positive sample pairs in the latent space, while making the positive samples distinguish the latent vectors of other magnetic ring samples in the same batch. End-to-end training of stacked isomorphic variational autoencoders is performed using the inherent loss of variational autoencoders and contrastive learning loss to obtain low-dimensional feature representations of the health status of magnetic rings. Using labeled magnetic ring sample data, a preliminary signal is generated for a classifier based on latent spatial features. Uncertainty sampling is used to select magnetic ring samples whose low-dimensional feature representation distribution is closest to a uniform distribution with respect to health status. Representative sampling is used to select magnetic ring samples with the greatest difference from labeled magnetic ring samples. The training set is expanded using dynamically selected samples, and the health status diagnosis model is dynamically optimized. During the online diagnosis phase, a pre-trained stacked isomorphic variational autoencoder is used to obtain a low-dimensional feature representation of each magnetic ring. This representation is then substituted into the health diagnosis model to obtain a probability distribution based on Softmax. The category with the highest probability is used as the preliminary diagnosis result, and the corresponding probability value is used as the preliminary health score to obtain the corresponding health level. The uncertainty of the prediction is obtained by the entropy value of the probability distribution, and the representativeness is obtained by calculating the average cosine similarity between the feature vector of the current magnetic ring and the feature vector of the labeled magnetic ring sample. Based on the uncertainty and representativeness, confidence and risk labels are obtained. The health level, confidence level, and risk marker are used as outputs to obtain the diagnostic results of the magnetic ring's health status.

5. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 1, characterized in that, Based on the depth partitioning ranging factor, a machine learning error prediction model is introduced to predict the depth measurement error of the encoder ranging system in future monitoring cycles. Feedforward compensation is applied to the depth partitioning ranging factor to generate an optimized anti-drift ranging factor, including: The system acquires the ambient temperature, cumulative running time of the wire rope, and historical ranging coefficient deviation sequence within the historical monitoring period, constructs a multi-dimensional time-series feature dataset related to the error, builds an error prediction model based on spatiotemporal graph neural network, uses different depth partitions as nodes, constructs an edge structure based on the physical distance and mechanical linkage relationship between nodes, and constructs a graph structure of ranging coefficient deviation based on the nodes and edge structure. The ambient temperature, the cumulative running time of the wire rope, and the graph structure are input together into the gated loop unit to obtain the graph structure change pattern of the ranging coefficient deviation throughout the entire depth range caused by changes in working conditions and load. By jointly training the spatiotemporal network, the nonlinear laws of the error in both depth and time dimensions are obtained. The ambient temperature, cumulative running time of the wire rope, and historical ranging coefficient deviation sequence within the current monitoring period are imported into the trained error prediction model. Obtain the predicted deviation of the ranging coefficients in each depth zone within a future monitoring period, and synthesize the depth zone ranging factor with the predicted deviation to generate an optimized anti-drift ranging factor for this monitoring.

6. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 1, characterized in that, After calibration and compensation are completed, the control probe is raised and lowered according to the optimized step size, and forward and reverse tilt angle data are collected respectively, including: Before the probe is lifted to collect forward data, the anti-drift ranging factor is loaded into the motion control unit. Based on the anti-drift ranging factor corresponding to the current depth, the pulse threshold of the drive motor is dynamically calculated and adjusted, and equal-step lifting is performed based on the pulse threshold. When the probe reaches the target depth, the built-in miniature inertial measurement unit monitors the probe's triaxial angular velocity and acceleration in real time. At the same time, the winch servo motor applies adaptive constant tension to suppress the swing of the wire rope. When the angular velocity and acceleration are both less than the preset threshold, the tilt sensor is triggered to collect data. The probe is raised to the tube opening and enters the rotation mechanism. A high-resolution optical encoder is used to provide real-time feedback on the rotation angle for position control. After rotation, the dual-axis tilt sensor inside the probe measures the static attitude angle after rotation and compares it with the attitude angle before rotation. The comparison result is used to perform a second verification of the rotation result. After completing the rotation calibration, the probe is lowered with a compensated step size, and reverse data acquisition is performed using the same steady-state sensing and triggering mechanism as the forward data acquisition.

7. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 1, characterized in that, Pairing and fusing forward and reverse dip angle data at the same depth, and using the anti-drift ranging factor and the fused dip angle data, a piecewise accumulation algorithm is employed to calculate the horizontal displacement at each depth point, generating monitoring results, including: Dynamic time warping is used to nonlinearly align the forward and reverse depth-dip angle curves to obtain data point pairs at the same depth position. Consistency verification and adaptive fusion are performed on the paired dip angle data pairs to obtain the true dip angle. The true physical step distance of each depth segment after calibration with the anti-drift ranging factor is obtained. The segmented horizontal displacement increment is calculated using the true tilt angle and the true physical step distance. The horizontal displacement increment of each depth segment is accumulated segment by segment to obtain the cumulative horizontal displacement of any depth point relative to the reference point, and the monitoring result is generated.

8. The inclinometer data processing and analysis method based on magnetic ring-encoder collaborative calibration according to claim 7, characterized in that, Trend prediction based on horizontal displacement monitoring results includes: A depth-horizontal displacement spatiotemporal matrix is ​​constructed from monitoring data of several consecutive periods. Empirical mode decomposition is used to decompose the spatiotemporal matrix to obtain deformation modes with different spatiotemporal characteristics. Modal feature vectors of different deformation modes are obtained. Multi-scale weighted similarity measurement is used. The similarity between the modal feature vectors and the feature vectors of each typical destruction mode in the feature library is calculated to identify the top N candidate modes with the highest similarity and their scores. Different levels of deformation warnings are generated based on modal similarity and modal evolution trends, and the deformation warnings are output and sent.

Images