Multi-source data fusion anchor rod length detection and imaging system and method

Abstract

The invention relates to the technical field of imaging processing, and provides a multi-source data fusion anchor rod length detection and imaging system and method, and the method comprises the steps: carrying out the space-time alignment of multi-source data through a dynamic path compensation algorithm, and generating first target data; wherein the first target data comprises ultrasonic data and electromagnetic data; performing feature extraction on the first target data through a cross-modal contrast learning model, and determining associated features; performing scene attribute optimization on the associated features through a meta-learning optimization framework, and determining imaging weights of different associated features; and according to the imaging weight, fusing and aligning the multi-stage associated features, and outputting an anchor rod length detection result and defect imaging.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of imaging processing, in particular to a multi-source data fusion anchor rod length detection and imaging method. BACKGROUND

[0002] The anchor rod is a core load-bearing component in geotechnical engineering such as tunnel support and slope reinforcement, and its quality determines the safety of the engineering project. The anchor rod is usually completely wrapped by concrete or geotechnical medium, and the traditional detection method cannot observe the internal state, so non-destructive detection technology is usually used to detect the length and defects of the anchor rod.

[0003] In the traditional technology, single induction detection is used for anchor rod detection, mainly ultrasonic detection or radar detection, but it is easy to miss detection.

[0004] Ultrasonic detection calculates the length of the anchor rod based on the time difference of the echo signal, and is sensitive to surface defects such as surface cracks and exposed section corrosion. However, the wave speed of ultrasonic waves is affected by temperature (frozen soil environment) and medium uniformity, and cannot penetrate deep medium, so internal cracks, grouting defects and other hidden defects of the anchor rod cannot be identified during detection. The ground penetrating radar detects by emitting high-frequency electromagnetic waves and receiving reflected signals, and can penetrate deep medium and is sensitive to internal defects, but the signal will be affected by electromagnetic interference (such as spatial deviation caused by shaking of the detection equipment), and the surface detail resolution is low.

[0005] In the fusion of the two kinds of induction, the ultrasonic and electromagnetic data are fused by time stamp alignment or feature splicing, and there is a problem of time and space asynchronization of multiple sensors:

[0006] There are differences in sampling frequency and signal propagation characteristics between ultrasonic and electromagnetic sensors, and the time jitter reaches the microsecond level, so the fused data do not have time and space correspondence;

[0007] The time sequence features of ultrasonic waves (such as echo peak sequence) and the spatial features of electromagnetic waves (such as radar wave reflection layer distribution);

[0008] Both of them do not belong to the same feature space, and there is no unified correlation mechanism, so the semantic consistency of the fused features is low, and the real state of the anchor rod cannot be highlighted. SUMMARY

[0009] The present application provides a multi-source data fusion anchor rod length detection and imaging system and method, which is used to solve the problems in traditional anchor rod detection, such as insufficient single modal information leading to missed detection of hidden defects, asynchronization of multiple sensors leading to fusion failure, poor adaptability to complex environment, real-time imaging limited by computing power, and low reliability of detection results, and through multi-source data fusion and intelligent algorithm optimization, high-precision, high-adaptability and high-real-time anchor rod length and defect detection are realized.

[0010] In a first aspect, a multi-source data fusion anchor rod length detection and imaging method is provided, comprising:

[0011] The multi-source data is aligned in space and time through a dynamic path compensation algorithm to generate first target data; wherein the first target data includes ultrasonic data and electromagnetic data;

[0012] The first target data is extracted through a cross-modal contrast learning model to determine the associated features;

[0013] The associated features are executed through a meta-learning optimization framework to optimize the scene attributes, and the imaging weights of different associated features are determined;

[0014] According to the imaging weight, the multi-stage associated features are fused and aligned to output the anchor rod length detection result and the defect imaging.

[0015] In combination with the first aspect, the multi-source data includes ultrasonic data, electromagnetic data and environmental data collected by a sensor array; wherein the sensor array is configured with a quantum time synchronization module, and the quantum time synchronization module includes a White Rabbit protocol based on FPGA.

[0016] In combination with the first aspect, the dynamic path compensation algorithm is as follows:

[0017]

[0018] Wherein, Δt is the time compensation, L actual is the actual length, L nominal is the nominal length, u s (T) is the wave speed related to temperature T, ε r is the relative dielectric constant, and c is a constant.

[0019] In combination with the first aspect, the cross-modal contrast learning model is composed of ResNet1D network and Transformer network mapped to a unified feature space through MLP network; wherein,

[0020] The ResNet1D network is configured in the ultrasonic encoder of the sensor array, and is used to extract the time sequence features of the ultrasonic data;

[0021] The Transformer network is configured in the electromagnetic encoder of the sensor array, and is used to extract the spatial features of the electromagnetic data;

[0022] The MLP network is configured in the projection head connected to the sensor array.

[0023] In combination with the first aspect, the associated features are executed through a meta-learning optimization framework to optimize the scene attributes, comprising:

[0024] Obtain the geological conditions of the current detection scene, and determine whether it belongs to a known scene in the historical training set; wherein, if it is a new scene, trigger the meta-learning optimization process: if it is a known scene, determine the imaging weight corresponding to the historical optimization parameter and the associated feature.

[0025] In combination with the first aspect, the multi-stage associated feature fusion alignment comprises:

[0026] Preprocess the associated features, and calculate the similarity between the preprocessed associated features and the reference features;

[0027] If the similarity is less than a preset first similarity threshold, adjust the preprocessing parameters based on the environmental data, and reprocess the associated features until the similarity is greater than or equal to the preset first similarity threshold;

[0028] Align the preprocessed associated features again, and calculate the spatio-temporal consistency of the aligned features;

[0029] If the consistency is not satisfied, call the dynamic path compensation algorithm to re-align;

[0030] Fuse the multi-stage features through the cross-modal attention mechanism, calculate the information entropy of the fused features, and if the information entropy is greater than an information entropy threshold, adjust the attention weight and re-fuse.

[0031] In combination with the first aspect, the multi-stage associated feature fusion alignment according to the imaging weight, and outputting the anchor rod length detection result and the defect imaging further comprises:

[0032] Real-time perception of the environmental state through the temperature and humidity sensor and the inertial navigation sensor;

[0033] If a low-temperature environment is detected, it is identified that the ultrasonic data may have high-frequency attenuation;

[0034] If device shaking is detected, it is identified that the electromagnetic data may have spatial errors;

[0035] In a low-temperature environment, reduce the weight of ultrasonic features and enhance the weight of electromagnetic penetration information; when the device is shaking, add spatial smoothing filtering to the electromagnetic data and add time window averaging to the ultrasonic data;

[0036] Verify the spatial continuity of the fused results after adjustment, and if it is not improved, roll back the original weight and mark that the environmental disturbance needs to be confirmed manually.

[0037] In combination with the first aspect, the multi-stage associated feature fusion alignment according to the imaging weight, and outputting the anchor rod length detection result and the defect imaging further comprises:

[0038] Evaluate the confidence of the associated features through the MC-Dropout model, and if the difference between multiple sampling results is significant, it is determined as low confidence;

[0039] triggering a reprocessing procedure when there is a low-confidence sampling result;

[0040] checking the spatio-temporal alignment quality of multi-source data, and re-aligning if the deviation exceeds the standard;

[0041] adjusting the cross-modal contrast learning model parameters and re-extracting the associated features if the confidence is still not improved after alignment;

[0042] outputting the result if the confidence meets the standard after the second reprocessing, otherwise marking the area as high-risk for review.

[0043] In combination with the first aspect, the associated features of multiple stages are fused and aligned according to the imaging weight, and the anchor rod length detection result and defect imaging are outputted, further comprising:

[0044] Based on the associated features, the defect type is identified, if the ultrasonic wave echo is suddenly changed, it is determined as a surface defect, if the electromagnetic radar wave penetration is abnormal, it is determined as a hidden defect, and if both exist, it is determined as a composite defect;

[0045] According to the defect type, the imaging strategy is adjusted, including: for surface defects, superimposing the ultrasonic edge detection result, for hidden defects, superimposing the electromagnetic penetration profile, and for composite defects, displaying in layers and distinguishing the severity through color gradient;

[0046] According to the imaging strategy, the consistency of upper and lower layer information is verified after imaging, and if there is a contradiction, the meta-learning optimization framework is called to recalibrate the imaging weight.

[0047] The second aspect is a multi-source data fusion anchor rod length detection and imaging system, comprising:

[0048] The spatio-temporal alignment module is used to perform spatio-temporal alignment on multi-source data through a dynamic path compensation algorithm to generate first target data; wherein the first target data includes ultrasonic data and electromagnetic data;

[0049] The feature extraction module is used to extract features from the first target data through a cross-modal contrast learning model to determine the associated features;

[0050] The weight analysis module is used to execute scene attribute optimization on the associated features through a meta-learning optimization framework, and determine the imaging weight of different associated features;

[0051] The imaging processing module is used to fuse and align the associated features of multiple stages according to the imaging weight, and output the anchor rod length detection result and defect imaging.

[0052] Other features and advantages of the present application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

[0053] The technical solutions of the present application are described in further detail below by means of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0054] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of this specification that illustrates the application, and together with the description serve to explain the principles of the application. The drawings are provided solely for purposes of illustrating the preferred embodiments of the present application, and are not to be construed as limiting the present application.

[0055] In the drawings:

[0056] Figure 1 A method flow chart of a multi-source data fusion anchor rod length detection and imaging method in an embodiment of the present application;

[0057] Figure 2 A system composition diagram of a multi-source data fusion anchor rod length detection and imaging system in an embodiment of the present application;

[0058] Figure 3 A composition diagram of a traditional anchor rod length detection and imaging system in an embodiment of the present application;

[0059] Figure 4 An implementation scenario diagram of a cross-modal contrast learning model in an embodiment of the present application;

[0060] Figure 5 An implementation flow chart of associated feature fusion alignment in an embodiment of the present application;

[0061] Figure 6 A process diagram of dynamic optimization of fusion strategies in a complex environment in an embodiment of the present application;

[0062] Figure 7 A process diagram of precise visualization of defect information in an embodiment of the present application. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0064] Anchor rod is the most basic component of roadway support in modern coal mine, which reinforces the surrounding rock of roadway together, so that the surrounding rock supports itself. Anchor rod is not only used in mines, but also used in engineering technology to reinforce the main body of slope, tunnel and dam. As a tension member penetrating into the stratum, one end of the anchor rod is connected with the engineering structure, and the other end penetrates into the stratum. The whole anchor rod is divided into a free section and an anchoring section. The free section is the area that transmits the tension at the anchor rod head to the anchoring body. Its function is to apply pre-stress to the anchor rod. The present application mainly detects the anchoring section.

[0065] The detection of anchor rod is mainly through data acquisition and data feature recognition to perform imaging of the anchor rod detection area. As shown in Figure 3 The conventional detection mainly includes defect detection of ultrasonic detection equipment and defect detection of radar detection equipment. In the processing equipment 10, if single sensing detection is used, the echo signal of ultrasonic detection or the electromagnetic signal of geological radar detection is calculated by the calculation equipment 20 to recognize the defect features and then output the anchor rod imaging diagram. However, the detail resolution is low due to the influence of temperature, medium and the like.

[0066] With the development of technology, through the software of the calculation equipment 20, the echo signal of ultrasonic detection and the electromagnetic signal of geological radar detection can be fused by time stamp / feature splicing. The fusion of time stamp is to determine the same feature detected by the echo signal of ultrasonic detection and the electromagnetic signal of geological radar detection in the same time stamp to realize anchor rod imaging. Feature splicing is to compare the same features according to the echo signal of ultrasonic detection and the electromagnetic signal of geological radar detection. If the echo signal of ultrasonic detection has feature defects in the real imaging of anchor rod, the existing feature defects can be compensated by the electromagnetic signal of geological radar detection. Or the features detected by the echo signal of ultrasonic detection are fused with the features detected by the electromagnetic signal of geological radar detection, and the two features are fused to find more details. However, only the time stamp alignment, the spatial position of the features, the propagation characteristics and the semantic consistency of the features in different detection methods are different, and the second imaging obtained is also of low quality. The processing equipment 10 and the calculation equipment 20 are part of the anchor rod imaging equipment.

[0067] In order to solve the above defects, the present application proposes a multi-source data fusion anchor rod length detection and imaging system. Referring to Figure 2, specifically including: a sensor array 50, a database 40, a user terminal 30, and a processing device 10 and a computing device 20. At the data acquisition level, the sensor array 50 collects environmental data through the environmental sensing unit 503, which can reduce the interference caused by environmental factors in the imaging results. In data fusion, the spatio-temporal alignment module 100 of the present application fuses a dynamic path compensation algorithm, which maps ultrasonic data, electromagnetic data and environmental data to the same spatio-temporal coordinate system through timestamp calibration and path transformation of spatial coordinates, preventing fusion errors caused by asynchronous imaging results in time and space, and also ensuring that the cross-modal contrast learning model performs feature extraction and ensures the consistency of imaging. Because ultrasonic data (time domain signal) and electromagnetic data (frequency domain signal) belong to different modalities, the feature distribution difference is large. Through the contrast learning model, positive sample pairs (ultrasonic-electromagnetic data at the same anchor position) and negative sample pairs (ultrasonic-electromagnetic data at different positions) are constructed to maximize the feature similarity of positive samples and minimize the similarity of negative samples, thereby extracting cross-modal correlation features, determining the influence of the environment on imaging, and thus reducing the influence of the environment on imaging. The meta-learning optimization framework 3001 dynamically adjusts the imaging weight of different correlation features by using the feature weight distribution of historical scenes and fine-tuning the weight with a small amount of new scene data, preventing the electromagnetic data weight from being too high when the anchor rod depth is large due to severe ultrasonic attenuation. The imaging processing module 400 integrates multi-dimensional information in multi-stage correlation feature fusion alignment, outputs the anchor rod length (feature peak position calculation) and defect imaging (feature space distribution mapping), and improves detection accuracy and imaging clarity by integrating multi-stage information. For example, a single cross-modal feature may miss scene adaptation information, and after fusion, both "feature correlation" and "scene adaptation" can be reflected, accurately identifying the anchor rod end position and internal defects (such as cracks and cavities).

[0068] Embodiment 1:

[0069] Referring to Figure 1 The present application provides a multi-source data fusion anchor rod length detection and imaging method, comprising:

[0070] Step 1001: performing spatio-temporal alignment on the multi-source data through a dynamic path compensation algorithm to generate first target data; wherein the first target data includes ultrasonic data and electromagnetic data;

[0071] The spatio-temporal alignment of multi-source data is based on a dynamic path compensation algorithm, which uses the White Rabbit protocol to achieve nanosecond-level synchronization of sensors, and combines temperature-dependent wave speed, medium dielectric constant and other parameters to compensate for time errors caused by length deviation, wave speed variation and medium delay, eliminate the spatio-temporal misalignment of multi-source data (ultrasonic and electromagnetic data), and determine the fusion reference of spatio-temporal synchronization.

[0072] Step 1002: feature extraction of the first target data through the cross-modal contrast learning model to determine the associated features;

[0073] The feature extraction of the cross-modal contrast learning model extracts the time sequence features (such as echo peak sequence) of the ultrasonic data through the ResNet1D network, extracts the spatial features (such as radar wave reflection layer distribution) of the electromagnetic data through the Transformer network, and then maps to a unified feature space through the MLP projection head. Through the contrast learning of maximizing the cross-modal similarity, the complementary information of the two heterogeneous data is associated, and the feature display of the hidden defects is improved.

[0074] Step 1003: scene attribute optimization of the associated features through the meta-learning optimization framework 3001, and determination of the imaging weights of different associated features;

[0075] The meta-learning optimization framework 3001 quickly adjusts the model parameters through multi-task meta-learning, and combines the online learning mechanism to fine-tune the model to quickly adapt to new geological conditions such as permafrost and aquifer using unlabelled data, without the need for a large amount of labelled data weight for low-quality training.

[0076] Step 1004: fusion and alignment of the multi-stage associated features according to the imaging weights, and output of the anchor rod length detection result and the defect imaging.

[0077] The fusion imaging of the multi-stage associated features is based on the scene optimization to determine the imaging weights, dynamically adjusts the fusion strategy (such as enhancing the weight of electromagnetic penetration information in low-temperature scenes) combined with the environmental data, quantifies the uncertainty through MC-Dropout, and generates high-precision and high-confidence anchor rod length and defect imaging results.

[0078] Embodiment 2:

[0079] Referring to Figure 2 The multi-source data is collected by a sensor array 50, which includes an ultrasonic detection module 501, an electromagnetic induction module 502, and an environmental sensing unit 503.

[0080] The ultrasonic detection module 501 collects ultrasonic data; the electromagnetic induction module 502 collects electromagnetic data; and the environmental sensing unit 503 collects environmental data; wherein the sensor array 50 is configured with a quantum time synchronization module 504, and the quantum time synchronization module 504 is configured with a White Rabbit protocol based on FPGA.

[0081] The ultrasonic data acquires echo signals of the anchor rod structure through high-frequency sampling, and displays the surface and near-field characteristics of the anchor rod;

[0082] The electromagnetic data acquires penetration signals of the internal medium through high-frequency radar waves, and the deep structure characteristics of the anchor rod;

[0083] The environmental data reflect external interference factors through temperature, humidity, device posture and other environmental parameters when acquiring and detecting temperature and humidity and inertial navigation sensors.

[0084] The three kinds of data are multi-dimensionally fused, and single modal information does not cause defect omission. The quantum time synchronization module 504 is based on the White Rabbit protocol of FPGA, and through an accurate timestamp synchronization mechanism, unifies the sampling time sequence of each sensor at a nanosecond level of precision, and ensures that the ultrasonic wave, electromagnetic and environmental data are strictly aligned on the time axis. Although the multi-sensor sampling is not synchronized, data misplacement does not occur.

[0085] The complementarity (surface details and internal penetration) of ultrasonic wave and electromagnetic data is completely retained through synchronous acquisition, and the environmental data can compensate for environmental interference (such as the influence of temperature on wave speed and the influence of device shaking on electromagnetic signals), and time and space misplacement does not cause feature correlation failure.

[0086] Embodiment 3:

[0087] The dynamic path compensation algorithm of the application is as follows:

[0088]

[0089] Wherein, Δt is the time compensation amount, L actual is the actual length, L nominal is the nominal length, u s (T) is the wave speed related to temperature T, ε r is the relative dielectric constant, and c is a constant.

[0090] The dynamic path compensation algorithm of the application fuses the time compensation amount, the dielectric constant, and the temperature and constant to determine the time compensation amount on the anchor rod, ultrasonic wave and electromagnetic wave, realizes time and space alignment, and performs length alignment in space.

[0091] Embodiment 4:

[0092] The cross-modal contrast learning model of the application is composed of a ResNet1D network 2001 and a Transformer network 2002 mapped to a unified feature space through an MLP network 4001; wherein,

[0093] The ResNet1D network 2001 is configured in the ultrasonic wave encoder of the sensor array 50, and is used for extracting the time sequence features of the ultrasonic wave data;

[0094] The Transformer network 2002 is configured in the electromagnetic encoder of the sensor array 50, and is used for extracting the spatial features of the electromagnetic data;

[0095] The MLP network 4001 is configured in a projection head connected to the sensor array 50.

[0096] Referring to Figure 4 The cross-modal contrast learning model is composed of the ResNet1D network 2001, the Transformer network 2002, and the MLP network 4001.

[0097] The ResNet1D network 2001 is configured in an ultrasonic encoder, and is used for processing time sequence characteristics of ultrasonic data by using a one-dimensional residual network, solving a gradient disappearance problem of a deep network through a residual connection, and capturing time sequence features (such as time distribution of echo peak values and change law of attenuation rate) of ultrasonic echo signals to determine time sequence detail information of an anchor surface structure and a near-field defect.

[0098] The Transformer network 2002 is configured in an electromagnetic encoder, and is used for processing spatial characteristics of electromagnetic data by using a self-attention mechanism, capturing a reflection signal dependency relationship (such as spatial distribution of a medium interface and location characteristics of a hidden defect) of radar waves at different depths and positions to determine spatial penetration information of an internal medium distribution and a deep defect of the anchor.

[0099] The MLP network 4001 is configured in a projection head, and is used for adjusting a feature dimension by using a multi-layer fully connected network, mapping time sequence features extracted by the ResNet1D and spatial features extracted by the Transformer to a feature space of the same dimension, eliminating feature distribution differences between different modalities, and making heterogeneous features comparable in semantics.

[0100] If only the ResNet1D or the Transformer single network is used, surface and internal multi-dimensional features cannot be covered at the same time; if the MLP mapping is lacked, the two kinds of features cannot be effectively associated due to dimension mismatch and large distribution difference, and the cross-modal similarity cannot be maximized by the contrast learning. After the combination of multiple data, the correlation accuracy of cross-modal features is improved, and features of hidden defects (such as internal cracks) are more clearly displayed.

[0101] Embodiment 5:

[0102] The application executes scene attribute optimization on the associated features by using the meta-learning optimization framework 3001, including:

[0103] Geological conditions of a current detection scene are obtained, and it is determined whether the current detection scene belongs to a known scene in a historical training set; if the current detection scene is a new scene, a meta-learning optimization process is triggered; if the current detection scene is a known scene, imaging weights corresponding to historical optimization parameters and associated features are determined.

[0104] Referring to Figure 2, the meta-learning optimization framework 3001 has the dynamic logic of scene judgment and parameter adjustment, and the self-adaptive ability to different geological conditions. Through learning the general rules of parameter adjustment from historical training tasks (such as known ordinary surrounding rock, aquifer, and other geological conditions), when facing a new scene (such as a permafrost environment that has never been encountered before), the scene type is judged based on the current associated features (multi-dimensional features extracted by cross-modal contrastive learning), and the meta-learning optimization process is triggered to quickly adjust the model parameters based on the experience of historical tasks to adapt to the new scene with a small amount of new data; if it is a known scene (such as a trained ordinary surrounding rock), the historical optimized parameters are directly called without repeated training.

[0105] The meta-learning optimization framework 3001 does not need to retrain a large amount of labeled data; the historical parameters in known scenes are reused to improve detection efficiency and avoid redundant calculations. The associated features (time sequence details of ultrasonic waves and spatial penetration information of electromagnetic waves) provide multi-dimensional basis for scene judgment, and the meta-learning optimization framework 3001 can accurately identify the differences between new scenes and known scenes; meta-learning optimization adjusts parameters based on features to make the imaging weight more suitable for the actual needs of the current scene (such as increasing the weight of electromagnetic penetration information in a permafrost scene).

[0106] Embodiment 6:

[0107] Referring to Figure 5 In the process of performing multi-stage associated feature fusion alignment, the present application includes the following steps:

[0108] The associated features are preprocessed, and the similarity between the preprocessed associated features and the reference features is calculated; the preprocessing of the associated features includes denoising, normalization, etc. The similarity between the preprocessed features and the reference features is calculated to evaluate the matching degree of the features with the ideal state.

[0109] If the similarity is less than a preset first similarity threshold, the preprocessing parameters are adjusted based on the environmental data to reprocess the associated features until the similarity is greater than or equal to the preset first similarity threshold; if the similarity is insufficient (such as feature distortion caused by environmental interference), the preprocessing parameters (such as filtering strength, normalization threshold) are adjusted based on the environmental data (temperature and humidity, device posture, etc.) until the standard is met.

[0110] The preprocessed associated features are aligned again, and the spatio-temporal consistency of the aligned features is calculated;

[0111] If the consistency is not satisfied, a dynamic path compensation algorithm is called to realign; the secondary alignment is to verify the spatio-temporal consistency of the features again after preprocessing (such as timestamp deviation, spatial position offset), if the consistency is not satisfied (new deviation may be introduced due to preprocessing), a dynamic path compensation algorithm is called to realign to ensure the accuracy of the features in the spatio-temporal dimension.

[0112] The multi-stage features are fused through a cross-modal attention mechanism, and the information entropy of the fused features is calculated; if the information entropy is greater than the information entropy threshold, the attention weight is adjusted and then the fusion is performed again. The cross-modal attention mechanism allocates weights by learning the dependency between features (e.g., paying more attention to the electromagnetic features corresponding to hidden defects), and the information entropy measures the amount of information of the fused features. If the information entropy is too high (indicating that there is redundancy or noise), the attention weight is adjusted (the weight of the low-quality modality is reduced), and the fusion effect is improved.

[0113] The preprocessing and similarity adjustment are used to determine the basic quality of the input features, preventing noise or distorted features from affecting the fusion;

[0114] The secondary alignment and dynamic path compensation can evaluate the spatio-temporal accuracy and prevent fusion errors caused by misalignment;

[0115] The cross-modal attention and information entropy adjustment optimize the correlation of multi-modal features, making the fusion result meet the actual detection requirements.

[0116] Embodiment 7:

[0117] Referring to Figure 6 , the present application fuses and aligns the multi-stage correlation features according to the imaging weight, outputs the anchor length detection result and defect imaging, and further includes:

[0118] The temperature and humidity sensor and the inertial navigation sensor perceive the environmental state in real time; through environmental perception, problem identification, strategy adjustment and verification rollback, a closed-loop logic is formed to realize fusion in complex environments. The temperature and humidity sensor and the inertial navigation sensor collect environmental parameters (such as temperature and device acceleration) in real time. There is an inherent correlation between the quality of the collected environmental parameters (such as temperature and device acceleration) and the quality of the ultrasonic and electromagnetic data. For example, a low-temperature environment can cause high-frequency signal attenuation of ultrasonic waves in a hardened medium, and device shaking can cause spatial sampling position deviation of electromagnetic radar waves, thereby reducing data quality.

[0119] If a low-temperature environment is detected, it is identified that the ultrasonic data may have high-frequency attenuation; if device shaking is detected, it is identified that the electromagnetic data may have spatial errors; by identifying problems such as ultrasonic wave attenuation corresponding to low temperature and shaking corresponding to electromagnetic spatial error in different environmental states,

[0120] Adjust the fusion strategy based on the environmental state: reduce the weight of ultrasonic features and increase the weight of electromagnetic penetration information in a low-temperature environment; when the device is shaking, increase spatial smoothing filtering for electromagnetic data and increase time window averaging for ultrasonic data; reduce the weight of the modality that is easily affected (such as ultrasonic waves in a low-temperature environment) in fusion, and increase the weight of the penetration information of the modality that needs to be enhanced (such as electromagnetic waves in a low-temperature environment); increase spatial smoothing filtering for the disturbed data (such as electromagnetic data under shaking) to reduce noise, and increase time window averaging for the other modality (such as ultrasonic waves) to stabilize the time sequence features.

[0121] After adjustment, verify the spatial continuity of the fusion results. If no improvement is seen, roll back to the original weights and mark them as requiring manual confirmation of environmental interference. After adjustment, evaluate the effect of the adjustment by verifying the spatial continuity of the fusion results (e.g., the clarity of the anchor boundary and the integrity of the defect outline). If no improvement is seen, roll back to the original weights and mark them as requiring manual confirmation to avoid invalid adjustments.

[0122] Environmental awareness is a prerequisite; without environmental data, problems cannot be identified. Problem identification is key to strategy development; without a clear understanding of the problem type, effective adjustment strategies cannot be developed. Targeted solutions address issues that hinder convergence. Verification rollback prevents distorted results from incorrect adjustments. Improve the authenticity of anchor length detection and defect imaging in low-temperature frozen soil and equipment shaking.

[0123] Example 8:

[0124] This application fuses and aligns the associated features of multiple stages based on the imaging weights, outputs the anchor length detection results and defect imaging, and also includes:

[0125] The confidence of associated features is evaluated through the MC-Dropout model. If the differences between multiple sampling results are significant, it is judged as low confidence. By linking confidence evaluation with the reprocessing process, the reliability of the detection results is dynamically improved.

[0126] The MC-Dropout model uses the Dropout layer to sample the associated features multiple times during the inference phase. The variance of the multiple sampling results can quantify the uncertainty of the detection results. The larger the variance, the more unstable the judgment of the area and the lower the confidence level; vice versa. If low confidence is detected (significant differences in multiple sampling results), the reprocessing process is triggered:

[0127] First, check the spatiotemporal alignment quality of multi-source data. If the deviation exceeds the standard, realign it. If the confidence level does not improve after alignment, adjust the parameters of the cross-modal contrastive learning model and re-extract related features.

[0128] During reprocessing, check the spatiotemporal alignment quality of multi-source data (such as the timestamp deviation and spatial position matching of ultrasonic and electromagnetic data). If the alignment deviation exceeds the standard (possibly due to environmental interference or synchronization error), call the dynamic path compensation algorithm to realign the data to eliminate the impact of spatiotemporal misalignment on features. If the confidence does not increase after alignment (indicating that the problem may lie in the feature extraction link), adjust the parameters of the cross-modal contrastive learning model (such as increasing the attention weight of the electromagnetic encoder) and re-extract related features to enhance the feature discrimination ability.

[0129] If the confidence level meets the standard after two reprocessings, the result will be output; otherwise, the area will be marked as high risk for review.

[0130] If the confidence level meets the standard after two reprocessing (the sampling results tend to be consistent), output the results;

[0131] If it still does not meet the standard, mark the area as high-risk for review and prompt manual intervention.

[0132] The confidence evaluation of MC-Dropout is used to judge the quality of the output results, preventing the system from only outputting "pass / fail" ambiguity; the reprocessing process corrects the specific reasons for low confidence (data alignment or feature extraction problems).

[0133] Example 9:

[0134] Referring to Figure 7 , the present application aligns and outputs anchor length detection results and defect imaging based on imaging weights, and further includes:

[0135] Based on the associated features, the defect type is identified. If the ultrasonic wave echo is suddenly changed, it is determined to be a surface defect. If the electromagnetic radar wave penetration is abnormal, it is determined to be a hidden defect. If both exist, it is determined to be a composite defect. The associated features (time sequence details of ultrasonic waves and spatial penetration information of electromagnetics) imply the multi-dimensional attributes of the defect - ultrasonic wave echo mutation (such as amplitude drop, waveform distortion) usually corresponds to physical damage (such as fracture, rust) on the surface of the anchor, because ultrasonic waves are sensitive to surface structure; electromagnetic radar wave penetration anomaly (such as reflection layer interruption, energy attenuation anomaly) usually corresponds to discontinuity (such as cracks, grouting not dense) of internal medium, because electromagnetic waves can penetrate the medium and feedback internal information. Pattern matching of associated features can identify defect types (surface, hidden, composite)

[0136] Adjust the imaging strategy according to the defect type: superimpose the edge detection result of the ultrasonic wave for surface defects, superimpose the electromagnetic penetration profile for hidden defects, and display the composite defects in layers and distinguish the severity by color gradient; superimpose the edge detection result of the ultrasonic wave (such as extracting the surface contour by Canny algorithm) for surface defects, use the high resolution of ultrasonic waves to highlight the position and shape of surface damage; superimpose the penetration profile of the electromagnetic wave (such as the depth distribution of the radar wave reflection layer) for hidden defects, use the penetration of the electromagnetic wave to display the depth and trend of internal defects; composite defects are displayed in layers (the upper layer is surface details and the lower layer is internal information), and the severity of the defects is distinguished by color gradient (such as red for severe and yellow for moderate), and the damage range is directly reflected. After imaging, the consistency of the upper and lower layer information needs to be verified (such as no surface anomaly but internal display of fracture may mean that the surface cover layer hides the internal damage), if there is a contradiction, the meta-learning optimization framework 3001 is called to recalibrate the imaging weight (such as increasing the weight of the electromagnetic feature), to ensure the synergy of multi-modal information.

[0137] After imaging, the consistency of the upper and lower layer information is verified, and if there is a contradiction, the meta-learning optimization framework 3001 is called to recalibrate the imaging weight.

[0138] The imaging strategy adjustment matches the surface details of the ultrasonic wave, the internal penetration of the electromagnetic wave and the defect type of the multi-modal feature, and prevents imaging deviation.

[0139] Embodiment 10:

[0140] Referring to Figure 2 The anchor rod length detection and imaging system based on multi-source data fusion provided in the present application comprises:

[0141] The space-time alignment module 100 is configured to perform space-time alignment on the multi-source data through a dynamic path compensation algorithm to generate first target data; the first target data comprises ultrasonic wave data and electromagnetic data; the space-time alignment of the multi-source data is based on the dynamic path compensation algorithm, which uses the White Rabbit protocol to achieve nanosecond-level synchronization of sensors, and combines parameters such as temperature-dependent wave speed and medium dielectric constant to compensate for time errors caused by length deviation, wave speed variation and medium delay, eliminate space-time misalignment of multi-source data (ultrasonic wave and electromagnetic data), and determine a fusion reference for space-time synchronization.

[0142] The feature extraction module 200 is configured to perform feature extraction on the first target data through a cross-modal contrast learning model to determine associated features; the feature extraction of the cross-modal contrast learning model extracts time sequence features (such as echo peak value sequences) of the ultrasonic wave data through a ResNet1D network 2001, extracts spatial features (such as radar wave reflection layer distribution) of the electromagnetic data through a Transformer network 2002, and then maps to a unified feature space through a projection head of an MLP network 4001, and through contrast learning of maximizing cross-modal similarity, associates complementary information of the two heterogeneous data to improve feature display of hidden defects.

[0143] The weight analysis module 300 is configured to perform scene attribute optimization on the associated features through a meta-learning optimization framework 3001, and determine imaging weights of different associated features; the meta-learning optimization framework 3001 quickly adjusts model parameters through multi-task meta-learning, and combines an online learning mechanism to fine-tune the model using unlabeled data so that the model can quickly adapt to new geological conditions such as permafrost and aquifer, without the need for a large amount of labeled data weight for low-quality training.

[0144] The imaging processing module 400 is used for fusing and aligning the multi-stage associated features according to the imaging weight, outputting the anchor rod length detection result and the defect imaging. The fusion and alignment network 4002 in the imaging processing module 400 fuses and aligns the multi-stage associated features, determines the imaging weight based on scene optimization, dynamically adjusts the fusion strategy in combination with environmental data (for example, enhancing the weight of electromagnetic penetration information in a low-temperature scene), quantifies the uncertainty through MC-Dropout, and generates the anchor rod length and defect imaging result with high precision and high confidence.

[0145] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, it is intended that the present application cover the modifications and changes as they come within the scope of the claims and their equivalents.

Claims

1. A method for detecting and imaging anchor length using multi-source data fusion, characterized in that: include: Performing spatiotemporal alignment on the multi-source data using a dynamic path compensation algorithm to generate first target data; wherein the first target data includes ultrasonic data and electromagnetic data; Extract features from the first target data using a cross-modal contrastive learning model to determine relevant features; The associated features are used to perform scene attribute optimization through a meta-learning optimization framework, and the imaging weights of different associated features are determined; According to the imaging weight, the multi-stage associated features are fused and aligned to output the anchor length detection results and defect imaging.

2. The anchor length detection and imaging method using multi-source data fusion according to claim 1, characterized in that: The multi-source data includes ultrasonic data, electromagnetic data, and environmental data collected by a sensor array; wherein the sensor array is configured with a quantum time synchronization module, and the quantum time synchronization module includes a White Rabbit protocol based on FPGA.

3. The anchor length detection and imaging method based on multi-source data fusion according to claim 1, characterized in that: The dynamic path compensation algorithm is as follows: Among them, Δt is the time compensation, L actual is the actual length, L nominal is the nominal length, u s (T) is the wave velocity related to temperature T, ε r is the relative dielectric constant, and c is a constant.

4. The anchor length detection and imaging method using multi-source data fusion according to claim 2, characterized in that: The cross-modal contrastive learning model is composed of a ResNet1D network and a Transformer network mapped to a unified feature space through an MLP network; in, The ResNet1D network is configured in the ultrasonic encoder of the sensor array to extract the temporal features of the ultrasonic data; The Transformer network is configured in the electromagnetic encoder of the sensor array to extract the spatial features of the electromagnetic data; The MLP network is configured in a projection head connected to a sensor array.

5. The anchor length detection and imaging method using multi-source data fusion according to claim 2, characterized in that: The method of performing scene attribute optimization on the associated features through a meta-learning optimization framework includes: Obtain the geological conditions of the current detection scene and determine whether it belongs to a known scene in the historical training set. If it is a new scene, trigger the meta-learning optimization process. If it is a known scene, determine the imaging weights corresponding to the historical optimization parameters and associated features.

6. The anchor rod length detection and imaging method using multi-source data fusion according to claim 2, characterized in that: The multi-stage correlation feature fusion alignment includes: Preprocess the associated features and calculate the similarity between the preprocessed associated features and the benchmark features; If the similarity is less than a preset first similarity threshold, adjusting the preprocessing parameters based on the environmental data, and reprocessing the associated features until the similarity is greater than or equal to the preset first similarity threshold; Perform secondary alignment on the preprocessed correlation features and calculate the spatiotemporal consistency of the aligned features; If the consistency is not satisfied, the dynamic path compensation algorithm is called to realign; Multi-stage features are fused through the cross-modal attention mechanism, and the information entropy of the fused features is calculated; if the information entropy is greater than the information entropy threshold, the attention weight is adjusted and then re-fused.

7. The anchor rod length detection and imaging method using multi-source data fusion according to claim 2, characterized in that: The method further includes fusing and aligning the associated features of multiple stages according to the imaging weights and outputting the anchor length detection results and defect imaging. Real-time perception of environmental conditions through temperature and humidity sensors and inertial navigation sensors; If a low temperature environment is detected, the ultrasonic data may be attenuated at high frequencies; If device shaking is detected, the identification electromagnetic data may contain spatial errors; In low-temperature environments, the weight of ultrasonic features is reduced and the weight of electromagnetic penetration information is increased. When the device is shaking, spatial smoothing filtering is added to the electromagnetic data and time window averaging is added to the ultrasonic data. After adjustment, verify the spatial continuity of the fusion result. If it is not improved, roll back the original weights and mark the environmental interference that needs manual confirmation.

8. The anchor rod length detection and imaging method using multi-source data fusion according to claim 2, characterized in that: The method further includes fusing and aligning the associated features of multiple stages according to the imaging weights and outputting the anchor length detection results and defect imaging. The confidence of the associated features is evaluated using the MC-Dropout model. If the differences between multiple sampling results are significant, it is judged as low confidence. When there are low-confidence sampling results, the reprocessing process is triggered; Check the spatiotemporal alignment quality of multi-source data and realign if the deviation exceeds the standard; If the confidence score does not improve after alignment, adjust the cross-modal contrastive learning model parameters and re-extract related features; If the confidence level meets the standard after two reprocessings, the result will be output; otherwise, the area will be marked as high risk for review.

9. The anchor rod length detection and imaging method using multi-source data fusion according to claim 2, characterized in that: The method further includes fusing and aligning the associated features of multiple stages according to the imaging weights and outputting the anchor length detection results and defect imaging. Identify defect types based on correlation features. If the ultrasonic echo abruptly changes, it is considered a surface defect. If the electromagnetic radar wave penetrates abnormally, it is considered a hidden defect. If both exist, it is considered a compound defect. Adjust imaging strategies based on defect types. Imaging strategies include: superimposing ultrasonic edge detection results on surface defects, superimposing electromagnetic penetration profiles on hidden defects, and displaying complex defects in layers and distinguishing their severity through color gradients. According to the imaging strategy, the consistency of the upper and lower layer information is verified after imaging. If there is any contradiction, the meta-learning optimization framework is called to recalibrate the imaging weights.

10. A multi-source data fusion anchor length detection and imaging system, characterized in that: include: A spatiotemporal alignment module is used to perform spatiotemporal alignment on multi-source data using a dynamic path compensation algorithm to generate first target data; wherein the first target data includes ultrasonic data and electromagnetic data; Feature extraction module: used to extract features from the first target data through a cross-modal contrastive learning model to determine relevant features; Weight analysis module: used to optimize scene attributes using the meta-learning optimization framework for correlation features and determine the imaging weights of different correlation features; Imaging processing module: used to fuse and align multi-stage associated features according to imaging weights, and output anchor length detection results and defect imaging.

Images