Soil layer interface recognition model construction system based on intelligent image recognition

Abstract

The invention discloses a soil layer interface recognition model construction system based on intelligent image recognition, and relates to the field of soil layer recognition, and the system comprises a data collection module which obtains soil layer depth image data through a sensing collection device, synchronously collects probe detection resistance, side wall friction resistance and pore water pressure signals of corresponding depths, and sends the data to a data processing module; time and space registration is carried out on all the collected data according to depth identification; the feature matching module is used for extracting soil features from images and probe signal features from signal data and accurately judging whether sudden change points or intervals represent a real soil layer interface or not, so that the false detection rate and the omission ratio are greatly reduced, breakthrough of soil layer interface recognition accuracy is achieved, local interference is merged or maintained, and the recognition accuracy of the soil layer interface is improved. High robustness is maintained in a complex stratum environment, a foundation bearing capacity attenuation coefficient of a specified period in the future is output based on environment load history and future load prediction, and a high threshold feature and a soil layer identifier which have the maximum influence are identified.

Description

Technical Field

[0001] This invention relates to the field of soil layer identification technology, specifically a soil layer interface identification model construction system based on intelligent image recognition. Background Technology

[0002] With the acceleration of urbanization and the continuous increase in infrastructure construction, the demand for soil layer feature identification and assessment is increasing. Establishing an efficient and accurate soil interface identification system can provide important soil characteristic data support for foundation engineering, civil engineering and other fields. In particular, the rapid development of computer vision, machine learning and big data technologies, image recognition and signal processing technologies have made significant progress in various fields. The application of such technologies has made artificial intelligence-based soil layer identification systems have greater potential in terms of accuracy and efficiency.

[0003] Traditional methods often process image data or probe signals separately, ignoring the spatiotemporal correlation of multimodal data, leading to misjudgment of soil interfaces. They are difficult to adapt to regional differences in soil characteristics, are susceptible to noise interference, resulting in false positive interface identification. They rely on fixed depth intervals or single feature thresholds, cannot adaptively merge or separate soil layer changes, and lack quantitative prediction of dynamic soil layer decay. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a soil interface recognition model construction system based on intelligent image recognition, which can effectively solve the problems of the existing technology.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This invention discloses a soil interface recognition model construction system based on intelligent image recognition, comprising:

[0009] The data acquisition module acquires soil depth image data through sensing acquisition devices, and simultaneously acquires probe detection resistance, side wall friction and pore water pressure signals at the corresponding depths. All acquired data are registered in time and space according to depth identifiers.

[0010] The feature matching module extracts soil features from images and probe signal features from signal data. It then performs one-to-one matching between the extracted soil features and probe signal features at the same depth to construct a fusion dataset under depth labeling.

[0011] The initial layering module divides the fused dataset into several initial soil layers according to the depth range. Each initial soil layer contains the depth range, image features, and probe signal features.

[0012] The calibration unit is used to receive fused data and preliminary soil layer division results, detect abrupt trend segments exceeding preset thresholds by jointly analyzing image and signal feature sequences, pre-train the calibration model, input the abrupt features extracted within the preset depth range of the abrupt segments into the trained calibration model, and output the soil layer change results of the current preliminary soil layer division results.

[0013] The soil layer redefinition module is used to receive the soil layer modification results output by the calibration unit, and to perform fine-grained correction, merging or splitting of the initially divided soil layers to generate adjustment acquisition data corresponding to each final accurate soil layer.

[0014] The identification unit is used to automatically identify soil characteristics based on the adjusted acquisition data obtained from the soil redefinition module, and output the performance coefficient of the soil layer, as well as the decay coefficient of the performance coefficient in the future cycle.

[0015] The attenuation tracing module is used to analyze the correlation strength between the attenuation coefficient and the internal characteristics and performance coefficients of each soil layer, and to identify and output one or more soil layer characteristics and their soil layer identifiers that have the greatest impact on the predicted attenuation coefficient and have a high preset threshold influence.

[0016] Furthermore, the soil features of the feature matching module include: texture, particle size, color distribution, and structural morphology; the probe signal features include: time-domain statistics, frequency-domain components, and signal morphology.

[0017] Furthermore, the calibration unit is equipped with sub-modules, including a mutation trend detection module, a mutation feature extraction module, and a deep learning modeling module. The mutation feature extraction module is interconnected with the mutation trend detection module and the deep learning modeling module via a wireless network.

[0018] The mutation trend detection module receives the fused dataset and, along a continuous depth direction, jointly analyzes the sequence of image features and the sequence of probe signal features to identify mutation trend segments in the joint feature sequence that exceed a preset dynamic threshold. The process of jointly analyzing and identifying mutation trend segments involves: strictly aligning the image feature sequence with the probe signal feature sequence at the same depth according to depth identifiers to ensure synchronization of multimodal data at each depth point; normalizing the texture gradient and granularity distribution change rate in the image feature sequence; and quantizing the temporal statistical volatility and frequency domain energy offset in the probe signal sequence. The algorithm employs a weighted fusion method to merge the normalized image feature changes and signal feature fluctuations into a single-dimensional joint feature sequence. The image feature weights and signal feature weights are dynamically allocated based on the feature contribution of historical soil interface data. In the continuous depth direction, a dynamic sliding window algorithm is applied to the joint feature sequence to calculate the standard deviation of feature values ​​within the window and the absolute value of feature gradients between adjacent windows in real time. When the standard deviation or gradient value continuously exceeds the dynamic threshold set based on the historical statistical values ​​of soil continuity, the window is determined to be a sudden change trend segment, and its starting depth, ending depth, and sudden change intensity index are output.

[0019] The mutation feature extraction module is used to extract all image features and signal feature data within a preset depth range above and below each mutation point or mutation interval identified by the mutation trend detection module, and extract the mutation features within this depth range.

[0020] The deep learning modeling module is used to receive mutation features within the depth range output by the mutation feature extraction module, receive the preliminary soil layer division results output by the preliminary layering module, pre-train and calibrate the model, use the mutation features within the depth range as the core input of the calibration model, and integrate and consider the preliminary soil layer context information in which they are located to determine whether the mutation point or interval represents the real soil layer interface.

[0021] Furthermore, the mutation feature extraction module extracts mutation features including: maximum feature gradient, gradient change rate, feature mutation amplitude, and mutation morphology pattern.

[0022] Furthermore, during the judgment phase, the deep learning modeling module determines a valid interface when the gradient change rate and mutation amplitude in the mutation feature simultaneously exceed the dynamic threshold, and the mutation morphology matches the typical pattern of the soil interface; when the mutation feature only locally exceeds the threshold but the context information indicates that the feature is continuous and stable, it is determined to be an internal soil disturbance; based on the judgment result, a soil layer modification instruction is output: soil layer splitting or boundary correction is performed on the valid interface, and soil layer merging or maintaining the original division is performed on the internal disturbance.

[0023] Furthermore, the identification unit has sub-modules deployed at its lower level, including a model training module, a performance calculation module, and a decay analysis module. The performance calculation module is interconnected with the model training module and the performance calculation module via a wireless network, wherein:

[0024] The model training module is used to build a recognition model using a convolutional neural network algorithm. It trains the model based on historical soil layer adjustment data and known engineering performance indicators. The recognition model receives adjustment data for a specific soil layer, automatically identifies the characteristics of that soil layer, and outputs a performance coefficient characterizing its key engineering properties. The model construction process involves: structuring the historical soil layer adjustment data according to soil layer identifiers; each soil layer's sample data includes its corresponding texture, grain size, color distribution features, and probe signal feature vectors, while also associating it with known engineering performance indicator labels for that soil layer; and constructing a two-branch convolutional neural network, the first branch... The first branch is configured to process image feature vectors, extracting spatial hierarchical features through convolutional layers. The second branch is configured to process signal feature vectors, extracting temporal pattern features through one-dimensional convolutional layers. The high-level features output from the two branches are concatenated and then input into a fully connected layer. The dual-branch network is trained end-to-end using sample data, with engineering performance index labels as supervision signals. An adaptive gradient optimization algorithm is used to minimize the mean square error between the network output value and the true performance index. After training, the network receives adjustment data collected from new input soil layers. After dual-branch feature extraction and fusion, the fully connected layer outputs continuous values ​​representing the key engineering performance of the soil layer, i.e., performance coefficients.

[0025] The performance calculation module is used to calculate the overall foundation bearing capacity based on the performance coefficients of each soil layer output by the model training module, combined with the depth and spatial location information of each soil layer.

[0026] The attenuation analysis module uses the overall foundation bearing capacity output by the performance calculation module as a benchmark value, integrates and analyzes historical environmental loads, future load predictions, and time factors, and obtains the attenuation coefficient of the overall foundation bearing capacity within a specified future period through a time-series analysis algorithm.

[0027] Furthermore, the formula for calculating the overall foundation bearing capacity of the performance calculation module is as follows:

[0028] H total );

[0029] In the formula, Q u κ represents the overall bearing capacity of the foundation, n represents the soil synergy coefficient, and P represents the total number of soil layers. i h represents the performance coefficient of the i-th soil layer. i d represents the thickness of the i-th soil layer, γ represents the depth strengthening factor, and d represents the depth strengthening factor. iD represents the depth of the midpoint of the i-th soil layer. ref The reference depth is represented by a fixed value, λ represents the interlayer interaction attenuation coefficient, and Δz i P represents the gradient difference in performance coefficients between the i-th layer and its adjacent soil layers, η represents the sensitivity weight of the weak layer, and P represents the sensitivity weight of the weak layer. j Represents the j-th weakest soil layer. H represents the set of weakest soil layers. total Represents the total computation depth.

[0030] Furthermore, the attenuation analysis module performs time-series analysis as follows:

[0031] Based on historical and future environmental load predictions, the performance coefficients of each soil layer at the end of a specified period are predicted using a pre-trained soil performance degradation model.

[0032] Based on the updated performance coefficients of each soil layer, combined with soil layer depth and spatial location information, the future overall foundation bearing capacity is calculated using the same calculation strategy as the performance calculation module.

[0033] The ratio of the future overall foundation bearing capacity to the current overall foundation bearing capacity is used as the attenuation coefficient.

[0034] Furthermore, the data acquisition module is interconnected with a data preprocessing module via a wireless network. The data preprocessing module performs grayscale normalization, anisotropic diffusion filtering, and illumination compensation on the soil depth image to eliminate acquisition noise and non-uniform illumination interference. It also performs three-level processing on the probe detection resistance, sidewall friction, and pore water pressure signals, including: suppressing impulse noise through sliding window midpoint filtering; eliminating high-frequency interference by wavelet threshold denoising; and normalizing the amplitude of multi-source signals using depth marking.

[0035] Furthermore, the feature matching module is interconnected with the data acquisition module and the preliminary stratification module via a wireless network, the calibration unit is interconnected with the preliminary stratification module and the soil layer redefinition module via a wireless network, and the identification unit is interconnected with the soil layer redefinition module and the attenuation tracing module via a wireless network.

[0036] (III) Beneficial Effects

[0037] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0038] 1. By synchronously acquiring soil depth images and probe resistance, sidewall friction, and pore water pressure signals, and accurately registering them in time and space, the soil features such as texture, grain size, color distribution, and structural morphology of the images are matched one-to-one with probe features such as time-domain statistics, frequency-domain components, and signal morphology to construct a fusion dataset under depth identification. This not only enables joint analysis of feature sequences in continuous depth directions, dynamic threshold mutation detection, and calibration model pre-training, but also allows for context capture and feature extraction within the depth range for mutation points or intervals through multi-level sub-modules. This accurately determines whether a mutation point represents a real soil interface, significantly reducing false detection and false negative rates, and achieving a breakthrough in the accuracy of soil interface identification.

[0039] 2. By dividing the soil layer division and correction process into three major stages—initial stratification, abrupt change calibration, and fine-grained adjustment—abrupt change segment identification, feature extraction, and calibration judgment are completed respectively. Using multi-dimensional information such as gradient change rate, abrupt change amplitude, and morphological pattern, the system automatically distinguishes between the real interface and internal interference. Finally, in the fine-grained adjustment stage, the effective interface is split or the boundary is corrected, and local interference is merged or maintained. This system maintains high robustness in complex geological environments and avoids over-segmentation or missed judgment caused by noise, uneven illumination, or probe disturbance.

[0040] 3. By using a convolutional neural network model to collect adjustment data for each soil layer, engineering performance coefficients are identified. Combined with depth and spatial location information, the overall foundation bearing capacity is calculated. Subsequently, a time-series analysis algorithm is used to output the foundation bearing capacity attenuation coefficient for a specified future period based on historical environmental loads and future load predictions. Sensitivity analysis can also be performed on the correlation strength between the attenuation coefficient and the internal characteristics and performance coefficients of the soil layer to identify the most influential high-threshold features and soil layer identifiers. This provides precise and targeted suggestions for engineering maintenance and reinforcement. The seamless connection from static identification to dynamic prediction greatly improves the scientificity and timeliness of foundation safety assessment and long-term operation and maintenance decisions. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0042] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0043] Figure 2 This is a schematic diagram of the calibration unit in this invention;

[0044] Figure 3This is a schematic diagram of the identification unit in this invention.

[0045] The labels in the diagram represent: 1. Data acquisition module; 2. Feature matching module; 3. Preliminary stratification module; 4. Calibration unit; 41. Abrupt change trend detection module; 42. Abrupt change feature extraction module; 43. Deep learning modeling module; 5. Soil layer redefinition module; 6. Identification unit; 61. Model training module; 62. Performance calculation module; 63. Attenuation analysis module; 7. Attenuation source tracing module; 8. Data preprocessing module. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] The present invention will be further described below with reference to embodiments.

[0048] Example 1

[0049] This embodiment presents a soil interface recognition model construction system based on intelligent image recognition, such as... Figures 1-2 As shown, it includes:

[0050] Data acquisition module 1 acquires soil depth image data through sensing acquisition equipment, and simultaneously acquires probe detection resistance, side wall friction resistance and pore water pressure signals at the corresponding depth, and performs time and space registration of all acquired data according to depth identifier.

[0051] Feature matching module 2 extracts soil features from images and probe signal features from signal data. It then performs one-to-one matching between the extracted soil features and probe signal features at the same depth to construct a fusion dataset under depth identification. Soil features include: texture, grain size, color distribution, and structural morphology. Probe signal features include: time domain statistics, frequency domain components, and signal morphology.

[0052] The preliminary stratification module 3 divides the fused dataset into several initial soil layers according to the depth range. Each initial soil layer contains the depth range, image features, and probe signal features. It automatically segments according to the depth range, eliminating the need for manual setting of fixed layer thickness and improving adaptability to complex strata.

[0053] Calibration unit 4 receives fused data and preliminary soil layer division results. Through joint analysis of image and signal feature sequences, it detects abrupt change trend segments exceeding a preset threshold, pre-trains a calibration model, inputs abrupt change features extracted within a preset depth range of the abrupt change segments into the trained calibration model, and outputs the soil layer change results of the current preliminary soil layer division. Calibration unit 4 has sub-modules, including an abrupt change trend detection module 41, an abrupt change feature extraction module 42, and a deep learning modeling module 43. The abrupt change feature extraction module 42 interacts with the abrupt change trend detection module 41 and the deep learning modeling module 43 via a wireless network.

[0054] The mutation trend detection module 41 receives the fused dataset and, along a continuous depth direction, jointly analyzes the sequence of image features and the sequence of probe signal features to identify mutation trend segments in the joint feature sequence that exceed a preset dynamic threshold. It identifies mutation trends exceeding the preset dynamic threshold in real time along a continuous depth direction, efficiently capturing potential interface locations. The process of jointly analyzing and identifying mutation trend segments involves: strictly aligning the image feature sequence with the probe signal feature sequence at the same depth according to depth identifiers to ensure synchronization of multimodal data at each depth point; normalizing the texture gradient and granularity distribution change rate in the image feature sequence, and simultaneously normalizing the temporal domain of the probe signal sequence. The statistical value volatility and frequency domain energy offset are quantitatively extracted. A weighted fusion algorithm is used to merge the normalized image feature changes and signal feature fluctuations into a single-dimensional joint feature sequence. The image feature weights and signal feature weights are dynamically allocated according to the feature contribution of historical soil layer interface data. In the continuous depth direction, a dynamic sliding window algorithm is applied to the joint feature sequence to calculate the standard deviation of feature values ​​within the window and the absolute value of feature gradients between adjacent windows in real time. When the standard deviation or gradient value continuously exceeds the dynamic threshold set according to the historical statistical values ​​of soil layer continuity, the window is determined to be a sudden change trend segment, and its starting depth, ending depth, and sudden change intensity index are output.

[0055] The mutation feature extraction module 42 is used to extract all image features and signal feature data within a preset depth range above and below each mutation point or mutation interval identified by the mutation trend detection module 41, and extract mutation features within this depth range; the mutation features include: maximum feature gradient, gradient change rate, feature mutation amplitude and mutation morphology pattern.

[0056] The deep learning modeling module 43 receives mutation features within the depth range output by the mutation feature extraction module and preliminary soil layer division results output by the preliminary stratification module 3. It pre-trains and calibrates the model, using mutation features within the depth range as the core input of the calibration model and incorporating the context information of the preliminary soil layer to determine whether the mutation point or interval represents the true soil interface. During the judgment phase, the deep learning modeling module 43 determines a valid interface when the gradient change rate and mutation amplitude in the mutation feature simultaneously exceed the dynamic threshold and the mutation morphology matches the typical pattern of the soil interface. When the mutation feature only locally exceeds the threshold but the context information indicates that the feature is continuous and stable, it is determined to be internal soil disturbance. Based on the judgment results, it outputs soil layer modification instructions: performing soil layer splitting or boundary correction on valid interfaces, and performing soil layer merging or maintaining the original division on internal disturbances. Combining mutation features and preliminary stratification context, it has online learning and self-optimization capabilities.

[0057] The soil layer redefinition module 5 is used to receive the soil layer modification results output by the calibration unit 4, and to perform fine-grained correction, merging or splitting of the initially divided soil layers, generating adjustment acquisition data corresponding to each final accurate soil layer. All original or aggregated image features within the depth range of the soil layer represent the overall or typical soil properties of the layer and the corresponding statistical signal data of probe detection resistance, sidewall friction, and pore water pressure.

[0058] The identification unit 6 is used to automatically identify soil characteristics based on the adjusted collection data obtained from the soil redefinition module 5, and output the performance coefficient of the soil layer, as well as the decay coefficient of the performance coefficient in the future cycle.

[0059] The attenuation tracing module 7 is used to analyze the correlation strength between the attenuation coefficient and the internal characteristics and performance coefficients of each soil layer, and to identify and output one or more soil layer characteristics and their soil layer identifiers that have the greatest impact on the predicted attenuation coefficient and have a high preset threshold influence.

[0060] The data acquisition module 1 is wirelessly connected to the data preprocessing module 8. The data preprocessing module 8 is used to perform grayscale normalization, anisotropic diffusion filtering, and illumination compensation on the soil depth image to eliminate acquisition noise and non-uniform illumination interference. It performs three-level processing on the probe detection resistance, sidewall friction resistance, and pore water pressure signals, including: suppressing impulse noise through sliding window midpoint filtering; eliminating high-frequency interference by wavelet threshold denoising; and normalizing the amplitude of multi-source signals by using depth marking.

[0061] Feature matching module 2 is interconnected with data acquisition module 1 and preliminary stratification module 3 via a wireless network; calibration unit 4 is interconnected with preliminary stratification module 3 and soil layer redefinition module 5 via a wireless network; and identification unit 6 is interconnected with soil layer redefinition module 5 and attenuation tracing module 7 via a wireless network.

[0062] Compared with existing technologies, this technology constructs an automated chain from raw data acquisition, preprocessing, and feature extraction to dynamic hierarchical calibration, performance evaluation, and factor analysis. The complementary use of images and mechanical signals overcomes the limitations of weak anti-interference capabilities of single sensors, significantly improves the accuracy of soil interface identification, detects abrupt change trends online, does not rely on manual experience thresholds, can self-calibrate and continuously optimize, and effectively distinguishes between real interfaces and internal interference by combining contextual information and morphological pattern judgment. It not only performs hierarchical identification but also outputs performance degradation predictions and key factor analyses, providing quantitative basis for engineering monitoring, maintenance, and optimization.

[0063] Example 2

[0064] At other levels, this embodiment also provides another optimization mechanism based on Embodiment 1, specifically a formula for calculating the overall foundation bearing capacity as follows:

[0065]

[0066] In the formula, Q u P represents the overall bearing capacity of the foundation, k represents the soil layer synergy coefficient k∈[0.8, 1.2], n represents the total number of soil layers, and P i h represents the performance coefficient of the i-th soil layer. i d represents the thickness of the i-th soil layer, γ represents the depth reinforcement factor γ∈[0.soil layer redefinition module 5, 0.15], d i D represents the depth of the midpoint of the i-th soil layer. ref The reference depth is represented by a fixed value, λ represents the interlayer interaction attenuation coefficient λ∈[0.1, 0.3], and Δz i P represents the gradient difference in performance coefficients between the i-th layer and its adjacent soil layers, η represents the sensitivity weight of the weak layer η∈[0.2, 0.5], and P j Represents the j-th weakest soil layer. H represents the set of weakest soil layers. total Represents the total computation depth.

[0067] In this embodiment, multiple factors such as soil performance coefficient, thickness, depth, interlayer differences, and the influence of the weakest layer are fully integrated to break through the limitations of traditional single index calculation. The depth enhancement factor reflects the enhancement effect of soil burial depth on bearing capacity. The interlayer interaction attenuation coefficient is introduced to accurately quantify the mutual influence of performance differences between adjacent soil layers. The logarithmic function is used to handle the influence of the weakest layer, which effectively captures the global constraint effect of key weak layers and avoids over-amplifying the influence of deeply buried weak layers.

[0068] Example 3

[0069] In this embodiment, as Figure 3As shown, the identification unit 6 has sub-modules deployed below it. These sub-modules include a model training module 61, a performance calculation module 62, and a decay analysis module 63. The performance calculation module 62 is interconnected with the model training module 61 and the performance calculation module 63 via a wireless network.

[0070] Model training module 61 is used to construct a recognition model using a convolutional neural network algorithm. It trains the recognition model based on historical soil layer adjustment data and known engineering performance indicators such as compression modulus, permeability coefficient, and shear strength. The recognition model receives adjustment data for a specific soil layer, automatically identifies the characteristics of that soil layer, and outputs a performance coefficient characterizing the key engineering performance of that soil layer. The recognition model construction process involves: structuring the historical soil layer adjustment data according to soil layer identifiers; each soil layer's sample data includes its corresponding texture, grain size, color distribution features, and probe signal feature vectors, while also associating it with known engineering performance indicator labels for that soil layer; and constructing a bi-branch convolutional neural network. The neural network has two branches: the first branch processes image feature vectors, extracting spatial hierarchical features through convolutional layers; the second branch processes signal feature vectors, extracting temporal pattern features through one-dimensional convolutional layers. The high-level features output from the two branches are concatenated and then input into a fully connected layer. The dual-branch network is trained end-to-end using sample data, with engineering performance index labels used as supervision signals. An adaptive gradient optimization algorithm is employed to minimize the mean square error between the network output value and the actual performance index. After training, the network receives adjustment data from a new input soil layer. After dual-branch feature extraction and fusion, the fully connected layer outputs a continuous value representing the key engineering performance of the soil layer, i.e., the performance coefficient.

[0071] The performance calculation module 62 is used to calculate the overall foundation bearing capacity based on the performance coefficients of each soil layer output by the model training module 61, combined with the depth and spatial location information of each soil layer.

[0072] The attenuation analysis module 63 uses the overall foundation bearing capacity output by the performance calculation module 62 as a benchmark value, integrates and analyzes historical environmental loads, future load predictions, and time factors, and obtains the attenuation coefficient of the overall foundation bearing capacity within a specified future period through a time-series analysis algorithm. The process of time-series analysis performed by the attenuation analysis module 63 is as follows:

[0073] Based on historical and future environmental load predictions, the performance coefficients of each soil layer at the end of a specified period are predicted using a pre-trained soil performance degradation model.

[0074] Based on the updated performance coefficients of each soil layer, combined with soil layer depth and spatial location information, the future overall foundation bearing capacity is calculated using the same calculation strategy as performance calculation module 62.

[0075] The ratio of the future overall foundation bearing capacity to the current overall foundation bearing capacity is used as the attenuation coefficient.

[0076] Compared with existing technologies, this method automatically extracts soil layer features and outputs performance coefficients based on a convolutional neural network recognition model. Combined with real-time collected soil layer depth and spatial location information, it accurately calculates the current overall foundation bearing capacity. Furthermore, it integrates historical environmental loads and future load predictions, and uses a time-series analysis algorithm to dynamically simulate soil layer performance. By observing the evolution over time, it obtains the foundation bearing capacity decay trend in future cycles. This achieves full-process automation and high-precision analysis from soil layer identification to bearing capacity assessment and decay prediction, providing scientific and reliable decision support for engineering design and operation and maintenance.

[0077] In summary, this invention utilizes images and probe signals to perform preprocessing, feature extraction, and depth registration on the images and signals sequentially, constructing a fused dataset. Based on dynamic threshold mutation detection and a pre-trained calibration model, it automatically completes the splitting and merging of the initial soil layers, ultimately generating accurate soil layer interfaces and their corresponding image and signal statistical features.

[0078] Based on this, the system predicts the key engineering performance coefficients of each soil layer through a convolutional neural network, and calculates the overall foundation bearing capacity and its future attenuation coefficient by combining a time series analysis model. At the same time, it explores the correlation strength between the attenuation coefficient and soil layer characteristics and identifies the key features with the greatest impact.

[0079] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A soil interface recognition model construction system based on intelligent image recognition, characterized in that, include: The data acquisition module acquires soil depth image data through sensing acquisition devices, and simultaneously acquires probe detection resistance, side wall friction and pore water pressure signals at the corresponding depths. All acquired data are registered in time and space according to depth identifiers. The feature matching module extracts soil features from images and probe signal features from signal data. It then performs one-to-one matching between the extracted soil features and probe signal features at the same depth to construct a fusion dataset under depth labeling. The initial layering module divides the fused dataset into several initial soil layers according to the depth range. Each initial soil layer contains the depth range, image features, and probe signal features. The calibration unit is used to receive fused data and preliminary soil layer division results, detect abrupt trend segments exceeding preset thresholds by jointly analyzing image and signal feature sequences, pre-train the calibration model, input the abrupt features extracted within the preset depth range of the abrupt segments into the trained calibration model, and output the soil layer change results of the current preliminary soil layer division results. The soil layer redefinition module is used to receive the soil layer modification results output by the calibration unit, and to perform fine-grained correction, merging or splitting of the initially divided soil layers to generate adjustment acquisition data corresponding to each final accurate soil layer. The identification unit is used to automatically identify soil characteristics based on the adjusted acquisition data obtained from the soil redefinition module, and output the performance coefficient of the soil layer, as well as the decay coefficient of the performance coefficient in the future cycle. The attenuation tracing module is used to analyze the correlation strength between the attenuation coefficient and the internal characteristics and performance coefficients of each soil layer, and to identify and output one or more soil layer characteristics and their soil layer identifiers that have the greatest impact on the predicted attenuation coefficient and have a high preset threshold influence.

2. The soil interface recognition model construction system based on intelligent image recognition according to claim 1, characterized in that, The soil features of the feature matching module include: texture, particle size, color distribution, and structural morphology; the probe signal features include: time-domain statistics, frequency-domain components, and signal morphology.

3. The soil interface recognition model construction system based on intelligent image recognition according to claim 1, characterized in that, The calibration unit has sub-modules deployed at its lower level, including a mutation trend detection module, a mutation feature extraction module, and a deep learning modeling module. The mutation feature extraction module interacts with the mutation trend detection module and the deep learning modeling module via a wireless network, wherein: The mutation trend detection module receives the fused dataset and, along a continuous depth direction, jointly analyzes the sequence of image features and the sequence of probe signal features to identify mutation trend segments in the joint feature sequence that exceed a preset dynamic threshold. The process of jointly analyzing and identifying mutation trend segments involves: strictly aligning the image feature sequence with the probe signal feature sequence at the same depth according to depth identifiers to ensure synchronization of multimodal data at each depth point; normalizing the texture gradient and granularity distribution change rate in the image feature sequence; and quantizing the temporal statistical volatility and frequency domain energy offset in the probe signal sequence. The algorithm employs a weighted fusion method to merge the normalized image feature changes and signal feature fluctuations into a single-dimensional joint feature sequence. The image feature weights and signal feature weights are dynamically allocated based on the feature contribution of historical soil interface data. In the continuous depth direction, a dynamic sliding window algorithm is applied to the joint feature sequence to calculate the standard deviation of feature values ​​within the window and the absolute value of feature gradients between adjacent windows in real time. When the standard deviation or gradient value continuously exceeds the dynamic threshold set based on the historical statistical values ​​of soil continuity, the window is determined to be a sudden change trend segment, and its starting depth, ending depth, and sudden change intensity index are output. The mutation feature extraction module is used to extract all image features and signal feature data within a preset depth range above and below each mutation point or mutation interval identified by the mutation trend detection module, and extract the mutation features within this depth range. The deep learning modeling module is used to receive mutation features within the depth range output by the mutation feature extraction module, receive the preliminary soil layer division results output by the preliminary layering module, pre-train and calibrate the model, use the mutation features within the depth range as the core input of the calibration model, and integrate and consider the preliminary soil layer context information in which they are located to determine whether the mutation point or interval represents the real soil layer interface.

4. The soil interface recognition model construction system based on intelligent image recognition according to claim 3, characterized in that, The mutation feature extraction module extracts mutation features including: maximum feature gradient, gradient change rate, feature mutation amplitude, and mutation morphology pattern.

5. The soil interface recognition model construction system based on intelligent image recognition according to claim 3, characterized in that, In the judgment phase, the deep learning modeling module determines a valid interface when the gradient change rate and mutation amplitude in the mutation feature simultaneously exceed the dynamic threshold and the mutation morphology matches the typical pattern of the soil interface; when the mutation feature only locally exceeds the threshold but the context information indicates that the feature is continuous and stable, it is determined to be an internal soil disturbance; based on the judgment result, a soil layer modification instruction is output: soil layer splitting or boundary correction is performed on the valid interface, and soil layer merging or maintaining the original division is performed on the internal disturbance.

6. The soil interface recognition model construction system based on intelligent image recognition according to claim 1, characterized in that, The identification unit has sub-modules deployed at its lower level. These sub-modules include a model training module, a performance calculation module, and a decay analysis module. The performance calculation module interacts with the model training module and the performance calculation module via a wireless network. The model training module is used to build a recognition model using a convolutional neural network algorithm. It trains the model based on historical soil layer adjustment data and known engineering performance indicators. The recognition model receives adjustment data for a specific soil layer, automatically identifies the characteristics of that soil layer, and outputs a performance coefficient characterizing its key engineering properties. The model construction process involves: structuring the historical soil layer adjustment data according to soil layer identifiers; each soil layer's sample data includes its corresponding texture, grain size, color distribution features, and probe signal feature vectors, while also associating it with known engineering performance indicator labels for that soil layer; and constructing a two-branch convolutional neural network, the first branch... The first branch is configured to process image feature vectors, extracting spatial hierarchical features through convolutional layers. The second branch is configured to process signal feature vectors, extracting temporal pattern features through one-dimensional convolutional layers. The high-level features output from the two branches are concatenated and then input into a fully connected layer. The dual-branch network is trained end-to-end using sample data, with engineering performance index labels as supervision signals. An adaptive gradient optimization algorithm is used to minimize the mean square error between the network output value and the true performance index. After training, the network receives adjustment data collected from new input soil layers. After dual-branch feature extraction and fusion, the fully connected layer outputs continuous values ​​representing the key engineering performance of the soil layer, i.e., performance coefficients. The performance calculation module is used to calculate the overall foundation bearing capacity based on the performance coefficients of each soil layer output by the model training module, combined with the depth and spatial location information of each soil layer. The attenuation analysis module uses the overall foundation bearing capacity output by the performance calculation module as a benchmark value, integrates and analyzes historical environmental loads, future load predictions, and time factors, and obtains the attenuation coefficient of the overall foundation bearing capacity within a specified future period through a time-series analysis algorithm.

7. The soil interface recognition model construction system based on intelligent image recognition according to claim 6, characterized in that, The formula for calculating the overall foundation bearing capacity of the performance calculation module is as follows: In the formula, Q u κ represents the overall bearing capacity of the foundation, n represents the soil synergy coefficient, and P represents the total number of soil layers. i h represents the performance coefficient of the i-th soil layer. i d represents the thickness of the i-th soil layer, γ represents the depth strengthening factor, and d represents the depth strengthening factor. i D represents the depth of the midpoint of the i-th soil layer. ref The reference depth is represented by a fixed value, λ represents the interlayer interaction attenuation coefficient, and Δz i P represents the gradient difference in performance coefficients between the i-th layer and its adjacent soil layers, η represents the sensitivity weight of the weak layer, and P represents the sensitivity weight of the weak layer. j Represents the j-th weakest soil layer. H represents the set of weakest soil layers. total Represents the total computation depth.

8. The soil interface recognition model construction system based on intelligent image recognition according to claim 6, characterized in that, The process of time-series analysis performed by the attenuation analysis module is as follows: Based on historical and future environmental load predictions, the performance coefficients of each soil layer at the end of a specified period are predicted using a pre-trained soil performance degradation model. Based on the updated performance coefficients of each soil layer, combined with soil layer depth and spatial location information, the future overall foundation bearing capacity is calculated using the same calculation strategy as the performance calculation module. The ratio of the future overall foundation bearing capacity to the current overall foundation bearing capacity is used as the attenuation coefficient.

9. The soil interface recognition model construction system based on intelligent image recognition according to claim 1, characterized in that, The data acquisition module is interconnected with a data preprocessing module via a wireless network. The data preprocessing module performs grayscale normalization, anisotropic diffusion filtering, and illumination compensation on the soil depth image to eliminate acquisition noise and non-uniform illumination interference. It also performs three-level processing on the probe detection resistance, sidewall friction, and pore water pressure signals, including: suppressing impulse noise through sliding window midpoint filtering; eliminating high-frequency interference by wavelet threshold denoising; and normalizing the amplitude of multi-source signals using depth marking.

10. The soil interface recognition model construction system based on intelligent image recognition according to claim 1, characterized in that, The feature matching module is interconnected with the data acquisition module and the preliminary stratification module via a wireless network. The calibration unit is interconnected with the preliminary stratification module and the soil layer redefinition module via a wireless network. The identification unit is interconnected with the soil layer redefinition module and the attenuation tracing module via a wireless network.

Images