DeepSeek-based concrete internal crack identification method and system

Abstract

The invention aims to provide a method and a system for identifying internal cracks of concrete based on a DeepSeek framework, and belongs to the technical field of nondestructive testing of concrete. The method comprises the following steps of: acquiring a multi-modal physical signal in the concrete to generate a standardized frequency spectrum characteristic pattern; constructing a DeepSeek-R1 transfer learning model based on a hybrid expert architecture, and performing field adaptation fine tuning on a gating network and an expert network group by freezing encoder module parameters, dynamically adjusting a learning rate and a cross entropy-Dice hybrid loss function, and optimizing the sensitivity of the model to crack characteristics; an edge-cloud distributed deployment architecture is adopted, an encoder module sinks to an edge computing node to realize real-time signal processing, and a gating network and an expert network group complete high-precision reasoning at the cloud. According to the invention, automatic and high-precision detection of concrete internal cracks is realized, and generalization ability and real-time performance of cross-engineering scenes are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete non-destructive testing, and in particular to a method and system for identifying internal cracks in concrete based on DeepSeek. Background Art

[0002] Detecting internal cracks in concrete is a crucial step in building structural safety assessments. Traditional methods rely on manual experience or single sensor technology, resulting in low efficiency and insufficient precision. While existing deep learning models attempt to improve detection capabilities through multi-sensor fusion, they generally face two major challenges: First, they rely on massive amounts of labeled data, but concrete crack samples are expensive to obtain and have large annotation differences across scenarios, limiting the model's generalization capabilities. Second, conventional transfer learning methods have poor adaptability to cross-modal data (such as acoustic and electromagnetic signals), and feature shifts are prone to occur after model fine-tuning, making it impossible to effectively extract the geometric and physical properties of cracks. Furthermore, traditional Mixed of Experts (MoE) models suffer from high computational redundancy in engineering inspection scenarios, making it difficult to meet real-time requirements. Furthermore, they lack parameter optimization for small sample data, limiting their practical application value.

[0003] To address these issues, an improved approach based on the DeepSeek framework provides a new technical path for concrete crack identification. The DeepSeek model, through its hybrid expert architecture and dynamic gating mechanism, flexibly adapts to the characteristic distribution of multimodal data. However, its original design is geared towards general tasks and is not optimized for the physical characteristics of concrete internal signals or small sample scenarios. For example, it fails to address issues such as sensor noise interference and the inefficiency of weight distribution when migrating across projects. Implementing domain-adapted transfer learning based on the DeepSeek framework and building a lightweight distributed inference system are key breakthroughs in improving the accuracy of automated inspection.

[0004] How to achieve high-precision, cross-scenario real-time detection of internal cracks in concrete with limited labeled data based on the hybrid expert architecture of the DeepSeek model through transfer learning strategies and edge-cloud collaborative computing has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a method and system for identifying internal cracks in concrete based on DeepSeek. Spectral features are generated by fusing ultrasonic and electromagnetic multimodal signals. The encoder parameters are frozen and the gated network and expert network group are dynamically optimized in combination with a transfer learning strategy to achieve high-precision extraction of crack features in small sample scenarios. An edge-cloud distributed architecture deployment model is adopted, with the encoder module sinking to the edge node to process real-time signals, and high-confidence inference is completed in the cloud. The three-dimensional parameters of the cracks are output by combining attention heat maps with adaptive threshold segmentation, significantly improving detection efficiency and cross-scenario generalization capabilities.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for identifying internal cracks in concrete based on DeepSeek, comprising the following steps:

[0008] Collect multimodal detection signals inside the concrete and perform preprocessing to generate standardized spectrum feature maps;

[0009] Construct a DeepSeek-R1 transfer learning model based on a hybrid expert architecture, which includes an encoder module, a gating network, and an expert network group. The encoder module is used to extract multi-scale spatial features from the spectral feature map, and the gating network selects the activated expert network through a weight distribution mechanism.

[0010] Freezing the parameters of the encoder module and performing transfer learning fine-tuning on the gating network and the expert network group based on a concrete crack dataset, wherein the transfer learning fine-tuning includes dynamically adjusting the learning rate and optimizing the weight distribution according to the sensitivity of the crack features;

[0011] The frequency spectrum feature map is input into the fine-tuned model to output the three-dimensional coordinate information of the crack.

[0012] Beneficial effects: By collecting multimodal detection signals inside the concrete and preprocessing them to generate standardized spectral feature maps, and then using the DeepSeek-R1 transfer learning model based on a hybrid expert architecture for crack identification, it can effectively integrate the advantages of multiple modal signals, make full use of the encoder module to extract multi-scale spatial features, and the synergy of the gated network and the expert network group. Combined with the transfer learning fine-tuning strategy of dynamically adjusting the learning rate and optimizing the weight distribution, it achieves high-precision extraction of internal crack features of concrete in small sample scenarios, providing a reliable basis for the subsequent accurate positioning and evaluation of cracks, and significantly improving the accuracy and efficiency of concrete crack detection.

[0013] Furthermore, the multimodal detection signal includes:

[0014] The acoustic wave signal collected by the ultrasonic sensor array has a sampling frequency range of 20kHz-1MHz;

[0015] The electromagnetic signal collected by the electromagnetic sensor array has an operating frequency range of 10MHz-1GHz.

[0016] Beneficial Effects: The multimodal detection signal is clearly defined as including acoustic signals collected by an ultrasonic sensor array with a sampling frequency range of 20kHz-1MHz and electromagnetic signals collected by an electromagnetic sensor array with an operating frequency range of 10MHz-1GHz, making the detection signal sources more diverse and targeted. Ultrasonic signals can detect deep into the concrete and are sensitive to information such as the depth of the crack; electromagnetic signals have a good response to the surface and near-surface characteristics of the crack. The combination of the two covers the characteristic information of concrete cracks at different locations and scales, further improving the comprehensiveness and accuracy of crack identification, and providing more sufficient data support for the precise analysis and judgment of subsequent models.

[0017] Furthermore, the preprocessing includes:

[0018] Perform time-frequency domain transformation on multimodal signals to generate a spectrum diagram;

[0019] The frequency bands whose energy proportion exceeds the preset threshold are extracted by wavelet packet decomposition, and the signal segments are intercepted by a sliding window with a window length of T seconds and a step size of 0.5T.

[0020] Beneficial effects: Transforming the multimodal signal in the time-frequency domain to generate a spectrum can convert the signal from the time domain to the frequency domain, facilitating subsequent analysis of the signal's characteristics at different frequencies and more clearly showing the manifestation of defects such as cracks in the frequency domain. By extracting frequency bands whose energy ratio exceeds a preset threshold through wavelet packet decomposition, noise interference can be effectively removed, focusing on key frequency band information related to crack characteristics; using a sliding window to intercept signal segments with a window length of T seconds and a step size of 0.5T, the signal can be processed in segments, allowing the model to analyze signal characteristics segment by segment and better capture the changing patterns of cracks in the time series, thereby further improving the quality and representativeness of the pre-processed signal and laying a solid foundation for the accurate identification of the model.

[0021] Furthermore, each expert network in the expert network group includes a cross attention mechanism and a residual convolution layer, and the activation function is a ReLU function.

[0022] Beneficial effects: Each expert network in the expert network group contains a cross-attention mechanism and a residual convolution layer, and the activation function is the ReLU function. The cross-attention mechanism enables different expert networks to collaborate with each other, focusing on different aspects of crack characteristics, enhancing the model's global perception of crack characteristics, and avoiding misjudgments caused by a single expert network's excessive focus on local features. The residual convolution layer can alleviate the gradient vanishing problem during deep network training, improve the model's training efficiency and stability, and enable the model to better learn the complex mapping relationships of crack characteristics. The ReLU activation function has the advantages of simple calculation, the ability to introduce nonlinear factors, and the avoidance of gradient vanishing when the neuron output is zero. It helps to improve the model's convergence speed and learning ability, further optimizing the model's structure and performance, and making it show higher accuracy and robustness in crack identification tasks.

[0023] Furthermore, the training strategy for transfer learning fine-tuning includes:

[0024] The weighted sum of cross entropy loss and Dice coefficient is used as the loss function;

[0025] The initial learning rate is set to η, and decays to 0.9η after every k batches of training.

[0026] Beneficial Effects: Using a weighted sum of cross-entropy loss and the Dice coefficient as the loss function, combined with an initial learning rate set to η and decaying to 0.9η after every k training batches, this method comprehensively considers both classification accuracy and segmentation quality, and can simultaneously optimize the model's classification of cracks and non-crack areas, as well as the segmentation accuracy of crack areas. Dynamically adjusting the learning rate allows for flexible adjustments to the learning step size based on the convergence during training, avoiding model oscillation caused by excessively high learning rates in the early stages and slow convergence caused by excessively low learning rates in the later stages. This allows the model to more efficiently approach the optimal solution during training, further improving the model's training effectiveness and generalization capabilities, and providing a strong guarantee for high-precision crack identification.

[0027] Furthermore, the weight coefficients α and β of the loss function satisfy α+β=1, wherein the value range of α is 0.6-0.8.

[0028] Beneficial Effects: Limiting the loss function's weight coefficients α and β to satisfy α + β = 1, with α ranging from 0.6 to 0.8, effectively balances the weights of the cross-entropy loss and the Dice coefficient in the loss function. Within this range, the model prioritizes the role of the cross-entropy loss, effectively improving the model's accuracy in crack classification while also optimizing the Dice coefficient for crack segmentation. This allows the model to accurately distinguish between cracks and non-crack areas and accurately segment crack boundaries in crack identification tasks, further improving the overall performance of crack identification and better meeting the dual demands for crack detection accuracy and segmentation quality in practical applications.

[0029] Furthermore, the three-dimensional coordinate information of the crack is generated by the following steps:

[0030] Generate an attention heat map based on the weight matrix output by the gating network;

[0031] The center point coordinates of the area exceeding the threshold γ in the heat map are calculated, and the crack depth is determined by combining the output of the residual convolution layer.

[0032] Beneficial Effects: By generating an attention heatmap based on the weight matrix output by the gating network, calculating the center point coordinates of regions in the heatmap exceeding the threshold γ, and combining this with the output of the residual convolutional layer to determine the crack depth, the weight distribution results of the gating network can be intuitively converted into a visual heatmap, facilitating analysis and understanding of the model's focus on crack characteristics. The center point coordinate calculation based on the heatmap accurately locates the crack position, and combined with the depth information output of the residual convolutional layer, the crack depth is further determined, achieving complete output of the three-dimensional coordinate information of the crack. This provides comprehensive and accurate parameter support for subsequent quantitative crack assessment and structural safety analysis, enhancing the application value of the crack identification results.

[0033] Furthermore, the threshold γ is dynamically adjusted based on the concrete material properties within a range of 0.6-0.8, allowing the threshold setting to better adapt to the differences in crack characteristics caused by different concrete material properties. Different concrete materials vary in composition, density, and internal structure, and cracks appear differently in the thermal map. By dynamically adjusting the threshold, areas related to crack characteristics can be more accurately screened, avoiding misidentifications or missed detections caused by a fixed threshold. This further improves the accuracy and adaptability of crack identification and enhances the model's generalization capabilities across diverse engineering scenarios.

[0034] In a second aspect, the present invention provides a concrete internal crack identification system based on DeepSeek, comprising:

[0035] a data acquisition module configured to acquire ultrasonic and electromagnetic sensor signals;

[0036] A preprocessing module configured to perform the preprocessing steps of the method as described in any of the preceding items;

[0037] The model inference module deploys the fine-tuned DeepSeek-R1 model and outputs the crack parameters.

[0038] Beneficial effects: A system for identifying internal cracks in concrete based on DeepSeek is provided, which includes a data acquisition module, a preprocessing module, and a model inference module, realizing full-process automation from data acquisition to crack parameter output. The data acquisition module can acquire ultrasonic and electromagnetic sensor signals to provide basic data for subsequent analysis; the preprocessing module can perform the preprocessing steps of any of the aforementioned methods to ensure the data quality of the input model; the model inference module deploys the fine-tuned DeepSeek-R1 model, which can output crack parameters efficiently and accurately. The system organically combines various functional modules to form a complete crack identification solution, which improves the automation level and work efficiency of concrete crack detection, reduces the cost of manual intervention, and has good prospects for engineering application.

[0039] Furthermore, the model inference module adopts a distributed architecture: the encoder module is deployed on the edge computing node, and the gated network and expert network group are deployed on the cloud server; the edge computing node and the cloud server transmit data through the 5G communication protocol.

[0040] Beneficial effects: The model inference module adopts a distributed architecture, the encoder module is deployed on the edge computing node, the gating network and expert network group are deployed on the cloud server, and the edge computing node and the cloud server transmit data via the 5G communication protocol. This architecture fully leverages the advantages of edge computing and cloud computing. The encoder module performs preliminary processing of real-time signals at the edge node, which can quickly respond to field data and reduce data transmission delays. The gating network and expert network group are deployed on the cloud server and can use powerful computing resources to complete high-confidence reasoning to ensure the accuracy and reliability of crack identification. The high-speed transmission characteristics of the 5G communication protocol ensure efficient transmission of data between the edge node and the cloud server, enabling the entire system to achieve more accurate crack identification and analysis while ensuring real-time performance, further improving the system's performance and practicality, and meeting the comprehensive requirements for real-time, high-precision, and efficient concrete crack detection in actual engineering.

[0041] In summary, compared with the existing technology, the present invention generates a standardized spectral feature map by fusing ultrasonic and electromagnetic multimodal signals, and constructs a DeepSeek-R1 transfer learning model based on a hybrid expert architecture, thereby achieving high-precision extraction of internal crack features in concrete in small sample scenarios. At the same time, the model is deployed using an edge-cloud distributed architecture, with the encoder module sinking to the edge node to process real-time signals, and high-confidence reasoning is completed in the cloud. The three-dimensional parameters of the cracks are output by combining attention heat maps and adaptive threshold segmentation, which significantly improves the detection efficiency and cross-scenario generalization capability. In addition, the present invention further optimizes the training effect and performance of the model through strategies such as dynamically adjusting the learning rate, optimizing weight distribution, and adopting a loss function that is a weighted sum of cross-entropy loss and Dice coefficient. It effectively solves the problems of low efficiency and insufficient accuracy caused by traditional methods relying on manual experience or single sensor technology, as well as the limited generalization capability of existing deep learning models under small sample data. It provides a new technology path for the automated detection of internal cracks in concrete that is efficient, accurate, and highly adaptable, and has important engineering application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present invention and, together with the description, serve to explain the principles of the present invention.

[0043] The present invention can be more clearly understood from the following detailed description with reference to the accompanying drawings, in which:

[0044] Figure 1 is an overall flow chart of a method for identifying internal cracks in concrete based on DeepSeek provided by an embodiment of the present invention;

[0045] Figure 2 This is a sub-flowchart of the preprocessing part of the method for identifying internal cracks in concrete based on DeepSeek provided by an embodiment of the present invention;

[0046] Figure 3 2 is a diagram of the DeepSeek-R1 transfer learning model architecture of the DeepSeek-based concrete internal crack identification method provided by an embodiment of the present invention;

[0047] Figure 4 4 is an architecture diagram of a concrete internal crack identification system based on DeepSeek provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations of the technical solution of the present application. In the absence of conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0049] The term "and / or" in this document simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " in this document generally indicates an "or" relationship between the related objects.

[0050] Example 1

[0051] like Figure 1 As shown, the embodiment of the present invention provides a method for identifying internal cracks in concrete based on DeepSeek, comprising the following steps:

[0052] Step S1: collecting multimodal acoustic wave signals and electromagnetic signals inside the concrete through an ultrasonic sensor array and an electromagnetic sensor array to generate original detection data;

[0053] Step S2: preprocessing the original detection data, including signal normalization, noise suppression and time-frequency domain transformation, to generate a standardized spectrum feature map;

[0054] Step S3: Figure 3 As shown, a DeepSeek-R1 transfer learning model based on a mixture of experts (MoE) architecture is constructed. The DeepSeek-R1 transfer learning model includes:

[0055] Encoder module, used to extract multi-scale spatial features in the spectral feature map;

[0056] Gating network, which selects the activated expert network through a weight distribution mechanism;

[0057] Expert network group, each expert network includes a cross-attention mechanism and a residual convolution layer;

[0058] In step S4, in the DeepSeek-R1 transfer learning model, the parameters of the encoder module are frozen, and only the gate network and the expert network group are fine-tuned, and the transfer learning strategy is used to optimize the model's sensitivity to crack characteristics;

[0059] Step S5: inputting the standardized spectral feature map into the fine-tuned DeepSeek-R1 model to output the three-dimensional coordinate information of the position, length and depth of the crack;

[0060] Step S6 generates a visual crack distribution map based on the output results, and screens the effective crack areas through a threshold segmentation algorithm.

[0061] Furthermore, in step S1:

[0062] The sampling frequency of the ultrasonic sensor array is 20kHz-1MHz, and the operating frequency of the electromagnetic sensor array is 10MHz-1GHz;

[0063] The multimodal acoustic wave signal includes time domain waveform data of longitudinal waves, shear waves and surface waves.

[0064] Further, if Figure 2 As shown, the preprocessing in step S2 further includes:

[0065] Performing wavelet packet decomposition on the original detection data to extract frequency bands whose energy ratio exceeds a preset threshold;

[0066] The frequency band is segmented and intercepted by a sliding window, with a window length of T seconds and a step length of 0.5T seconds. Furthermore, the activation function of the expert network group in step S3 is:

[0067] f(x)=ReLU(W i x+b i )

[0068] Among them, W i is the weight matrix (weight parameter set) of the i-th expert network, and its dimension is the output feature dimension d out ×Input feature dimension d in ; b i is the bias term (offset) of the i-th expert network, with a dimension of d out ×1; x is the input feature vector, dimension is d in ×1; · is the matrix multiplication operator; ReLU is the rectified linear unit activation function, defined as max(0,·).

[0069] Furthermore, the transfer learning strategy of step S4 includes:

[0070] The gated network and the expert network group are trained using the concrete crack dataset, and the loss function is the weighted sum of the cross entropy loss and the Dice coefficient:

[0071] L=αL CE +βL Dice

[0072] Among them, L is the total loss function value; α is the cross entropy loss L CEThe weight coefficient (α∈[0,1]); β is the Dice coefficient loss L Dice Weight coefficient (β∈[0,1], and α+β=1); L CE is the cross entropy loss, and its calculation formula is:

[0073]

[0074] N is the total number of samples; C is the number of crack categories (such as “no cracks”, “minor cracks”, “serious cracks”, etc.); y c is the indicator value (0 or 1) of category c in the true label; p c is the probability of the model predicting category c; L Dice is the Dice coefficient loss, and the calculation formula is:

[0075]

[0076] p i is the predicted value of the i-th sample (continuous value, range [0,1]); y i is the true label of the i-th sample (0 or 1); ∈ is the smoothing coefficient (to prevent the denominator from being 0, the default is 1×10 -5 );

[0077] During the training process, a dynamic learning rate adjustment strategy is adopted. The initial learning rate is η, which decays to 0.9η after each k batch of training, that is, η new =0.9η old , η old is the current learning rate value; η new is the adjusted learning rate value; k is the learning rate decay cycle (triggering decay after every k batches of training); 0.9 is the decay coefficient (the learning rate is multiplied by this coefficient).

[0078] Furthermore, the method for locating the crack position in step S5 is:

[0079] Generate an attention heat map through the weight matrix output by the gating network;

[0080] The areas in the thermal map that exceed the threshold γ are marked as candidate crack areas, and the coordinates of their center points are calculated.

[0081] Furthermore, the threshold segmentation algorithm in step S6 adopts the adaptive Otsu algorithm, including:

[0082] The segmentation threshold is dynamically adjusted according to the crack depth distribution histogram and is calculated using the following formula:

[0083]

[0084] t is the segmentation threshold to be optimized; H(d) is the crack depth distribution histogram, d represents the depth value; ω0(t) is the pixel ratio below the threshold t. ω1(t) is the percentage of pixels above the threshold t, ω1(t) = 1-ω0(t); μ0(t) is the average depth of the area below the threshold t, μ1(t) is the average depth of the area above the threshold t, d max is the maximum crack depth value;

[0085] The segmented area is subjected to a morphological closing operation to eliminate noise points. The morphological closing operation specifically adopts the following formula:

[0086]

[0087] I is the input binary image (0 represents background, 1 represents crack); B is the structural element (usually a 3×3 rectangle); For expansion operations (enlargement of the crack area); is an erosion operation (shrinking the region to remove noise);

[0088] As an example, Figure 4 As shown, the present invention provides a concrete internal crack identification system based on DeepSeek, comprising:

[0089] A data acquisition module, used for acquiring multimodal detection signals through an ultrasonic sensor array and an electromagnetic sensor array;

[0090] Preprocessing module, used to normalize, suppress noise and transform the original signal into time-frequency domain;

[0091] A model inference module deploys any of the aforementioned DeepSeek-R1 transfer learning models and outputs crack parameters;

[0092] Visualization module for generating 3D crack distribution maps and marking dangerous areas.

[0093] Furthermore, the model reasoning module adopts a distributed computing architecture:

[0094] The encoder module is deployed on the edge computing node;

[0095] The gated network and expert network groups are deployed on cloud servers.

[0096] Preferably, the DeepSeek-based concrete internal crack identification system further includes:

[0097] Data augmentation unit, used to perform elastic deformation, Gaussian noise injection and frequency domain masking operations on the training dataset;

[0098] Hardware acceleration unit, used to implement parallel computing of model inference.

[0099] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying internal cracks in concrete based on DeepSeek, characterized in that: include: Collect multimodal detection signals inside the concrete and perform preprocessing to generate standardized spectrum feature maps; Construct a DeepSeek-R1 transfer learning model based on a hybrid expert architecture, which includes an encoder module, a gating network, and an expert network group. The encoder module is used to extract multi-scale spatial features from the spectral feature map, and the gating network selects the activated expert network through a weight distribution mechanism. Freezing the parameters of the encoder module and performing transfer learning fine-tuning on the gating network and the expert network group based on a concrete crack dataset, wherein the transfer learning fine-tuning includes dynamically adjusting the learning rate and optimizing the weight distribution according to the sensitivity of the crack features; The frequency spectrum feature map is input into the fine-tuned model to output the three-dimensional coordinate information of the crack.

2. The method for identifying internal cracks in concrete based on DeepSeek according to claim 1, characterized in that: The multimodal detection signal includes: The acoustic wave signal collected by the ultrasonic sensor array has a sampling frequency range of 20kHz-1MHz; The electromagnetic signal collected by the electromagnetic sensor array has an operating frequency range of 10MHz-1GHz.

3. The method for identifying internal cracks in concrete based on DeepSeek according to claim 1, characterized in that: The pretreatment includes: Perform time-frequency domain transformation on multimodal signals to generate a spectrum diagram; The frequency bands whose energy proportion exceeds the preset threshold are extracted by wavelet packet decomposition, and the signal segments are intercepted by a sliding window with a window length of T seconds and a step size of 0.5T.

4. The method for identifying internal cracks in concrete based on DeepSeek according to claim 1, characterized in that: Each expert network in the expert network group includes a cross attention mechanism and a residual convolution layer, and the activation function is a ReLU function.

5. The method for identifying internal cracks in concrete based on DeepSeek according to claim 1, characterized in that: The training strategy for transfer learning fine-tuning includes: The weighted sum of cross entropy loss and Dice coefficient is used as the loss function; The initial learning rate is set to η, and decays to 0.9η after every k batches of training.

6. The method for identifying internal cracks in concrete based on DeepSeek according to claim 5, characterized in that: The weight coefficients α and β of the loss function satisfy α+β=1, wherein the value range of α is 0.6-0.

8.

7. The method for identifying internal cracks in concrete based on DeepSeek according to claim 1, characterized in that: The three-dimensional coordinate information of the crack is generated by the following steps: Generate an attention heat map based on the weight matrix output by the gating network; The center point coordinates of the area exceeding the threshold γ in the heat map are calculated, and the crack depth is determined by combining the output of the residual convolution layer.

8. The method for identifying internal cracks in concrete based on DeepSeek according to claim 7, characterized in that: The threshold γ is dynamically adjusted according to the concrete material properties, and the adjustment range is 0.6-0.

8.

9. A concrete internal crack identification system based on DeepSeek, characterized in that: include: a data acquisition module configured to acquire ultrasonic and electromagnetic sensor signals; a preprocessing module configured to perform the preprocessing step of the method according to any one of claims 1 to 8; The model inference module deploys the fine-tuned DeepSeek-R1 model and outputs the crack parameters.

10. The system according to claim 9, characterized in that The model inference module adopts a distributed architecture: the encoder module is deployed on the edge computing node, and the gated network and expert network group are deployed on the cloud server; the edge computing node and the cloud server transmit data via the 5G communication protocol.

Images