Improved evidence theory multi-modal tunnel anomaly identification method

Abstract

The application discloses an improved evidence theory multi-modal tunnel anomaly identification method, and belongs to the technical field of underground engineering safety monitoring; different modal data are collected in the robot inspection process, and the space-time synchronous registration of the modal data is realized through unified time stamp and external parameter calibration; laser radar three-dimensional structure features, visible light texture edge features and infrared temperature field features are respectively extracted, and basic probability distribution functions of the modes are constructed; a modal confidence coefficient and an environment adjustment coefficient are introduced to double-dynamically correct the data, a conflict adaptive attenuation mechanism is used to improve D-S evidence fusion rules, and the modal data fusion is realized; finally, the anomaly identification result and the early warning information are output based on the fusion trust function. The problems of low tunnel anomaly identification precision, weak environment anti-interference ability and high conflict information fusion loss in the prior art are effectively solved, the precision and stability of the tunnel anomaly identification in the complex underground environment are greatly improved, and the method has a good engineering application prospect.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering safety monitoring technology, specifically involving an improved evidence theory-based multimodal roadway anomaly identification method, which can be widely applied to roadway structural safety and environmental anomaly monitoring in fields such as mining, underground tunnel engineering, and urban underground space development. Background Technology

[0002] Underground tunnels are core structures in mining, transportation tunnels, and underground utility tunnels. Their structural safety and environmental stability directly determine the operational safety of the project and the safety of personnel. During long-term service, underground tunnels are susceptible to various safety hazards due to multiple factors such as stress redistribution, groundwater seepage, weathering and deterioration of surrounding rock, and construction disturbances. These hazards include the expansion of surrounding rock cracks, localized structural spalling, water leakage, and thermal anomalies. If these hazards are not identified and addressed in a timely manner, they can easily lead to major safety accidents such as collapses and water inrushes.

[0003] Currently, anomaly detection in tunnels mainly relies on two methods: manual inspection or single-sensor detection. Manual inspection suffers from inherent drawbacks such as low efficiency, high labor intensity, high risk in hazardous environments, and significant susceptibility to subjective experience, making it unsuitable for long-distance, routine tunnel inspections. Single-sensor detection has significant limitations: visible light image detection is easily affected by environmental factors such as dust, sudden changes in lighting, and shadows within the tunnel, resulting in a substantial decrease in accuracy in low-visibility environments; infrared thermal imaging can only detect temperature anomalies and cannot accurately identify geometric anomalies such as structural cracks and spalling in the surrounding rock; while lidar can acquire high-precision three-dimensional structural data of the tunnel, enabling accurate detection of structural anomalies, it cannot detect abnormal temperature environments.

[0004] In recent years, multi-sensor fusion technology has been gradually introduced into the field of underground environmental monitoring to improve anomaly detection capabilities through the complementarity of multi-source information. However, existing multi-sensor fusion methods mostly adopt shallow fusion methods such as simple weighted averaging and fixed weight superposition, lacking a dynamic evaluation mechanism for the reliability of different modal detections. They also fail to consider the differentiated impact of environmental interferences such as dust, vibration, and lighting changes in the tunnel on the detection performance of each modality, and cannot achieve adaptive dynamic adjustment of fusion weights. At the same time, in complex scenarios where there are significant conflicts between multi-source information, traditional DS evidence theory fusion methods are prone to conflict amplification, leading to serious distortion of fusion results, and even misjudgment or missed judgment, making it difficult to meet the high-precision and high-robustness anomaly identification requirements in complex underground tunnel environments.

[0005] Therefore, developing a multimodal roadway anomaly identification method with dynamic evaluation of modal credibility, adaptive adjustment of environmental interference, and robust fusion of highly conflicting evidence has become an urgent technical problem to be solved in this field. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides an improved evidence-based multimodal roadway anomaly identification method, which can effectively solve the technical problems of low roadway anomaly identification accuracy, weak environmental anti-interference ability, easy distortion of high-conflict multi-source information fusion, and poor security of manual inspection in the prior art, and realize high-precision, high-robustness, and automated identification of roadway anomalies in complex underground environments.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: an improved evidence-based multimodal tunnel anomaly identification method, applied to a tracked mobile robot platform equipped with an infrared thermal imaging camera, a visible light camera, and a lidar sensor, comprising the following steps: S1. Synchronous Acquisition Control: Controls different sensors to synchronously start the acquisition of roadway inspection data, and assigns a unified timestamp to each acquired modal data to complete the time synchronization of each modal data.

[0008] S2. Spatial registration processing: Calibrate and obtain the extrinsic parameter matrix from the lidar to the camera coordinate system and the extrinsic parameter matrix from the infrared thermal imaging camera to the visible light camera coordinate system. Transform the lidar point cloud coordinates to a unified coordinate system. Combine the camera intrinsic parameter matrix to complete the projection of 3D points onto the image plane and realize the spatial registration of various modal data.

[0009] S3. Multimodal feature extraction: Feature extraction is performed on each modal data after registration to obtain the three-dimensional structural features of the lidar point cloud, the texture and edge features of the visible light image, and the temperature field distribution features of the infrared thermal imaging data, and feature vectors of each modality are constructed respectively.

[0010] S4. Construction of basic probability allocation function: Input the feature vectors of each modality into the corresponding pre-trained classification model to obtain the predicted probability of the preset abnormal category set, and construct the basic probability allocation function of each modality based on the predicted probability.

[0011] S5. Modal confidence correction: Calculate the confidence coefficient of each modality based on its historical recognition accuracy. Then, correct the basic probability allocation function of each modality using the confidence coefficient to obtain the basic probability allocation function after confidence correction.

[0012] S6. Environmental Adaptive Correction: Calculate the environmental disturbance index, generate an environmental adjustment factor based on the environmental disturbance index, and perform a second correction on the basic probability allocation function after the first correction through the environmental adjustment factor to obtain the basic probability allocation function after environmental adaptive correction.

[0013] S7. Improved DS evidence fusion: Calculate the conflict factor between each modality data, introduce a conflict attenuation coefficient to adaptively attenuate the conflict factor to obtain an improved conflict factor, optimize the DS evidence fusion rule based on the improved conflict factor, and fuse the basic probability allocation function of each modality after secondary correction through the optimized DS evidence fusion rule to obtain the fused global basic probability allocation function.

[0014] S8. Anomaly Identification and Early Warning: Based on the fused global basic probability allocation function, calculate the trust function of each anomaly category, select the anomaly category with the highest trust as the final anomaly identification result, and output the corresponding early warning information.

[0015] Furthermore, in step S1, the formula for calculating the unified timestamp is: Where t0 is the system initial time, Δt is the sampling period, and k is the sampling sequence number. Through a dual mechanism of unified hardware triggering and timestamp assignment, precise synchronization of data from each modality in the time dimension is ensured, eliminating fusion errors caused by time differences in multi-sensor sampling.

[0016] Further, in step S3, the extraction process of the three-dimensional structural features is as follows: a local neighborhood covariance matrix is ​​constructed for the registered lidar point cloud data; eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​λ0, λ1, and λ2; the point cloud curvature is calculated based on the eigenvalues; the mean curvature, variance of curvature, and plane residual are extracted to construct the lidar three-dimensional structural feature vector; the extraction process of the texture and edge features is as follows: the visible light image is preprocessed by grayscale conversion and Gaussian filtering; edge information is extracted using the Canny operator; edge density is calculated; and the texture features extracted by the grayscale co-occurrence matrix are combined to construct the visible light feature vector; the extraction process of the temperature field distribution features is as follows: the mean temperature, variance of temperature, and amplitude of temperature gradient are calculated for the infrared thermal imaging data; the area ratio of abnormal hot spots is extracted; and the infrared temperature field feature vector is constructed.

[0017] Furthermore, in step S4, the formula for constructing the basic probability allocation function is: Where Θ={A1,A2,…,A n} represents a pre-defined set of anomaly categories, including at least four types of anomalies: surrounding rock cracks, structural spalling, roadway water accumulation, and seepage heat anomalies, while also including normal operating condition categories; m i (A j Let ) represent the anomaly category A of the i-th modality pair. j The basic probability assignment value, pi(A) j Let ) represent the anomaly category A of the i-th modality pair.j The predicted probability, m i (Θ) represents the uncertainty of the i-th mode.

[0018] Furthermore, in step S5, the formula for calculating the credibility coefficient is: Among them, Acc i Let be the historical recognition accuracy of the i-th modality, and max(Acc) be the maximum historical recognition accuracy of all modalities; the basic probability assignment function after one correction. for: .

[0019] By using the reliability coefficients of each modality, the data of low-reliability modalities are weighted down to reduce the interference of modalities with high misjudgment probability on the fusion results.

[0020] Furthermore, the formula for calculating the environmental disturbance index is as follows: Where D is the dust concentration in the tunnel, V is the vibration intensity during the inspection process, L is the change in illumination, and w1, w2, and w3 are preset weighting coefficients, satisfying w1 + w2 + w3 = 1; the calculation formula for the environmental adjustment factor is: Where γ is the preset environmental impact coefficient; the basic probability allocation function after secondary correction is: By using environmental adjustment factors, dynamic adaptive correction of the data is achieved. The stronger the environmental interference, the lower the data weight of the corresponding modality, thus improving the environmental anti-interference capability of the fusion method.

[0021] Furthermore, in step S7, the formula for calculating the conflict factor between each modality of data is: in, , Let B be the basic probability assignment function after secondary correction for two different modalities; B and C are subsets of the two modalities in the preset abnormal category set Θ.

[0022] The formula for calculating the conflict attenuation coefficient is: Where δ is the preset conflict adjustment coefficient, with a value range of [1,5]; K is the conflict factor; the improved conflict factor is: By adaptively suppressing high-conflict factors through a conflict attenuation coefficient, the fusion distortion problem of traditional DS evidence theory in high-conflict scenarios is solved.

[0023] Furthermore, in step S7, the optimized DS evidence fusion rule is specifically formulated as follows: Where Θ is a preset set of anomaly categories, that is, the set of all possible anomaly categories; A is a certain anomaly category in the preset set of anomaly categories Θ; B, C, and E are the subsets corresponding to the three modalities in the preset set of anomaly categories Θ, respectively representing the set of propositions supported by the evidence of the three different modalities; These are the basic probability assignment functions after quadratic correction for the three modes, respectively; K represents the basic probability assignment value for anomaly category A after trimodal fusion. 123 This represents the conflict factor between the fusion results of the first two modalities and the data of the third modality. This is a preset conflict adjustment coefficient used to control the attenuation intensity of conflict evidence.

[0024] Furthermore, in step S8, the formula for calculating the trust function is: in, The trust level for anomaly category A; The global basic probability assignment function after the fusion of subset B; The formula for selecting the final anomaly detection result is: .

[0025] in, This determines the final anomaly category. Based on the risk level of the anomaly category, corresponding warning information is output, including the anomaly location, anomaly type, risk level, and handling recommendations.

[0026] Compared with the prior art, the present invention has the following advantages: 1. This invention achieves full-dimensional coverage of roadway structural anomalies, texture anomalies, and temperature anomalies through the fusion detection of three-modal sensors: infrared thermal imaging camera, visible light camera, and lidar. It fully leverages the technical advantages of each sensor, complements the detection blind spots of a single sensor, and solves the problem that existing technologies cannot simultaneously identify structural and environmental anomalies. Furthermore, it achieves time synchronization of data from each modality through a unified hardware trigger signal and a unified timestamp mechanism, and achieves spatial registration of the three-modal data through multi-coordinate system extrinsic parameter calibration and coordinate transformation. This fundamentally eliminates the fusion error caused by the spatiotemporal mismatch of multi-source data, laying the foundation for high-precision multimodal fusion.

[0027] 2. This invention innovatively introduces modal credibility coefficient and environmental adjustment factor to perform dual dynamic correction on each modal data. This not only achieves the weight reduction processing of low reliability modal evidence, but also achieves adaptive adjustment to environmental interference, which greatly improves the anti-interference ability and robustness of the fusion method in complex roadway environments.

[0028] 3. This invention improves the traditional DS evidence fusion rules through a conflict adaptive attenuation mechanism. By adaptively suppressing high-conflict data through an improved conflict factor, it solves the technical pain points of distortion and high misjudgment rate in the fusion results of traditional DS evidence theory in high-conflict scenarios, and significantly improves the fusion accuracy of conflict information of various modalities.

[0029] 4. This invention achieves automated inspection based on a tracked mobile robot platform, replacing the traditional manual inspection method, which greatly improves the efficiency of tunnel inspection, avoids the risks of manual operation in high-risk environments, has good engineering applicability and promotion value, and can be widely used in safety monitoring scenarios of various underground tunnel projects. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the process of the present invention.

[0031] Figure 2 This is a structural design diagram of the tracked mobile robot platform in this invention. Detailed Implementation

[0032] The present invention will be further described below.

[0033] This embodiment is applied to a safety inspection scenario in underground tunnels, and the platform is a mobile robot. The mobile robot integrates an infrared thermal imaging sensor, a visible light camera, and a lidar sensor, such as... Figure 2 As shown, it is used to acquire infrared temperature field information, visible light image information, and three-dimensional structural information of the roadway environment. When applied to safety inspection scenarios in underground coal mine roadways, the mobile robot is preferably an explosion-proof tracked mobile robot, which should integrate an explosion-proof infrared thermal imaging camera, a visible light explosion-proof camera, and a multi-line lidar sensor, while also carrying a dust concentration sensor, a light sensor, etc., to acquire environmental parameters. A preset set of abnormality categories Θ={A1,A2,A3,A4,A5} is defined, where A1 is normal operating conditions, A2 is surrounding rock cracks, A3 is structural spalling, A4 is roadway water accumulation, and A5 is seepage thermal anomalies.

[0034] like Figure 1 As shown, the specific implementation steps are as follows: S1. Synchronous Acquisition Control: The FPGA module of the robot's main control unit sends a unified hardware trigger signal to the infrared thermal imaging camera, visible light camera, and lidar sensor to control the three sensors to synchronously start data acquisition. Simultaneously, a unified timestamp is assigned to each frame of acquired three-modal data. The timestamp calculation formula is as follows: In this embodiment, the initial system time t0 is the inspection start time, the sampling period Δt=0.1s, and k is the sampling sequence number, realizing microsecond-level time synchronization of three-modal data.

[0035] S2. Spatial Registration Processing: Before inspection, all sensors are calibrated. The intrinsic parameter matrices of the visible light camera and infrared camera are obtained using Zhang's calibration method, and the extrinsic parameter matrix T of the lidar to the visible light camera coordinate system is obtained using the hand-eye calibration method. L→C And the extrinsic parameter matrix T from the infrared thermal imaging camera to the visible light camera coordinate system. I→C The original point cloud coordinates P of the lidar were... L Through the formula: The coordinates are transformed to the unified coordinate system of the visible light camera, and the pixel-level projection of the 3D point cloud onto the visible light image plane is completed by combining the camera intrinsic parameter matrix. This achieves spatial registration of the three-modal data and ensures that the structure, texture and temperature information at the same spatial location correspond one-to-one.

[0036] S3. Multimodal Feature Extraction: For the registered LiDAR point cloud data, construct the k-neighborhood (k=20 in this embodiment) covariance matrix for each point. The specific formula is as follows: In the formula, C is the covariance matrix of the point cloud neighborhood; Let be the three-dimensional coordinates of the i-th point in the neighborhood; The centroid coordinates of the neighboring point cloud; This represents the number of neighboring points.

[0037] Eigenvalue decomposition of the covariance matrix yields λ0, λ1, and λ2; point cloud curvature is then calculated. We extract three core features—mean curvature, variance curvature, and plane residual—within the neighborhood to construct a three-dimensional structural feature vector for the lidar.

[0038] For the registered visible light image, grayscale conversion and 5×5 Gaussian filtering preprocessing are performed to eliminate image noise. The Canny operator is then used to extract image edge information and calculate edge density. , where N edge N represents the number of edge pixels. total The total number of pixels in the image is denoted as 1. At the same time, four texture features—contrast, energy, entropy, and correlation—are extracted through the gray-level co-occurrence matrix, and a visible light feature vector is constructed by combining it with the edge density.

[0039] Calculate the mean temperature T and temperature variance within the monitored area using the registered infrared thermal imaging data. Temperature gradient amplitude: Simultaneously, the proportion of abnormal hot spots exceeding the normal temperature threshold is extracted to construct an infrared temperature field feature vector.

[0040] S4. Construction of the basic probability assignment function: Input the feature vectors of the three modalities into the pre-trained SVM classification model, where the lidar features correspond to the structural anomaly classification model, the visible light features correspond to the texture anomaly classification model, and the infrared features correspond to the temperature anomaly classification model. Output the predicted probability p of each modality for the five preset categories. i (A j Based on the predicted probabilities, a basic probability assignment function for each mode is constructed, and the formula is as follows: , , where i=1,2,3 correspond to the three modes of lidar point cloud data, visible light image, and infrared thermal imaging data, respectively.

[0041] S5. Modal Confidence Correction: Based on the historical recognition accuracy of each mode in the downhole test dataset, in this embodiment, the historical recognition accuracy of the lidar is Acc1=92.3%, visible light Acc2=85.7%, and infrared Acc3=89.1%. The confidence coefficient of each mode is calculated as follows: We obtain α1=1, α2=0.928, and α3=0.965; We then perform a first correction on the basic probability assignment function for each mode using the confidence coefficient. And the uncertainty is corrected synchronously according to the modified basic probability allocation function. .

[0042] S6. Environmental Adaptive Correction: The environmental sensors on board acquire real-time data on the dust concentration D in the tunnel, the robot vibration intensity V, and the change in illumination L. In this embodiment, the preset weights are w1=0.5, w2=0.3, and w3=0.2, and the environmental interference index E is calculated. i =0.5D+0.3V+0.2L; Set the environmental impact coefficient γ=0.8 and calculate the environmental adjustment factor. The basic probability allocation function, after a first correction, is further modified using environmental adjustment factors to obtain... And the uncertainty is corrected synchronously according to the modified basic probability allocation function. .

[0043] S7. Improved DS evidence fusion: First, calculate the conflict factor between the two modalities of visible light image and infrared thermal imaging data. Calculate the collision attenuation coefficient: δ is the preset conflict adjustment coefficient, and K is the conflict factor; thus, the improved conflict factor is obtained. .

[0044] The first round of fusion of the two modes is performed using the following formula: The intermediate fusion results are obtained: Next, the conflict factor K between the intermediate fusion result and the lidar data is calculated. 123 And the corresponding improved conflict factor is obtained by adjusting the conflict attenuation coefficient: .

[0045] The second round of fusion will be conducted using the following formula: Will After expansion, the improved global basic probability assignment function for DS fusion is obtained: .

[0046] S8. Anomaly Identification and Early Warning: Based on the fused global basic probability allocation function, calculate the trust function for each anomaly category. The category with the highest trust level was selected as the final identification result. Based on the identification results, if it is an anomaly, the system will output an early warning message containing the anomaly location, anomaly type, risk level, and handling suggestions, and simultaneously upload it to the ground monitoring platform, in conjunction with the robot's positioning information.

[0047] This embodiment also provides an improved evidence theory multimodal roadway anomaly identification device for implementing the above method, comprising: a tracked mobile robot platform, serving as a carrier for inspection operations and hardware mounting, possessing adaptive walking, autonomous navigation, and obstacle avoidance capabilities on complex road surfaces; a multi-sensor acquisition module, including an infrared thermal imaging camera, a visible light camera, and a lidar sensor integrated on the tracked mobile robot platform, for acquiring three modal raw data of roadway inspection; a synchronous acquisition control unit, for sending a unified hardware trigger signal to the multi-sensor acquisition module to control the three sensors to synchronously start acquisition, and assigning a unified timestamp to each acquired modal data to complete the time synchronization of each modal data; a spatial registration unit, for realizing the spatial registration of the three modal data; a feature extraction unit, for performing feature extraction on the registered three modal data respectively, obtaining the three-dimensional structural features of the lidar point cloud, the texture and edge features of the visible light image, and the temperature field distribution features of the infrared thermal imaging data, and constructing the feature vectors corresponding to each modality respectively; and a basic probability allocation construction unit, for inputting the feature vectors of each modality into the corresponding... The system employs a pre-trained classification model to obtain predicted probabilities for a preset set of anomaly categories, and constructs basic probability allocation functions for each modality based on these predicted probabilities. A primary correction unit calculates modality confidence coefficients based on the historical recognition accuracy of each modality, and uses these coefficients to perform a primary correction on the basic probability allocation functions for each modality. A secondary correction unit constructs an environmental impact model to calculate environmental interference indicators, generates environmental adjustment factors based on these indicators, and uses these adjustment factors to perform a secondary correction on the primary-corrected basic probability allocation functions. An improved evidence fusion unit calculates conflict factors between data from different modalities, introduces a conflict attenuation coefficient to adaptively attenuate these factors to obtain improved conflict factors, optimizes the DS evidence fusion rules based on these improved conflict factors, and fuses the secondary-corrected three-modality basic probability allocation functions to obtain a fused global basic probability allocation function. An anomaly identification and early warning unit calculates the trust function for each anomaly category based on the fused global basic probability allocation function, selects the anomaly category with the highest trust as the final anomaly identification result, and outputs the corresponding early warning information. All of these units are integrated into the main control industrial computer and ground monitoring host of the tracked mobile robot platform.

[0048] Through field testing in this embodiment, in the complex environment of high dust, low light, and strong vibration in underground coal mines, the comprehensive identification accuracy of various anomalies in the roadway by this invention reaches 96.8%, which is more than 15% higher than the single sensor detection method, more than 10% higher than the traditional fixed weight fusion method, and 85% lower than the traditional DS evidence fusion method in high-conflict scenarios, demonstrating excellent detection accuracy and environmental robustness.

[0049] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An improved evidence-based multimodal roadway anomaly identification method, characterized in that, Includes the following steps: S1. During the inspection process, data from each modality are collected synchronously, and a unified timestamp is assigned to each modality data. S2. Determine the transformation relationship between each modal data to achieve spatial registration of each modal data; S3. Extract features from each registered modal data and construct feature vectors for each modality. S4. Based on the feature vectors of each modality, obtain the predicted probability of the preset abnormal category set, and construct the basic probability allocation function of each modality based on the predicted probability; S5. Calculate the confidence coefficient of each modality based on its historical recognition accuracy, and use it to make a correction to the basic probability allocation function of each modality. S6. Calculate the environmental disturbance index and generate the environmental adjustment factor. Then, use the environmental adjustment factor to perform a second correction on the basic probability allocation function after the first correction. S7. Introduce a conflict attenuation coefficient to adaptively attenuate the conflict factors between each modality data to obtain an improved conflict factor. Optimize the DS evidence fusion rule based on the improved conflict factor, and fuse the basic probability allocation functions of each modality after the second correction through this rule to obtain the fused global basic probability allocation function. S8. Calculate the trust function of each anomaly category based on the fused global basic probability allocation function, select the anomaly category with the highest trust as the final anomaly identification result, and output the warning information.

2. The improved evidence-based multimodal roadway anomaly identification method according to claim 1, characterized in that, In step S1, the formula for calculating the unified timestamp is: Where t0 is the initial system time, Δt is the sampling period, and k is the sampling sequence number.

3. The improved evidence-based multimodal roadway anomaly identification method according to claim 1, characterized in that, The modal data in step S1 include infrared thermal imaging data collected by an infrared thermal imaging camera, visible light data collected by a visible light camera, and lidar point cloud data collected by a lidar sensor.

4. The improved evidence-based multimodal roadway anomaly identification method according to claim 3, characterized in that, In step S3, feature extraction specifically involves: 3D structural feature extraction: Construct a local neighborhood covariance matrix for the registered lidar point cloud data, perform eigenvalue decomposition on the matrix to obtain eigenvalues, calculate the point cloud curvature based on the eigenvalues, extract the mean curvature, variance of curvature and plane residual, and construct the lidar 3D structural feature vector. Texture and edge feature extraction: Visible light data is preprocessed, and edge information is extracted using the Canny operator. Edge density is calculated, and visible light feature vectors are constructed by combining texture features. Temperature field distribution feature extraction: Calculate the mean temperature, temperature variance, and temperature gradient amplitude from the infrared thermal imaging data, extract the area ratio of abnormal hot spots, and construct an infrared temperature field feature vector.

5. The improved evidence-based multimodal roadway anomaly identification method according to claim 3, characterized in that, In step S4, the formula for constructing the basic probability allocation function is: Where Θ={A1,A2,…,A n } represents the preset set of exception categories, m i (A j Let ) represent the anomaly category A of the i-th modality pair. j The basic probability assignment value, pi(A) j Let ) represent the anomaly category A of the i-th modality pair. j The predicted probability, m i (Θ) represents the uncertainty of the i-th mode.

6. The improved evidence-based multimodal roadway anomaly identification method according to claim 5, characterized in that, In step S5, the formula for calculating the credibility coefficient is: Among them, Acc i Let be the historical recognition accuracy of the i-th modality, and max(Acc) be the maximum value of the historical recognition accuracy of all modalities. Basic probability assignment function after one correction for: 。 7. The improved evidence-based multimodal roadway anomaly identification method according to claim 6, characterized in that, In step S6, the formula for calculating the environmental disturbance index is: Where D is the dust concentration in the tunnel, V is the vibration intensity during the inspection process, L is the change in illumination, and w1, w2, and w3 are preset weighting coefficients. The formula for calculating environmental regulation factors is: Wherein, γ is the preset environmental impact coefficient; The basic probability assignment function after the second correction is: 。 8. The improved evidence-based multimodal roadway anomaly identification method according to claim 7, characterized in that, In step S7, the formula for calculating the conflict factor between modal data is as follows: in, , Let B and C be the basic probability assignment functions after secondary correction for two different modalities, where B and C are subsets of the two modalities in the preset set of anomaly categories Θ. The formula for calculating the conflict attenuation coefficient is: Where δ is the preset conflict adjustment coefficient; K is the conflict factor; The improved conflict factor is then obtained as follows: 。 9. The improved evidence-based multimodal roadway anomaly identification method according to claim 8, characterized in that, In step S7, the optimized DS evidence fusion rule is specifically formulated as follows: Where Θ is a preset set of anomaly categories; A is a certain anomaly category in the preset set of anomaly categories Θ; B, C, and E are subsets corresponding to the three modalities in the preset set of anomaly categories Θ, respectively; These are the basic probability assignment functions after quadratic correction for the three modes, respectively; K represents the basic probability assignment value for anomaly category A after trimodal fusion. 123 This represents the conflict factor between the fusion results of the first two modalities and the data of the third modality. This is the preset conflict adjustment coefficient.

10. The improved evidence-based multimodal roadway anomaly identification method according to claim 1, characterized in that, In step S8, the formula for calculating the trust function is: in, The trust level for anomaly category A; The global basic probability assignment function after the fusion of subset B; The formula for selecting the final anomaly detection result is: in, This is the final identified anomaly category.

Images