A concrete crack detection method and system combined with a timing neural network

Abstract

The application discloses a concrete crack detection method and system combined with a time sequence neural network, and mainly relates to the technical field of concrete detection. The method comprises the following steps: a detection component is used to perform real-time detection on concrete cracks of a target application scene, and real-time concrete crack detection results are obtained; a target scene environment clue template is called based on the target application scene, and scene environment clue perception data sequences are obtained; a long-short time sequence neural network is used to perform environment interference trend feature interaction analysis on the scene environment clue perception data sequences, and environment interference trend features are obtained; causal perception fusion is performed, and fused concrete crack detection results are obtained. The application has the beneficial effect of solving the technical problem that the existing concrete crack detection ignores potential influencing factors, leading to the inability to conduct in-depth detection analysis on the actual development of concrete cracks, achieving the technical effect of obtaining crack detection results containing environment interference trends and improving the reliability of concrete crack detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of concrete detection, and particularly relates to a concrete crack detection method and system combining a time sequence neural network. BACKGROUND

[0002] The concrete crack detection technology is widely applied to the health monitoring of building structures, and the purpose is to find cracks in time and evaluate the influence of the cracks on the safety of the structures. The traditional concrete crack detection method usually depends on image processing, sensor detection, and manual inspection. However, the detection data under the traditional detection method lacks the continuity of the time dimension, and the interference of environmental factors is ignored, resulting in the lag of the concrete crack detection results.

[0003] The prior art has the technical problem that the concrete crack detection ignores potential influencing factors, and thus cannot deeply analyze the actual development of the concrete cracks. SUMMARY

[0004] The present application provides a concrete crack detection method and system combining a time sequence neural network, which is used to solve the technical problem that the concrete crack detection in the prior art ignores potential influencing factors, and thus cannot deeply analyze the actual development of the concrete cracks.

[0005] In view of the above problems, the present application provides a concrete crack detection method and system combining a time sequence neural network.

[0006] In a first aspect, the present application provides a concrete crack detection method combining a time sequence neural network, and the method comprises the following steps:

[0007] The detection component is used to detect the concrete cracks in real time in a target application scene, and real-time concrete crack detection results are obtained, wherein the concrete crack detection results include crack width-depth-length and crack extension direction; a target scene environment clue template is called based on the target application scene, environment continuous perception data is extracted based on the target scene environment clue template, and scene environment clue perception data sequences are obtained; a long-short time sequence neural network is used to analyze the interaction of the environment interference trend characteristics of the scene environment clue perception data sequences, and environment interference trend characteristics are obtained; and the crack width-depth-length and the crack extension direction are fused based on the environment interference trend characteristics, and fused concrete crack detection results are obtained.

[0008] Optionally, based on the target application scenario, a scene environment clue template is called, and scene environment clue perception data sequence is obtained based on the target scene environment clue template, including: constructing a scene environment clue template library, wherein each scene environment clue template includes a scene feature vector; performing feature extraction on target application scenario information of the target application scenario according to a preset scene index, and constructing a target application scenario feature vector; and performing approximate matching based on the target application scenario feature vector and the scene feature vector of each scene environment clue template in the scene environment clue template library, and taking the scene environment clue template corresponding to the maximum matching similarity as the target scene environment clue template.

[0009] Optionally, the scene environment clue template library is constructed, wherein each scene environment clue template includes a scene feature vector, including: obtaining a sample application scenario information set and a corresponding sample abnormal environment clue set; performing feature extraction on the sample application scenario information set according to a preset scene index, obtaining a sample application scenario feature vector set, and performing same-class aggregation on the sample application scenario feature vector set to obtain a K aggregated sample application scenario feature vector set, wherein K is a positive integer; performing mapping aggregation on the sample abnormal environment clue set according to the K aggregated sample application scenario feature vector set to obtain a K aggregated sample abnormal environment clue set, and performing set union calculation on the K aggregated sample abnormal environment clue set to obtain K integrated sample abnormal environment clues; and performing centralized integration on the K aggregated sample application scenario feature vector set to obtain K scene feature vectors, and using the K scene feature vectors to identify the K integrated sample abnormal environment clues to construct the scene environment clue template library.

[0010] Optionally, the K scene feature vectors are obtained by performing centralized integration on the K aggregated sample application scenario feature vector set, including: extracting a first aggregated sample application scenario feature vector set from the K aggregated sample application scenario feature vector set; performing pairwise combination on the first aggregated sample application scenario feature vector set, respectively calculating combination similarity to obtain a combination similarity set; extracting the combination similarity set with any one of the first aggregated sample application scenario feature vectors in the first aggregated sample application scenario feature vector set, and performing mean value calculation on the extraction result to obtain an integrated similarity set, wherein each integrated similarity corresponds to a first aggregated sample application scenario feature vector; taking the first aggregated sample application scenario feature vector corresponding to the maximum value in the integrated similarity set as a first scene feature vector; respectively performing combination similarity calculation on the K aggregated sample application scenario feature vector set, and integrating according to the calculation result to obtain K scene feature vectors.

[0011] Optionally, the environment interference trend feature is obtained by performing interaction analysis on the long-time sequence environment interference trend feature and the short-time sequence environment interference trend feature, including: calculating a feature interaction similarity set of the long-time sequence environment interference trend feature and the short-time sequence environment interference trend feature by using a cosine similarity formula; constructing an interaction analysis matrix based on the feature interaction similarity set, and enhancing the short-time sequence environment interference trend feature by using the interaction analysis matrix to obtain the environment interference trend feature.

[0012] Optionally, the environment interference trend feature is obtained by performing interaction analysis on the long-time sequence environment interference trend feature and the short-time sequence environment interference trend feature, including: calculating a feature interaction similarity set of the long-time sequence environment interference trend feature and the short-time sequence environment interference trend feature by using a cosine similarity formula; constructing an interaction analysis matrix based on the feature interaction similarity set, and enhancing the short-time sequence environment interference trend feature by using the interaction analysis matrix to obtain the environment interference trend feature.

[0013] Optionally, the system further includes: a causal perception fusion device; and the environment interference trend feature, the crack width-depth-length, and the crack extension direction are analyzed by using the causal perception fusion device to obtain the fusion concrete crack detection result.

[0014] In a second aspect, the application provides a concrete crack detection system combined with a time sequence neural network, including:

[0015] The crack detection result obtaining module is configured to perform real-time concrete crack detection on the target application scene by using the detection component to obtain a real-time concrete crack detection result, wherein the concrete crack detection result includes a crack width-depth-length and a crack extension direction.

[0016] The one or more technical solutions provided in the application have at least the following technical effects or advantages:

[0017] The application utilizes a detection assembly to perform real-time concrete crack detection on a target application scene, and obtains a real-time concrete crack detection result, wherein the concrete crack detection result includes crack width-depth-length and crack extension direction; a target scene environment clue template is called based on the target application scene, environment continuous perception data extraction is performed based on the target scene environment clue template, and a scene environment clue perception data sequence is obtained; environment interference trend characteristics are obtained by utilizing a long-short time sequence neural network to perform environment interference trend characteristic interaction analysis on the scene environment clue perception data sequence; and crack width-depth-length and crack extension direction are subjected to causal perception fusion based on the environment interference trend characteristics, and a fused concrete crack detection result is obtained. The technical effect of obtaining a crack detection result containing an environment interference trend and improving the reliability of concrete crack detection is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0018] FIG. 1 is a flowchart of a concrete crack detection method provided by an embodiment of the application. Figure 1 FIG. 2 is a structural diagram of a concrete crack detection system provided by an embodiment of the application.

[0019] FIG. 3 is a structural diagram of a concrete crack detection system provided by an embodiment of the application. Figure 2 FIG. 4 is a structural diagram of a concrete crack detection system provided by an embodiment of the application.

[0020] Reference signs in the drawings:

[0021] A crack detection result obtaining module 11, a perception data sequence obtaining module 12, an environment interference trend characteristic obtaining module 13, and a fused concrete crack detection result obtaining module 14. DETAILED DESCRIPTION

[0022] The application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not to limit the scope of the application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content taught by the application, and these equivalent forms also fall within the scope defined by the appended claims of the application. It should be noted that the terms “include” and “have” are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] Embodiment one, as shown in FIG. 1, the application provides a concrete crack detection method combined with a time sequence neural network, the method comprises: Figure 1

[0024] ​S1: Use the detection component to perform real-time concrete crack detection on the target application scenario and obtain real-time concrete crack detection results, including crack width-depth-length and crack extension direction.

[0025] In one possible embodiment, the detection component refers to equipment or sensors used for concrete crack detection, including crack width gauges, crack depth gauges, laser scanners, etc. The detection component is used to acquire different characteristic data of concrete cracks. The target application scenario refers to the specific building or structural environment requiring crack detection, such as bridges, floor slabs, tunnels, etc., where the concrete may have cracks, affecting the structural safety. The concrete crack detection results reflect the real-time status of concrete cracks in the target application scenario, including crack width-depth-length and crack extension direction.

[0026] The width is the transverse size of the crack. The depth is the distance from the surface of the crack to its deepest point. The length is the distance the crack extends along the surface. The direction of extension is the direction in which the crack extends along the surface of the structure, usually expressed as vertical, horizontal, or oblique.

[0027] Preferably, the device measures the width of a concrete crack by installing a crack width gauge on either side or around the crack. A crack depth gauge measures the depth of a crack using probes, ultrasonic waves, or laser technology. During measurement, the probe or laser beam of the crack depth gauge is inserted into the crack and precisely measures its depth. A laser scanner scans the concrete surface by emitting a laser beam, using the reflected wave data to accurately construct a three-dimensional image of the crack. Laser scanners can capture the entire length and direction of the crack; this process is typically fully automated, providing the crack's length and its path of propagation on the surface.

[0028] By combining multiple detection devices (crack width gauge, crack depth gauge, laser scanner, etc.) to comprehensively measure the geometric characteristics of concrete cracks, the technical effect of providing basic information for the subsequent integration of crack detection results and environmental interference is achieved.

[0029] S2: Based on the target application scenario, call the target scenario environment clue template, and extract continuous environmental perception data based on the target scenario environment clue template to obtain a scene environment clue perception data sequence;

[0030] Furthermore, based on the target application scenario, a scene environment clue template is invoked, and based on the target scene environment clue template, continuous environmental perception data is extracted to obtain a scene environment clue perception data sequence. Step S2 in this embodiment further includes:

[0031] Construct a scene environment cue template library, where each scene environment cue template includes a scene feature vector;

[0032] feature extraction is performed on the target application scene information of the target application scene according to preset scene indicators, and a target application scene feature vector is constructed;

[0033] Approximate matching is performed between the target application scene feature vector and a scene feature vector of each scene environment clue template in the scene environment clue template library, and a scene environment clue template corresponding to a maximum matching similarity value is taken as a target scene environment clue template.

[0034] In one possible embodiment, the target scene environment clue template is a template of an environmental factor that interferes with a concrete crack in the target application scene, and includes environmental factor features (such as temperature, humidity, load, etc.) that can affect crack development in the scene. Each environment clue template includes a scene feature vector for expressing the performance of these environmental factors in the scene.

[0035] The target scene environment clue template is taken as an object of perception data extraction, and environment continuous perception data extraction is performed on an environment sensor arranged in the target application scene, to obtain a scene environment clue perception data sequence. The scene environment clue perception data sequence reflects the change of environment data associated with the concrete crack in the target application scene over time. By obtaining the scene environment clue perception data sequence, a technical effect of providing data support for subsequent analysis of the environmental interference trend feature of the target application scene is achieved.

[0036] Preferably, a scene environment clue template library is constructed in advance, and each scene environment clue template includes a scene feature vector. The scene feature vector is used to describe the representation of each scene environment clue template in a corresponding type of application scene. The preset scene indicators are indicators used to describe the environment of the application scene, including temperature, humidity, load, etc. The target application scene information is used to describe the environment of the target application scene.

[0037] In one embodiment, the target application scene information is subjected to feature extraction with the preset scene indicators as indexes, and the extracted indicators are filled into an initially empty vector to obtain the target application scene feature vector. The target application scene feature vector reflects the actual environment of the target application scene.

[0038] Then, the cosine similarity formula is used to calculate the similarity between the target application scene feature vector and the scene feature vector of each scene environment clue template in the scene environment clue template library, thus obtaining a matching similarity set. This matching similarity set reflects the degree of similarity between the target application scene feature vector and the scene feature vector corresponding to the template in the scene environment clue template library. The scene environment clue template corresponding to the maximum matching similarity is extracted and used as the target scene environment clue template.

[0039] Furthermore, a scene environment clue template library is constructed, wherein each scene environment clue template includes a scene feature vector. Step S2 in this embodiment of the application further includes:

[0040] Obtain a set of sample application scenario information and a corresponding set of sample abnormal environment clues;

[0041] The sample application scenario information set is subjected to feature extraction according to the preset scenario indicators to obtain a sample application scenario feature vector set. The sample application scenario feature vector set is then aggregated to obtain K aggregated sample application scenario feature vector sets, where K is a positive integer.

[0042] Based on the set of application scenario feature vectors of K aggregated samples, the set of abnormal environment clues of the samples is mapped and aggregated to obtain K aggregated sample abnormal environment clue sets. Then, the union of the K aggregated sample abnormal environment clue sets is obtained to obtain K integrated sample abnormal environment clues.

[0043] The application scenario feature vector sets of K aggregated samples are centrally integrated to obtain K scenario feature vectors. The K scenario feature vectors are then used to identify abnormal environmental clues in the K integrated samples, and a scenario environmental clue template library is constructed.

[0044] In one possible embodiment, the sample application scenario information set refers to the collection of environmental information related to crack detection gathered in multiple different application scenarios. Each sample scenario typically includes parameters such as temperature, humidity, load, and vibration, which affect crack evolution. The sample abnormal environmental clue set refers to abnormal environmental signals or characteristics collected in the sample application scenarios, typically manifested as environmental fluctuations or factors that influence crack propagation under specific conditions, such as drastic temperature fluctuations or load overload.

[0045] Using preset scenario indicators as indexes, features are extracted from the sample application scenario information set to obtain a sample application scenario feature vector set. Then, the sample application scenario feature vector set is aggregated into similar sets, that is, similar sample application scenario feature vectors are grouped into one set to obtain the K aggregated sample application scenario feature vector sets.

[0046] Preferably, a first sample application scenario feature vector is randomly extracted from the sample application scenario feature vector set, the similarity between the sample application scenario feature vector set and the first sample application scenario feature vector is calculated, and the sample application scenario feature vectors in the calculation result that meet the preset similarity threshold are added to the first aggregated sample application scenario feature vector set. Further, the first aggregated sample application scenario feature vector set is removed from the sample application scenario feature vector set, a second sample application scenario feature vector is randomly extracted again, and the same type of aggregation is performed based on the same principle as obtaining the first aggregated sample application scenario feature vector set to obtain a second aggregated sample application scenario feature vector set. In this way, K aggregated sample application scenario feature vector sets are obtained.

[0047] Since the sample application scenario feature vector and the sample abnormal environment clue are one-to-one corresponding, the K aggregated sample application scenario feature vector sets can be used to map and aggregate the sample abnormal environment clue set to obtain corresponding K aggregated sample abnormal environment clue sets. Only one of the repeated abnormal environment clue factors in the K aggregated sample abnormal environment clue sets is retained, that is, the union of the K aggregated sample abnormal environment clue sets is obtained to obtain the K integrated sample abnormal environment clues. The K integrated sample abnormal environment clues are used to describe the abnormal environment factors that are easy to interfere with the concrete cracks in different application scenarios.

[0048] Further, the K aggregated sample application scenario feature vector sets are centrally integrated to obtain feature vectors that can represent each aggregated set, that is, the K scene feature vectors. Then, the K integrated sample abnormal environment clues are identified according to the K scene feature vectors, and the identified K integrated sample abnormal environment clues are respectively taken as scene environment clue templates and are summarized into a scene environment clue template library.

[0049] By constructing the scene environment clue template library, a customized environment analysis template is provided for each type of application scenario, data support is provided for subsequent determination of environment interference factors in different application scenarios, and the technical effect of improving the environment interference analysis efficiency is achieved.

[0050] Further, the K aggregated sample application scenario feature vector sets are centrally integrated to obtain K scene feature vectors, and the embodiment of the application further includes the following steps.

[0051] The first aggregated sample application scenario feature vector set is extracted from the K aggregated sample application scenario feature vector sets.

[0052] apply a set of scene feature vectors on the first aggregated sample application scene to calculate a combination similarity respectively, and obtain a set of combination similarities;

[0053] extract the set of combination similarities with any one of the first aggregated sample application scene feature vectors in the set of first aggregated sample application scene feature vectors respectively, and obtain a set of integrated similarities by averaging the extraction results, wherein each integrated similarity corresponds to a first aggregated sample application scene feature vector;

[0054] take the first aggregated sample application scene feature vector corresponding to the maximum value in the set of integrated similarities as the first scene feature vector;

[0055] calculate the combination similarity of each set of K aggregated sample application scene feature vectors respectively, and obtain K scene feature vectors by integrating the calculation results.

[0056] In an embodiment of the present application, a set of first aggregated sample application scene feature vectors is extracted from the K sets of aggregated sample application scene feature vectors, and then the first aggregated sample application scene feature vectors in the set of first aggregated sample application scene feature vectors are combined two by two to obtain a combination set. Further, the similarity of each combination is calculated using the cosine similarity formula to obtain a set of combination similarities. Each combination similarity reflects the similarity of two first aggregated sample application scene feature vectors in a combination.

[0057] Further, any one of the first aggregated sample application scene feature vectors in the set of first aggregated sample application scene feature vectors is taken as an index to extract the set of combination similarities, and the combination similarity related to the first aggregated sample application scene feature vector is obtained as the extraction result. The extraction results are averaged to obtain the integrated similarity corresponding to the first aggregated sample application scene feature vector. Based on the same principle, the set of first aggregated sample application scene feature vectors is extracted and analyzed to obtain the set of integrated similarities.

[0058] Further, the first aggregated sample application scene feature vector corresponding to the maximum value in the set of integrated similarities is taken as the first scene feature vector, that is, the first scene feature vector is the first aggregated sample application scene feature vector with the deepest average association degree with other first aggregated sample application scene feature vectors. Based on the same principle as obtaining the first scene feature vector, the combination similarity of each set of K aggregated sample application scene feature vectors is calculated, and then integrated to obtain K scene feature vectors.

[0059] S3: perform environment interference trend feature interaction analysis on the scene environment clue perception data sequence using a long-short time sequence neural network to obtain an environment interference trend feature;

[0060] Further, the long-short time sequence neural network is used for environment interference trend feature interaction analysis on the scene environment clue perception data sequence, to obtain the environment interference trend feature. In an embodiment of the present application, step S3 further includes:

[0061] The long time sequence trend feature extraction branch of the long-short time sequence neural network is used for feature extraction on the scene environment clue perception data sequence, to obtain the long time sequence environment interference trend feature.

[0062] The short time sequence trend feature extraction branch of the long-short time sequence neural network is used for feature extraction on the scene environment clue perception data sequence, to obtain the short time sequence environment interference trend feature.

[0063] The long time sequence environment interference trend feature and the short time sequence environment interference trend feature are subjected to interaction analysis, to obtain the environment interference trend feature.

[0064] Further, the long time sequence environment interference trend feature and the short time sequence environment interference trend feature are subjected to interaction analysis, to obtain the environment interference trend feature. In an embodiment of the present application, step S3 further includes:

[0065] The cosine similarity formula is used to calculate a feature interaction similarity set of the long time sequence environment interference trend feature and the short time sequence environment interference trend feature.

[0066] Based on the feature interaction similarity set, an interaction analysis matrix is constructed, and the short time sequence environment interference trend feature is enhanced by using the interaction analysis matrix, to obtain the environment interference trend feature.

[0067] In an embodiment, the long-short time sequence neural network is a neural network architecture specially used for processing and predicting time sequence data, and is capable of capturing long-term dependency. By using the long time sequence trend feature extraction branch for long time sequence data feature extraction, and using the short time sequence trend feature extraction branch for short time sequence (e.g. short time environment change) data feature extraction. The environment interference trend feature is used for comprehensive long-short time sequence analysis, and can reliably reflect the trend of the environment factors that interfere with the concrete cracks in the target application scene.

[0068] Preferably, the long short-term sequence neural network long-term trend feature extraction branch and the short-term trend feature extraction branch are connected in parallel. A plurality of sample scene environment clue perception data sequences, a plurality of sample long-term sequence environment interference trend features, and a plurality of sample short-term sequence environment interference trend features are obtained as a training data set. The plurality of sample scene environment clue perception data sequences and the plurality of sample long-term sequence environment interference trend features in the training data set are extracted to supervise the training of the framework based on the time sequence neural network, until the training converges, and the long-term trend feature extraction branch is obtained after the training is completed. The plurality of sample scene environment clue perception data sequences and the plurality of sample short-term sequence environment interference trend features in the training data set are extracted to supervise the training of the framework based on the time sequence neural network, and the network parameters of the framework are updated according to the output results during the training, until the training converges, and the short-term trend feature extraction branch is obtained after the training is completed.

[0069] The cosine similarity formula is used to calculate the feature interaction similarity set of the long-term sequence environment interference trend feature and the short-term sequence environment interference trend feature. The feature interaction similarity set reflects the feature similarity degree of the long-term sequence environment interference trend feature and the short-term sequence environment interference trend feature. Then, the softmax formula is used to normalize the feature interaction similarity set, and the processed data is filled into an initially empty matrix to obtain an interaction analysis matrix, wherein the interaction analysis matrix is a diagonal matrix.

[0070] The interaction analysis matrix is used to enhance the short-term sequence environment interference trend feature to obtain the environment interference trend feature. Preferably, the interaction analysis matrix is used as a weight matrix, and then the short-term sequence environment interference trend feature is weighted and calculated to obtain the environment interference trend feature. The feature situation under long-term and short-term sequence analysis is comprehensively considered, the reliability of the environment interference trend analysis is improved, and the reliability of the subsequent concrete crack detection result is improved.

[0071] S4: Based on the environment interference trend feature, the crack width-depth-length and the crack extension direction are causally perceived and fused to obtain a fused concrete crack detection result.

[0072] Further, the step S4 of the embodiment of the application further comprises:

[0073] A causal perception fusioner is constructed.

[0074] The environment interference trend feature, the crack width-depth-length, and the crack extension direction are analyzed by using the causal perception fusioner to obtain the fused concrete crack detection result.

[0075] Preferably, the causal perception fusion unit is a functional unit used to perform causal trend fusion of real-time concrete crack detection results based on environmental interference. Multiple environmental interference trend features, crack width-depth-length, and crack extension directions from multiple samples, along with corresponding fused concrete crack detection results, are acquired as training data for the fusion unit. A feedforward neural network architecture is obtained, comprising an input layer, hidden layers, and an output layer. The environmental interference trend features, crack width-depth-length, and fused crack detection results for each sample are used as inputs to the network. Forward propagation is performed through the feedforward neural network to calculate the predicted crack propagation result. The error between the predicted result and the actual label (e.g., true crack width, depth, etc.) is calculated; commonly used loss functions include mean squared error (MSE). The error is calculated using a backpropagation algorithm, and the network parameters (weights and biases) are updated according to the gradient. Optimization algorithms (such as gradient descent or its variants, such as the Adam optimizer) are typically used to update the model weights until training converges, resulting in the trained causal perception fusion unit.

[0076] The trained causal perception fusion engine is used to analyze the environmental disturbance trend characteristics, crack width-depth-length, and crack extension direction to obtain the fused concrete crack detection results. These fused concrete crack detection results are concrete crack detection results that incorporate the environmental disturbance trend. This achieves the technical effect of improving the reliability of crack detection results.

[0077] In summary, the embodiments of this application have at least the following technical effects:

[0078] 1. This application, through real-time data acquisition and the application of temporal neural networks, introduces interference analysis of environmental factors, which can effectively capture the influence of factors such as temperature, humidity, and load on crack development, thereby achieving the technical effect of improving the accuracy of crack detection results.

[0079] 2. This application achieves the technical effect of further improving the accuracy of crack detection results by fusing the geometric features of cracks with environmental disturbance trends to obtain crack detection results that include implicit environmental disturbances.

[0080] Example 2, based on the same inventive concept as the concrete crack detection method combining temporal neural networks in the foregoing examples, as shown in the appendix. Figure 2 As shown, this application provides a concrete crack detection system combining a temporal neural network. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0081] The crack detection result obtaining module 11 is configured to perform real-time concrete crack detection on the target application scene by using the detection assembly, and obtain a real-time concrete crack detection result, wherein the concrete crack detection result includes crack width-depth-length and crack extension direction.

[0082] The perception data sequence obtaining module 12 is configured to call a target scene environment clue template based on the target application scene, extract environment continuous perception data based on the target scene environment clue template, and obtain a scene environment clue perception data sequence.

[0083] The environment interference trend feature obtaining module 13 is configured to perform environment interference trend feature interaction analysis on the scene environment clue perception data sequence by using a long-short time sequence neural network, and obtain an environment interference trend feature.

[0084] The fused concrete crack detection result obtaining module 14 is configured to perform causal perception fusion on crack width-depth-length and crack extension direction based on the environment interference trend feature, and obtain a fused concrete crack detection result.

[0085] Further, the perception data sequence obtaining module 12 is configured to perform the following steps:

[0086] A scene environment clue template library is constructed, wherein each scene environment clue template includes a scene feature vector.

[0087] The target application scene information of the target application scene is feature extracted according to a preset scene index, and a target application scene feature vector is constructed.

[0088] Approximate matching is performed between the target application scene feature vector and the scene feature vector of each scene environment clue template in the scene environment clue template library, and the scene environment clue template corresponding to the maximum matching similarity value is taken as the target scene environment clue template.

[0089] Further, the perception data sequence obtaining module 12 is configured to perform the following steps:

[0090] A sample application scene information set and a corresponding sample abnormal environment clue set are obtained.

[0091] The sample application scene information set is feature extracted according to a preset scene index, a sample application scene feature vector set is obtained, and the sample application scene feature vector set is aggregated in the same category to obtain a K aggregated sample application scene feature vector set, wherein K is a positive integer.

[0092] mapping and aggregating the sample anomaly environment clue set according to the K sets of aggregated sample application scene feature vectors, obtaining K sets of aggregated sample anomaly environment clues, and performing a union set operation on the K sets of aggregated sample anomaly environment clues, to obtain K integrated sample anomaly environment clues;

[0093] integrating the K sets of aggregated sample application scene feature vectors, to obtain K scene feature vectors, and using the K scene feature vectors to identify the K integrated sample anomaly environment clues, to construct a scene environment clue template library.

[0094] Further, the perception data sequence obtaining module 12 is configured to perform the following steps:

[0095] extracting a first set of aggregated sample application scene feature vectors from the K sets of aggregated sample application scene feature vectors;

[0096] pairwise combining the first set of aggregated sample application scene feature vectors, respectively calculating the combination similarity, and obtaining a combination similarity set;

[0097] extracting the combination similarity set with any one of the first set of aggregated sample application scene feature vectors, and performing mean value calculation on the extraction result, to obtain an integrated similarity set, wherein each integrated similarity corresponds to a first aggregated sample application scene feature vector;

[0098] taking the first aggregated sample application scene feature vector corresponding to the maximum value in the integrated similarity set as a first scene feature vector;

[0099] respectively calculating the combination similarity of the K sets of aggregated sample application scene feature vectors, and integrating according to the calculation result, to obtain K scene feature vectors.

[0100] Further, the environment interference trend feature obtaining module 13 is configured to perform the following steps:

[0101] extracting the long-time sequence trend feature extraction branch of the long-short time sequence neural network to perform feature extraction on the scene environment clue perception data sequence, to obtain a long-time sequence environment interference trend feature;

[0102] extracting the short-time sequence trend feature extraction branch of the long-short time sequence neural network to perform feature extraction on the scene environment clue perception data sequence, to obtain a short-time sequence environment interference trend feature;

[0103] performing interactive analysis on the long-time sequence environment interference trend feature and the short-time sequence environment interference trend feature, to obtain the environment interference trend feature.

[0104] Further, the environment interference trend feature acquisition module 13 is configured to perform the following steps:

[0105] Calculate the feature interaction similarity set of the long-time sequence environment interference trend feature and the short-time sequence environment interference trend feature by using the cosine similarity formula.

[0106] Construct an interaction analysis matrix based on the feature interaction similarity set, and enhance the short-time sequence environment interference trend feature by using the interaction analysis matrix to obtain the environment interference trend feature.

[0107] Further, the fused concrete crack detection result acquisition module 14 is configured to perform the following steps:

[0108] Construct a causal perception fusioner.

[0109] Analyze the environment interference trend feature, the crack width-depth-length and the crack extension direction by using the causal perception fusioner to obtain the fused concrete crack detection result.

[0110] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above describes the specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0111] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0112] The present application and the drawings are only exemplary descriptions of the present application, and should be considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalents, the present application intends to include these modifications and changes.

Claims

1. A method for detecting concrete cracks using a temporal neural network, characterized in that, The method includes: The detection component is used to perform real-time concrete crack detection on the target application scenario and obtain real-time concrete crack detection results, which include crack width-depth-length and crack extension direction. Based on the target application scenario, the target scenario environment clue template is invoked, and environmental continuous perception data is extracted based on the target scenario environment clue template to obtain a scene environment clue perception data sequence. The environmental interference trend features are obtained by using a long and short temporal neural network to perform interactive analysis on the scene environmental cue perception data sequence. Based on the aforementioned environmental disturbance trend characteristics, causal perception fusion is performed on crack width-depth-length and crack extension direction to obtain fused concrete crack detection results. Specifically, based on the target application scenario, a scene environment clue template is invoked; based on the target scene environment clue template, continuous environmental perception data extraction is performed to obtain a scene environment clue perception data sequence, including: Construct a scene environment cue template library, where each scene environment cue template includes a scene feature vector; Based on preset scenario indicators, feature extraction is performed on the target application scenario information of the target application scenario to construct a target application scenario feature vector; Approximate matching is performed between the target application scenario feature vector and the scenario feature vector of each scenario environment clue template in the scenario environment clue template library, and the scenario environment clue template corresponding to the maximum matching similarity is taken as the target scenario environment clue template.

2. The concrete crack detection method combining a temporal neural network as described in claim 1, characterized in that, Construct a scene environment cue template library, where each scene environment cue template includes a scene feature vector, including: Obtain a set of sample application scenario information and a corresponding set of sample abnormal environment clues; The sample application scenario information set is subjected to feature extraction according to the preset scenario indicators to obtain a sample application scenario feature vector set. The sample application scenario feature vector set is then aggregated to obtain K aggregated sample application scenario feature vector sets, where K is a positive integer. Based on the set of application scenario feature vectors of K aggregated samples, the set of abnormal environment clues of the samples is mapped and aggregated to obtain K aggregated sample abnormal environment clue sets. Then, the union of the K aggregated sample abnormal environment clue sets is obtained to obtain K integrated sample abnormal environment clues. The application scenario feature vector sets of K aggregated samples are centrally integrated to obtain K scenario feature vectors. The K scenario feature vectors are then used to identify abnormal environmental clues in the K integrated samples, and a scenario environmental clue template library is constructed.

3. The concrete crack detection method combining a temporal neural network as described in claim 2, characterized in that, The application scenario feature vector sets of K aggregated samples are centrally integrated to obtain K scenario feature vectors, including: Extract the first aggregated sample application scenario feature vector set from the K aggregated sample application scenario feature vector sets; The first aggregated sample application scenario feature vector set is combined pairwise, and the combination similarity is calculated for each pairwise combination to obtain a combination similarity set. The combined similarity set is extracted by taking any one of the first aggregated sample application scenario feature vectors from the first aggregated sample application scenario feature vector set, and the mean of the extraction results is calculated to obtain the integrated similarity set, where each integrated similarity corresponds to a first aggregated sample application scenario feature vector; The first aggregated sample application scenario feature vector corresponding to the maximum value in the integrated similarity set is used as the first scenario feature vector; The similarity of the application scenario feature vector sets of K aggregated samples is calculated separately, and the results are integrated to obtain K scenario feature vectors.

4. The concrete crack detection method combining a temporal neural network as described in claim 1, characterized in that, Using long and short-term neural networks, environmental interference trend features are interactively analyzed on the scene environmental cue perception data sequence to obtain environmental interference trend features, including: The long-term trend feature extraction branch of the long-short-term neural network is used to extract features from the scene environment cue perception data sequence to obtain long-term environmental interference trend features. The short-term trend feature extraction branch of the long-short-term neural network is used to extract features from the scene environment cue perception data sequence to obtain short-term environmental interference trend features. Interactive analysis is performed on the long-term and short-term environmental interference trend characteristics to obtain the environmental interference trend characteristics.

5. The concrete crack detection method combining a temporal neural network as described in claim 4, characterized in that, Interactive analysis is performed on the long-term and short-term environmental interference trend characteristics to obtain the environmental interference trend characteristics, including: The cosine similarity formula is used to calculate the feature interaction similarity set of long-term and short-term environmental disturbance trend features; An interaction analysis matrix is ​​constructed based on the feature interaction similarity set, and the interaction analysis matrix is ​​used to enhance the short-term environmental interference trend features to obtain the environmental interference trend features.

6. The concrete crack detection method combining a temporal neural network as described in claim 1, characterized in that, include: Construct a causal perception fusion engine; The environmental disturbance trend characteristics, crack width-depth-length, and crack extension direction are analyzed using a causal perception fusion device to obtain the crack detection results of the fused concrete.

7. A concrete crack detection system incorporating a temporal neural network, characterized in that, The system is used to implement the concrete crack detection method combining a temporal neural network as described in any one of claims 1-6, the system comprising: The crack detection result acquisition module is used to perform real-time concrete crack detection on the target application scenario using the detection component and obtain real-time concrete crack detection results, which include crack width-depth-length and crack extension direction. The perception data sequence acquisition module is used to call the target scene environment clue template based on the target application scenario, and perform continuous environmental perception data extraction based on the target scene environment clue template to obtain the scene environment clue perception data sequence. The environmental interference trend feature acquisition module is used to perform interactive analysis of the environmental interference trend features of the scene environmental clue perception data sequence using a long and short time series neural network to acquire environmental interference trend features. The module for obtaining the results of the fused concrete crack detection is used to perform causal perception fusion of crack width-depth-length and crack extension direction based on the environmental disturbance trend characteristics, so as to obtain the results of the fused concrete crack detection.

Images