Intelligent concrete detection method and system based on machine learning

Abstract

The present invention provides a method and system for intelligent detection of concrete based on machine learning. The first sample concrete sample image data sequence is used to perform parameter learning on each network module unit of the basic concrete intelligent detection network one by one, and a temporary concrete intelligent detection network is generated, which significantly improves the learning efficiency and accuracy of concrete characteristics. Subsequently, through module interaction processing, the key module units in the prior machine learning network and the temporary concrete intelligent detection network are integrated to form multiple fusion network module units. Finally, the second sample concrete sample image data sequence is used to perform migration parameter learning on the second interactive network module unit in the fusion network module unit until the training convergence requirements are met, and a target concrete intelligent detection network that completes module parameter learning is generated, thereby improving the intelligence level of concrete detection and significantly enhancing the generalization ability and detection accuracy of the concrete intelligent detection network.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method and system for intelligent concrete detection based on machine learning. Background Art

[0002] In the field of concrete quality testing, traditional testing methods often rely on manual experience, physical tests or chemical analysis. Although these methods have certain accuracy, they have disadvantages such as long testing cycle, high cost and great influence of human factors. With the rapid development of computer technology and machine learning, intelligent concrete testing methods based on machine learning have gradually emerged, providing a new way for the rapid and accurate assessment of concrete quality.

[0003] However, existing machine learning-based intelligent concrete detection methods often use a single machine learning model, which makes it difficult to fully utilize prior knowledge and domain experience, resulting in limited generalization ability and detection accuracy of the model. In order to overcome the above shortcomings, a feasible idea is to organically integrate the prior machine learning network and the basic concrete intelligent detection network, and make full use of the advantages of both in a modular way. However, how to achieve effective decomposition and reconstruction of network modules, and how to establish effective mapping connections and parameter learning mechanisms between modules are technical problems that need to be solved urgently. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for intelligent detection of concrete based on machine learning, the method comprising:

[0005] Decomposing the priori machine learning network and the basic concrete intelligent detection network into a plurality of network module units respectively; constructing mapping connections respectively between each network module unit in the priori machine learning network and the basic concrete intelligent detection network;

[0006] Acquire a first sample concrete sample image data sequence, traverse each network module unit of the basic concrete intelligent detection network, and perform module parameter learning on the traversed network module units one by one based on the first sample concrete sample image data sequence, the target network module units in the priori machine learning network and corresponding to the traversed network module units, until the module parameter learning of all network module units of the basic concrete intelligent detection network is completed, thereby generating a temporary concrete intelligent detection network;

[0007] The a priori machine learning network and the temporary concrete intelligent detection network are subjected to module interaction processing to generate a plurality of fusion network module units; the fusion network module unit comprises a plurality of first interaction network module units extracted from the a priori machine learning network, and a plurality of second interaction network module units corresponding to the plurality of first interaction network module units extracted from the temporary concrete intelligent detection network;

[0008] Acquire a second sample concrete sample image data sequence, traverse each of the multiple fusion network module units, and based on the second sample concrete sample image data sequence and the multiple first interactive network module units in the traversed fusion network module units, perform migration parameter learning on the multiple second interactive network module units in the fusion network module units until each of the fusion network module units meets the training convergence requirements, thereby generating a target concrete intelligent detection network that completes module parameter learning.

[0009] On the other hand, an embodiment of the present invention further provides a machine learning-based intelligent concrete detection system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present application realizes the effective mapping and connection between the network module units by decomposing and reconstructing the prior machine learning network and the basic concrete intelligent detection network. The method first uses the first example concrete sample image data sequence to perform parameter learning on each network module unit of the basic concrete intelligent detection network one by one, generates a temporary concrete intelligent detection network, and significantly improves the learning efficiency and accuracy of concrete characteristics. Subsequently, through module interaction processing, the key module units in the prior machine learning network and the temporary concrete intelligent detection network are integrated to form a plurality of fusion network module units. Finally, using the second example concrete sample image data sequence, the second interactive network module unit in the fusion network module unit is transferred to learn the parameters until the training convergence requirements are met, and the target concrete intelligent detection network that completes the module parameter learning is generated. This method not only improves the intelligent level of concrete detection, but also significantly enhances the generalization ability and detection accuracy of the concrete intelligent detection network, thereby facilitating the rapid and accurate evaluation of concrete quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the intelligent concrete detection method based on machine learning provided by an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of the intelligent concrete detection system based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The figure is a flow chart of a method for intelligent concrete detection based on machine learning provided by an embodiment of the present invention. The method for intelligent concrete detection based on machine learning is introduced in detail below.

[0014] Step S110, decomposing the priori machine learning network and the foundation concrete intelligent detection network into a plurality of network module units respectively. Each network module unit in the priori machine learning network and the foundation concrete intelligent detection network respectively constructs a mapping relationship.

[0015] In this embodiment, a priori machine learning network and a basic concrete intelligent detection network are pre-configured. The priori machine learning network may be a general network that has been pre-trained with a large amount of data, such as a convolutional neural network pre-trained on image data related to a variety of building materials. The basic concrete intelligent detection network is a network specially designed for concrete detection, but its initial performance may need to be improved.

[0016] The priori machine learning network has a complex architecture, including multiple network layers, such as convolutional layers, pooling layers, fully connected layers, etc. This embodiment decomposes at least every two adjacent network modules of the priori machine learning network into a network module unit based on the network layered architecture. For example, a convolutional layer and the pooling layer following it can be regarded as a network module unit. Assuming that there are 10 such adjacent module combinations in the priori machine learning network, 10 network module units are generated.

[0017] For the basic concrete intelligent detection network, it may be composed of modules specifically used to detect different properties of concrete (such as strength, porosity, etc.). In this embodiment, each network module of the basic concrete intelligent detection network is decomposed into a network module unit. For example, there is a module specifically used to detect concrete strength, which is used as an independent network module unit; there is also a module for detecting porosity, which is also used as a separate network module unit. Assuming that there are 5 such modules, 5 network module units are generated.

[0018] Then, this embodiment constructs a mapping relationship for these network module units. Taking a network module unit (including a convolution layer and a pooling layer) in the priori machine learning network as an example, it constructs a mapping relationship with a network module unit in the basic concrete intelligent detection network (such as a module unit for detecting concrete strength). This mapping relationship may be based on functional similarity or similarity of data processing processes. For example, this unit in the priori machine learning network extracts and preliminarily screens image features when processing image data, and the strength detection unit in the basic concrete intelligent detection network also needs to process relevant features in the concrete sample image, so a mapping relationship is established between them. In this way, a mapping relationship is constructed between each network module unit in the priori machine learning network and the basic concrete intelligent detection network, laying the foundation for subsequent learning and fusion operations.

[0019] Step S120, obtaining a first sample concrete sample image data sequence, traversing each network module unit of the basic concrete intelligent detection network, and based on the first sample concrete sample image data sequence, the target network module unit in the priori machine learning network and corresponding to the traversed network module unit, performing module parameter learning on the traversed network module unit one by one until the module parameter learning of all network module units of the basic concrete intelligent detection network is completed, thereby generating a temporary concrete intelligent detection network.

[0020] In the actual scenario of concrete production and quality control, this embodiment needs to obtain a high-quality first-instance concrete sample image data sequence. Assume that this embodiment collects samples from multiple concrete production batches, and each sample is captured at different angles and under different lighting conditions. These images constitute the first-instance concrete sample image data sequence. This data sequence contains a variety of original data information related to concrete, such as the surface texture, color, and whether there are cracks in the concrete.

[0021] This embodiment begins to traverse each network module unit of the basic concrete intelligent detection network. Take a network module unit in the basic concrete intelligent detection network as an example, which is a module unit for detecting concrete cracks. First, this embodiment loads the first example concrete sample image data sequence into the priori machine learning network and the target network module unit corresponding to the traversed crack detection network module unit. This target network module unit may be a unit with similar functions in the priori machine learning network, such as a unit that has similarities in processing image features.

[0022] Then, the target network module unit is used to perform feature detection on the first sample concrete sample image data sequence. The input layer of the target network module unit is composed of multiple neurons, and each neuron corresponds to receiving part of the data information in the first sample concrete sample image data sequence. For example, for a concrete sample image, the data information of different areas of the image is assigned to different neurons. In the input layer, the neurons perform linear transformation on the sample concrete sample image data in the first sample concrete sample image data sequence. Each neuron performs a weighted sum operation on the sample concrete sample image data according to the weight and bias parameters corresponding to the neuron and then transmits it to the hidden layer of the target network module unit. The hidden layer is composed of multiple hidden neurons, and the hidden neurons perform nonlinear transformation on the data transmitted by the input layer. The hidden neurons are interconnected and transmit information, and the output of each hidden neuron will serve as the input of the next layer of neurons. After being processed by multiple layers of hidden neurons in the hidden layer, the initial feature set is obtained.

[0023] Next, the information gain method is used to evaluate the feature importance of the initial feature set. For each feature in the initial feature set, the difference in information entropy of the sample concrete sample image data with and without the feature is calculated. This difference is the information gain of the feature. For example, when detecting concrete cracks, certain features of the concrete surface texture may have a high information gain for distinguishing whether there are cracks. The initial feature set is screened according to the results of the feature importance evaluation, and the screened features are combined using a rule-based feature combination method to generate a combined feature set.

[0024] The combined feature set is input again into the hidden layer of the target network module unit for deep mining. During the deep mining process, residual connections are used between the hidden layers. The residual connection allows the information of the combined feature set to be directly transferred from the previous hidden layer to the next hidden layer. After deep mining processing of multiple hidden layers, a deep mining feature set is obtained.

[0025] The deep mining feature set is then input into the output layer of the target network module unit, and the deep mining feature set is linearly combined through the weights and bias parameters of the neurons in the output layer. The output of each neuron in the output layer represents the possibility of the feature category corresponding to the neuron under the current input deep mining feature set. Finally, the Softmax classification algorithm is used to classify the outputs of each neuron in the output layer to convert the output value of each neuron in the output layer into a probability distribution between each category, thereby obtaining the first feature detection result generated by the target network module unit.

[0026] In this process, this embodiment locks the neuron weight information of other network module units in the basic concrete intelligent detection network, and only operates on the traversed crack detection network module unit. Based on the first example concrete sample image data sequence and the first feature detection result of the target network module unit, the neuron weight information of the traversed crack detection network module unit is module-by-module parameter learning.

[0027] Specifically, the first concrete sample image data sequence, the first feature detection result of the target network module unit, and the predefined first mapping relationship array are loaded into the traversed crack detection network module unit to obtain the second feature detection result generated by this network module unit. A matching operation is performed on the second feature detection result of the traversed crack detection network module unit and the predefined second mapping relationship array to obtain a first matching result. The module parameter learning error between the first matching result and the first feature detection result of the target network module unit is determined. If this module parameter learning error meets the training convergence requirement, the module parameter learning of the traversed crack detection network module unit is completed one by one.

[0028] In this way, each network module unit of the basic concrete intelligent detection network is traversed, such as the module unit for detecting concrete strength, porosity, etc., until the module parameter learning of all network module units of the basic concrete intelligent detection network is completed one by one, and finally a temporary concrete intelligent detection network is generated. This temporary concrete intelligent detection network integrates the knowledge of the prior machine learning network to a certain extent, and the detection ability of concrete sample image data has been initially improved.

[0029] Step S130, performing module interaction processing on the priori machine learning network and the temporary concrete intelligent detection network to generate a plurality of fused network module units. The fused network module unit includes a plurality of first interactive network module units extracted from the priori machine learning network, and a plurality of second interactive network module units corresponding to the plurality of first interactive network module units extracted from the temporary concrete intelligent detection network.

[0030] This embodiment has obtained a temporary concrete intelligent detection network and a priori machine learning network. The two networks are subjected to module interaction processing to generate multiple fusion network module units.

[0031] Take a network module unit in the priori machine learning network as an example. This unit has a unique ability to process image texture features. This network module unit is extracted from the priori machine learning network as the first interactive network module unit. Then, a corresponding network module unit is found in the temporary concrete intelligent detection network. This unit may also be a detection unit related to the concrete surface texture, and it is extracted as the second interactive network module unit corresponding to the above-mentioned first interactive network module unit. In this way, multiple first interactive network module units with different functions and characteristics are extracted from the priori machine learning network, and multiple second interactive network module units corresponding to these first interactive network module units are extracted from the temporary concrete intelligent detection network, thereby generating multiple fusion network module units.

[0032] These fusion network module units combine the advantages of the prior machine learning network and the temporary concrete intelligent detection network. For example, the first interactive network module unit in the prior machine learning network may have a more advanced feature extraction algorithm, while the second interactive network module unit in the temporary concrete intelligent detection network has a more specific optimization for the concrete detection task. Their fusion can make the new fusion network module unit show better performance in concrete detection.

[0033] Step S140, obtaining a second sample concrete sample image data sequence, traversing each of the multiple fusion network module units, and based on the second sample concrete sample image data sequence and the multiple first interactive network module units in the traversed fusion network module units, performing migration parameter learning on the multiple second interactive network module units in the fusion network module units until each of the fusion network module units meets the training convergence requirements, thereby generating a target concrete intelligent detection network that completes module parameter learning.

[0034] In the actual application scenario of concrete quality monitoring, this embodiment obtains a second example concrete sample image data sequence. This second example concrete sample image data sequence is also collected from different concrete samples and contains concrete image information in various quality states.

[0035] This embodiment begins to traverse each of the multiple fusion network module units. Take one of the fusion network module units as an example, this unit fuses the first interactive network module unit processing image edge features in the priori machine learning network and the second interactive network module unit related to concrete crack detection in the temporary concrete intelligent detection network.

[0036] First, the second sample concrete sample image data sequence is obtained and loaded into the first network module unit among the multiple first interactive network module units in the traversed fusion network module unit. The first network module unit starts to process the data and generates a first transfer learning result generated by each network module unit in the multiple first interactive network module units. For example, for the first interactive network module unit that processes the edge features of the image, information related to the edge features can be extracted based on the image information in the second sample concrete sample image data sequence and converted into the first transfer learning result.

[0037] Then, based on the second sample concrete sample image data sequence and the first transfer learning result generated by each network module unit in the plurality of first interactive network module units, transfer parameter learning is performed on each network module unit in the plurality of second interactive network module units. Specifically, the second sample concrete sample image data sequence is loaded into the first network module unit in the plurality of second interactive network module units in the traversed fusion network module unit, and the first mapping relationship array and the first transfer learning result of each network module unit in the plurality of first interactive network module units are loaded into each network module unit in the plurality of second interactive network module units, respectively, to generate the second transfer learning result generated by each network module unit in the plurality of second interactive network module units.

[0038] A matching operation is performed on the second transfer learning result of the last network module unit in the plurality of second interactive network module units and the second mapping relationship array to obtain a second matching result. For example, if the last network module unit is a unit related to concrete crack detection, then the second matching result reflects the degree of matching between the output of the network module unit related to crack detection and the expected result under the current transfer learning situation.

[0039] Next, the first transfer learning result of the last network module unit in the plurality of first interactive network module units and the transfer learning error between the second matching result are determined. If the transfer learning error meets the transfer learning convergence requirement, the transfer parameter learning of the traversed fusion network module unit is completed.

[0040] In this way, all fusion network module units are traversed, such as those related to concrete strength detection, porosity detection, etc., until each fusion network module unit meets the training convergence requirements, and finally a target concrete intelligent detection network with completed module parameter learning is generated. This target concrete intelligent detection network has highly optimized parameters and can accurately detect concrete sample image data, thus facilitating concrete quality control.

[0041] In addition, the present embodiment can also obtain the neuron weight information of the target concrete intelligent detection network that has completed the module parameter learning. The neuron weight information of the target concrete intelligent detection network is mapped and converted, and the neuron weight information of the target concrete intelligent detection network is mapped from the continuous numerical domain to the discrete numerical interval to obtain the mapped and converted target concrete intelligent detection network. Then, based on the mapped and converted target concrete intelligent detection network, concrete intelligent detection is performed on any input concrete sample image data.

[0042] When extracting neuron weight information from the target concrete intelligent detection network, it is necessary to take into account that the target concrete intelligent detection network contains multiple network layers, each network layer is composed of multiple neurons, and the connections between neurons have corresponding weight information. These weight information reflects the target concrete intelligent detection network's degree of cognition of different features and data relationships during the learning process.

[0043] The extracted neuron weight information is structurally analyzed to determine the distribution information of the neuron weight information in the target concrete intelligent detection network, such as the difference in the size range of weights between different network layers and the positive and negative distribution of weights. At the same time, the neuron weight information is statistically analyzed to determine the mean, median, standard deviation and other statistics of the neuron weight information. These statistics and distribution information constitute the analysis results of the neuron weight information.

[0044] According to the analysis results, the initial range of the discrete numerical interval is determined. For example, the range of the discrete numerical interval is determined based on the multiples of the standard deviation of the neuron weight information, with the mean value of the neuron weight information as the center. Then, according to the defined task detection indicators for the target concrete intelligent detection network and the boundary conditions of the discrete numerical interval, the initial range of the discrete numerical interval is fine-tuned to generate the final discrete numerical interval.

[0045] The mapping strategy is determined based on the analysis results and the discrete numerical interval. If the mapping strategy is a linear mapping strategy, the corresponding value of each neuron weight information in the discrete numerical interval is calculated based on the proportional relationship between each neuron weight information and the continuous numerical domain range. Alternatively, under the nonlinear mapping strategy, the corresponding mapping rules are used to convert the segmented range of each neuron weight information. During the mapping conversion process, the mapping relationship between the original value of each neuron weight information and the mapped conversion value is recorded.

[0046] Finally, according to the neuron weight information after mapping conversion, the structure of the target concrete intelligent detection network is adjusted, and the pre-defined verification data set is used to perform functional verification on the adjusted target concrete intelligent detection network. According to the functional verification results, the target concrete intelligent detection network is fine-tuned, and finally the target concrete intelligent detection network after mapping conversion is obtained, thereby achieving more accurate and efficient detection of concrete sample image data.

[0047] Based on the above steps, the embodiment of the present application realizes the effective mapping and connection between the network module units by decomposing and reconstructing the priori machine learning network and the basic concrete intelligent detection network. The method first uses the first example concrete sample image data sequence to perform parameter learning on each network module unit of the basic concrete intelligent detection network one by one, generates a temporary concrete intelligent detection network, and significantly improves the learning efficiency and accuracy of concrete characteristics. Subsequently, through module interaction processing, the key module units in the priori machine learning network and the temporary concrete intelligent detection network are integrated to form a plurality of fusion network module units. Finally, using the second example concrete sample image data sequence, the second interactive network module unit in the fusion network module unit is transferred and learned until the training convergence requirements are met, and the target concrete intelligent detection network that completes the module parameter learning is generated. This method not only improves the intelligent level of concrete detection, but also significantly enhances the generalization ability and detection accuracy of the concrete intelligent detection network, thereby facilitating the rapid and accurate evaluation of concrete quality.

[0048] In a possible implementation, step S120 includes:

[0049] Step S121, obtaining a first sample concrete sample image data sequence, loading the first sample concrete sample image data sequence into the priori machine learning network and the target network module unit corresponding to the traversed network module unit, using the target network module unit to perform feature detection on the first sample concrete sample image data sequence, and obtaining a first feature detection result generated by the target network module unit.

[0050] In this embodiment, the first step is to obtain a first-example concrete sample image data sequence, which is derived from multiple actual concrete engineering samples. For example, concrete test block samples collected from different construction sites may be in different curing stages, have different mix ratios, and are made under different environmental conditions. Multiple sets of images are taken for each sample, which comprehensively record the appearance characteristics of the concrete sample from different angles and under different lighting conditions. These images together constitute the first-example concrete sample image data sequence, which contains various raw data information related to concrete, such as the texture and color of the concrete surface, and whether there are tiny cracks and holes.

[0051] Next, each network module unit of the basic concrete intelligent detection network is traversed. This basic concrete intelligent detection network is specially constructed for detecting various characteristics of concrete. It contains multiple network module units, each of which is responsible for detecting a specific aspect of concrete or performing a specific data processing function. For example, one of the network module units is responsible for detecting cracks on the concrete surface, and another network module unit is responsible for analyzing the porosity of the concrete. When this embodiment traverses to the network module unit for detecting concrete cracks, the first example concrete sample image data sequence will be loaded into the priori machine learning network and the target network module unit corresponding to the traversed network module unit. The priori machine learning network is a network that has been pre-trained on a large amount of data, and it has certain feature extraction and analysis capabilities. This target network module unit has a certain association with the crack detection unit in the basic concrete intelligent detection network, which may be similar in function or data processing flow. Then the target network module unit is used to perform feature detection on the first example concrete sample image data sequence.

[0052] The target network module unit starts to process the first sample concrete sample image data sequence loaded in. The input layer of the target network module unit is composed of multiple neurons, which receive part of the data information in the first sample concrete sample image data sequence according to specific rules. For example, data related to a certain area or a certain feature in the image will be received by a specific neuron. In the input layer, the neurons will perform linear transformation on the sample concrete sample image data in the first sample concrete sample image data sequence. Each neuron performs a weighted summation operation on the sample concrete sample image data according to its corresponding weight and bias parameters, and then passes the result to the hidden layer of the target network module unit. The hidden layer contains multiple hidden neurons, which perform nonlinear transformation on the data transmitted by the input layer. The hidden neurons are interconnected and transmit information. The output of each hidden neuron will be used as the input of the next layer of neurons. After being processed by multiple layers of hidden neurons in the hidden layer, the initial feature set is obtained.

[0053] Then, the information gain method is used to evaluate the feature importance of the initial feature set. For each feature in the initial feature set, the information gain is determined by calculating the difference in information entropy of the sample concrete sample image data when the feature is included and when the feature is not included. This difference can reflect the importance of the feature for distinguishing different sample concrete sample image data. For example, when detecting concrete cracks, certain specific patterns of concrete surface texture have a high information gain for judging whether cracks exist. The initial feature set is screened according to the results of the feature importance evaluation, and the screened features are combined using a rule-based feature combination method to generate a combined feature set. The combined feature set is then input into the hidden layer of the target network module unit for deep mining. During the deep mining process, residual connections are used between the hidden layers. This residual connection allows the information of the combined feature set to be directly transferred from the previous hidden layer to the subsequent hidden layer. After deep mining processing of multiple hidden layers, a deep mining feature set is obtained. Finally, the deep mining feature set is input into the output layer of the target network module unit. The deep mining feature set is linearly combined through the weights and bias parameters of the neurons in the output layer. The output of each neuron in the output layer represents the possibility of the feature category corresponding to the neuron under the current input deep mining feature set. The Softmax classification algorithm is used to classify the outputs of each neuron in the output layer, and the output value of each neuron in the output layer is converted into a probability distribution between each category, thereby obtaining the first feature detection result generated by the target network module unit.

[0054] Step S122, traverse each network module unit of the basic concrete intelligent detection network, lock the neuron weight information of other network module units, and based on the first sample concrete sample image data sequence and the first feature detection result of the target network module unit, perform module-by-module parameter learning on the neuron weight information of the traversed network module units until the module-by-module parameter learning of all network module units of the basic concrete intelligent detection network is completed to generate a temporary concrete intelligent detection network.

[0055] In a possible implementation, step S122 includes:

[0056] Step S1221, loading the first sample concrete sample image data sequence, the first feature detection result of the target network module unit, and the predefined first mapping relationship array into the traversed network module unit to obtain the second feature detection result generated by the traversed network module unit, wherein the first sample concrete sample image data sequence includes various original data information related to concrete.

[0057] Step S1222: performing a matching operation on the second feature detection result of the traversed network module unit and a predefined second mapping relationship array to obtain a first matching result.

[0058] Step S1223, determining a module parameter learning error between the first matching result and the first feature detection result of the target network module unit.

[0059] Step S1224: If the module parameter learning error meets the training convergence requirement, the module parameter learning of the traversed network module unit is completed one by one.

[0060] When traversing each network module unit of the basic concrete intelligent detection network, the neuron weight information of other network module units can be locked, and only the current traversed network module unit can be focused on. Based on the first example concrete sample image data sequence and the first feature detection result of the target network module unit, the neuron weight information of the traversed network module unit is module-by-module parameter learning. Specifically, the first example concrete sample image data sequence, the first feature detection result of the target network module unit, and the predefined first mapping relationship array are loaded into the traversed network module unit. This first mapping relationship array contains some pre-set rule information about data processing and feature mapping. After loading, the traversed network module unit will process these input information according to its own algorithm and structure, thereby obtaining the second feature detection result generated by the traversed network module unit.

[0061] Then, a matching operation is performed on the second feature detection result of the traversed network module unit and the predefined second mapping relationship array to obtain a first matching result. This second mapping relationship array also contains rule information related to data processing and result matching. The first matching result reflects the degree of matching between the processing result of the traversed network module unit under the current input and rule and the expected result. Then, the module parameter learning error between the first matching result and the first feature detection result of the target network module unit is determined. This error is an important indicator to measure the gap between the learning effect of the traversed network module unit and the target network module unit. If this module parameter learning error meets the training convergence requirements, it means that the traversed network module unit has achieved the expected effect in the current learning process, and the module parameter learning of the traversed network module unit is completed one by one. In this way, all network module units in the foundation concrete intelligent detection network, such as network module units for detecting concrete strength, porosity, etc., are traversed in turn until the module parameter learning of all network module units in the foundation concrete intelligent detection network is completed one by one, and finally a temporary concrete intelligent detection network is generated. This temporary concrete intelligent detection network integrates the knowledge of the prior machine learning network to a certain extent, so that its performance in concrete detection is improved.

[0062] In a possible implementation, step S121 includes:

[0063] Step S1211: input the first sample concrete sample image data sequence into the input layer of the target network module unit, wherein the input layer of the target network module unit is composed of a plurality of neurons, and each neuron corresponds to a portion of the data information in the first sample concrete sample image data sequence received.

[0064] In this embodiment, a first concrete sample image data sequence is input to the input layer of the target network module unit. This target network module unit is a part of the priori machine learning network, and its input layer is composed of multiple neurons. These neurons have specific settings in structure and function, and each neuron corresponds to receiving part of the data information in the first concrete sample image data sequence. For example, when processing concrete sample images, data related to different areas or different attributes of the image will be assigned to different neurons. This is based on the structured processing requirements for concrete sample image data, so as to facilitate more efficient data processing and feature extraction in the subsequent process.

[0065] Step S1212: In the input layer, linear transformation is performed on the sample concrete sample image data in the first sample concrete sample image data sequence through neurons, wherein each neuron performs a weighted sum operation on the sample concrete sample image data according to the weight and bias parameters corresponding to the neuron and then transmits the data to the hidden layer of the target network module unit, the hidden layer is composed of a plurality of hidden neurons, and the hidden neurons perform nonlinear transformation on the data transmitted from the input layer. The hidden neurons are interconnected and transmit information, and the output of each hidden neuron is used as the input of the next layer of neurons. After being processed by the multiple layers of hidden neurons in the hidden layer, an initial feature set is obtained.

[0066] In the input layer, neurons begin to perform linear transformation on the sample concrete sample image data in the first sample concrete sample image data sequence. Each neuron has specific weight and bias parameters, which are determined during the pre-training or initialization process of the target network module unit. According to these weight and bias parameters, neurons perform weighted summation operation on the sample concrete sample image data. This weighted summation operation is based on the functional mechanism of neurons and aims to highlight or weaken certain data features. The result after the weighted summation operation will be passed to the hidden layer of the target network module unit.

[0067] The hidden layer of the target network module unit is composed of multiple hidden neurons. The main function of these hidden neurons is to perform nonlinear transformation on the data transmitted by the input layer. Each hidden neuron receives data from the input layer and processes it according to its own algorithm and parameter settings. The hidden neurons are connected to each other and transmit information. This connection structure and information transmission method constitute a complex neural network interaction system. The output of each hidden neuron will serve as the input of the next layer of neurons. This data flow method allows the data to be processed in multiple layers in the hidden layer. After being processed by the multiple layers of hidden neurons in the hidden layer, the initial feature set is obtained. This initial feature set contains various features extracted from the original concrete sample image data. These features are an abstract representation of the concrete sample image data, which may involve information such as the surface texture, color distribution, and structural characteristics of the concrete.

[0068] Step S1213, using the information gain method to evaluate the feature importance of the initial feature set, and generating the importance of each feature in the initial feature set to the sample concrete sample image data as a feature importance evaluation result. For each feature in the initial feature set, the difference in information entropy of the sample concrete sample image data when the feature is included and when the feature is not included is calculated, and the difference is the information gain of the feature. The larger the information gain, the more important the feature is for distinguishing different sample concrete sample image data.

[0069] For example, when detecting cracks in concrete sample images, certain texture features on the concrete surface may have a greater impact on the information entropy of the sample image data in the presence and absence of cracks, thus having a higher information gain. The calculation of this information gain is based on the principle of information theory, which measures the importance of features by quantifying their impact on data uncertainty. Based on the information gain of each feature, the importance of each feature in the initial feature set for the sample concrete sample image data can be generated as a feature importance evaluation result.

[0070] Step S1214, screening the initial feature set according to the result of the feature importance evaluation, and combining the screened features using a rule-based feature combination method to generate a combined feature set.

[0071] In this process, according to the pre-set screening criteria, the features that contribute less to distinguishing different sample concrete sample image data are removed, and the features with higher importance are retained. Then, the screened features are combined using a rule-based feature combination method. This rule-based feature combination method is designed according to the specific needs of concrete detection and the algorithm of the target network module unit. For example, some features may be related to the strength of concrete, while others may be related to the porosity of concrete. According to the intrinsic relationship between these features and the detection target, they are combined into a new feature set, namely the combined feature set.

[0072] Step S1215, the combined feature set is input again into the hidden layer of the target network module unit for deep mining. During the deep mining process, residual connections are used between the hidden layers, wherein the residual connections allow the information of the combined feature set to be directly transferred from the previous hidden layer to the subsequent hidden layer. After deep mining processing of multiple hidden layers, a deep mining feature set is obtained.

[0073] Residual connection is a special neural network connection method that allows the information of the combined feature set to be directly passed from the previous hidden layer to the next hidden layer. This connection method helps to solve the gradient vanishing problem of deep neural networks during training, while better retaining the original feature information. After deep mining processing of multiple hidden layers, a deep mining feature set is obtained. This deep mining feature set contains the image features of concrete samples after deep processing and mining. These features are more abstract and advanced, and can better reflect the intrinsic characteristics of concrete samples.

[0074] Step S1216, input the deep mining feature set to the output layer of the target network module unit, and perform linear combination on the deep mining feature set through the weights and bias parameters of the neurons in the output layer, wherein the output of each neuron in the output layer represents the possibility of the feature category corresponding to the neuron under the currently input deep mining feature set.

[0075] In this linear combination process, the output of each neuron in the output layer represents the possibility of the feature category corresponding to the neuron under the current input deep mining feature set. For example, for different quality categories of concrete (such as high quality, medium quality, low quality) or different performance indicators (such as strength grade, porosity range, etc.), the output of each neuron reflects the degree of association between the current deep mining feature set and the feature category.

[0076] Step S1217, using the Softmax classification algorithm to classify the outputs of each neuron in the output layer, so as to convert the output value of each neuron in the output layer into a probability distribution between various categories, and generate the first feature detection result.

[0077] The Softmax classification algorithm is a commonly used multi-classification algorithm that converts the output value of each neuron in the output layer into a probability distribution between categories. Through this conversion, a vector representing the probability of different categories can be obtained. This vector is the first feature detection result generated by the target network module unit. For example, in concrete detection, this result may indicate the probability distribution of concrete samples belonging to different strength grades or different quality categories, thereby providing a basis for subsequent analysis and decision-making. This neural network-based feature detection process effectively extracts feature information related to concrete characteristics through the step-by-step processing and analysis of concrete sample image data, laying the foundation for intelligent detection of concrete.

[0078] In a possible implementation, step S140 includes:

[0079] Step S141, obtaining a second sample concrete sample image data sequence, loading the second sample concrete sample image data sequence into the first network module unit among the multiple first interactive network module units in the traversed fusion network module unit, and generating a first transfer learning result generated by each network module unit in the multiple first interactive network module units.

[0080] In this embodiment, the second sample concrete sample image data sequence is also collected from actual concrete samples, and similar to the first sample concrete sample image data sequence, contains rich concrete sample information. For example, these samples come from different construction sites and different concrete production batches, and they differ in raw materials, mix ratios, curing conditions, etc., thus reflecting different concrete properties. These samples are photographed into images, acquired from multiple angles and under different lighting conditions, to comprehensively cover various appearance features that may appear in concrete, such as surface texture, color, cracks, pores, etc. These images together constitute the second sample concrete sample image data sequence.

[0081] Step S142: Based on the second example concrete sample image data sequence and the first transfer learning result generated by each network module unit in the plurality of first interactive network module units, transfer parameter learning is performed on each network module unit in the plurality of second interactive network module units.

[0082] In a possible implementation, step S142 includes:

[0083] Step S1421, loading the second sample concrete sample image data sequence into the first network module unit among the multiple second interactive network module units in the traversed fusion network module unit, and loading the first mapping relationship array and the first transfer learning result of each network module unit in the multiple first interactive network module units into each network module unit in the multiple second interactive network module units respectively, to generate the second transfer learning result generated by each network module unit in the multiple second interactive network module units.

[0084] Take a fusion network module unit as an example. This fusion network module unit integrates the first interactive network module unit related to concrete surface texture analysis in the priori machine learning network and the second interactive network module unit related to concrete crack detection in the temporary concrete intelligent detection network. In this process, the first network module unit starts to process the data. Since this network module unit is part of the priori machine learning network, it has its own algorithm and parameter settings. After receiving the second sample concrete sample image data sequence, it analyzes and processes the data based on its own neuron structure and connection weights. For example, for the texture information in the image, the neuron will calculate according to the pre-set weights and bias parameters to extract the features related to the texture. After a series of calculations and processing, the first transfer learning results generated by each network module unit in multiple first interactive network module units are generated. These results reflect the preliminary analysis and feature extraction of the second sample concrete sample image data sequence by the first network module unit, and these results will serve as an important basis for subsequent operations.

[0085] Specifically, the second sample concrete sample image data sequence is loaded into the first network module unit of the multiple second interactive network module units in the traversed fusion network module unit. At the same time, the first mapping relationship array and the first transfer learning result of each network module unit in the multiple first interactive network module units are respectively loaded into each network module unit in the multiple second interactive network module units. This first mapping relationship array contains some predefined rule information about data mapping and conversion, which matches the structure and function of the entire network. After loading these data and information, each network module unit in the multiple second interactive network module units begins to calculate and process. For example, in the second interactive network module unit related to concrete crack detection, according to the image information in the received second sample concrete sample image data sequence, the first mapping relationship array and the texture features in the first transfer learning result, etc., calculation can be performed according to its own algorithm. After calculation, the second transfer learning result generated by each network module unit in the multiple second interactive network module units is generated. These results are outputs after fusing multiple information, reflecting the processing status of the second interactive network module unit under the current input.

[0086] Step S1422, performing a matching operation on the second transfer learning result of the last network module unit in the plurality of second interactive network module units and the second mapping relationship array to obtain a second matching result.

[0087] This second mapping relationship array also contains rule information related to result matching, which is specially set to evaluate the output result of the second interactive network module unit. When performing the matching operation, the second transfer learning result of the last network module unit will be compared and evaluated according to the rules in the second mapping relationship array. For example, if the last network module unit is related to concrete crack detection, then the data related to the crack characteristics in the second transfer learning result will be matched according to the rules for crack detection result evaluation in the second mapping relationship array. This matching operation may involve evaluation criteria for crack location, length, width and other features, and the second matching result is obtained by comparing the actual second transfer learning result with the expected rule requirements. This second matching result reflects the degree of conformity between the output of the last second interactive network module unit and the expected result.

[0088] Step S1423, determining the first transfer learning result of the last network module unit in the plurality of first interactive network module units and the transfer learning error between the second matching result.

[0089] This transfer learning error is an important indicator to measure the difference between the learning effect of the second interactive network module unit in the transfer learning process and the first interactive network module unit. For example, in the module unit related to concrete crack detection, if the first transfer learning result of the first interactive network module unit on the relationship between texture features and cracks is different from the second matching result of the second interactive network module unit on crack detection, then this difference constitutes a transfer learning error. Through a specific calculation method, the specific value of this transfer learning error is determined according to the degree of difference between the two.

[0090] Step S1424: If the transfer learning error meets the transfer learning convergence requirement, the transfer parameter learning of the traversed fusion network module unit is completed.

[0091] If this transfer learning error meets the transfer learning convergence requirements, it means that in the current transfer learning process, multiple second interactive network module units have achieved the expected learning effect. The migration parameter learning of the traversed fusion network module unit is completed. In this way, each fusion network module unit is traversed. For example, the first interactive network module unit related to concrete strength detection in the priori machine learning network and the fusion network module unit related to concrete porosity detection in the temporary concrete intelligent detection network are also operated according to the same process. Until each fusion network module unit meets the transfer learning convergence requirements, the entire transfer parameter learning process is completed, and the target concrete intelligent detection network with completed module parameter learning is generated. After the above transfer parameter learning process, this target concrete intelligent detection network can detect concrete sample image data more accurately and improve the accuracy and reliability of concrete quality detection.

[0092] In a possible implementation, step S110 includes:

[0093] According to the network layered architecture, at least every two adjacent network modules of the priori machine learning network are decomposed into a network module unit to generate multiple network module units corresponding to the priori machine learning network, and each network module of the basic concrete intelligent detection network is decomposed into a network module unit to generate multiple network module units corresponding to the basic concrete intelligent detection network.

[0094] The method also includes: constructing the priori machine learning network and the corresponding relationship between each network module unit in the foundation concrete intelligent detection network according to the network layered architecture.

[0095] In this embodiment, the priori machine learning network usually includes multiple network modules, such as convolutional layers, pooling layers, fully connected layers, etc., and these modules are connected in a specific order and logic. According to the principle that at least every two adjacent network modules are decomposed into a network module unit, for example, a convolutional layer and the pooling layer following it are regarded as a whole to form a network module unit. Assuming that there are multiple such adjacent module combinations in the priori machine learning network, after decomposition, multiple network module units corresponding to the priori machine learning network are generated.

[0096] Taking a specific priori machine learning network as an example, it may be a deep convolutional neural network with certain pre-training results in the field of image recognition. The front end of this deep convolutional neural network may be a plurality of convolutional layers and pooling layers arranged alternately, and the first convolutional layer and the first pooling layer may be combined into the first network module unit. In this unit, the convolutional layer is responsible for feature extraction of the input concrete sample image data, such as detecting edges, textures and other features in the image, and the pooling layer downsamples the feature map output by the convolutional layer to reduce the amount of data while retaining important features. Subsequent adjacent convolutional layers and pooling layers also form different network module units in the same way.

[0097] For the basic concrete intelligent detection network, each network module is decomposed into a network module unit. The basic concrete intelligent detection network is specially constructed for concrete detection, and its network modules may have different functions. For example, some modules are responsible for detecting the strength of concrete, and some modules are responsible for detecting the porosity of concrete. Each network module with a specific function is divided as a network module unit, thereby generating multiple network module units corresponding to the basic concrete intelligent detection network.

[0098] For example, in the basic concrete intelligent detection network, there is a module dedicated to analyzing concrete strength. This module determines the strength grade of concrete by processing the input concrete sample image data with a specific algorithm. This module is decomposed into an independent network module unit.

[0099] After completing the above decomposition, the correspondence between the prior machine learning network and the network module units in the basic concrete intelligent detection network is constructed according to the network layered architecture. This correspondence is based on the similarity of the network module units in terms of function, data processing flow, etc.

[0100] Taking a network module unit in the priori machine learning network (including convolutional layers and pooling layers, mainly for preliminary feature extraction of images) as an example, a corresponding relationship can be established with a network module unit in the basic concrete intelligent detection network (responsible for detecting the surface texture features of concrete, and the surface texture features are related to image feature extraction). The determination of this corresponding relationship helps to better utilize the knowledge in the priori machine learning network to improve the performance of the basic concrete intelligent detection network in subsequent operations, such as parameter learning and fusion processing. For another example, the network module unit related to feature screening in the priori machine learning network may establish a corresponding relationship with the network module unit for analyzing important concrete features in the basic concrete intelligent detection network, so as to achieve effective transmission and sharing of information in the data processing process, thereby improving the accuracy and efficiency of the entire concrete intelligent detection system.

[0101] In a possible implementation, the method further includes:

[0102] Step S150, obtaining neuron weight information of the target concrete intelligent detection network that has completed module parameter learning.

[0103] For example, when testing the strength and porosity of concrete, the target concrete intelligent detection network will adjust the connection weights between neurons during the learning process based on the relationship between different features (such as texture features, color features, etc. in the concrete sample image) and the target detection results (values ​​of strength and porosity). For features that are highly relevant to strength detection, the network will give higher weights to the corresponding connections, indicating that these features are important in determining strength; while for features that are not closely related to porosity, their connection weights may be lower.

[0104] Step S160, performing a mapping conversion operation on the neuron weight information of the target concrete intelligent detection network, mapping the neuron weight information of the target concrete intelligent detection network from a continuous numerical domain to a discrete numerical interval, and obtaining a mapped and converted target concrete intelligent detection network;

[0105] Step S170: performing intelligent concrete detection on any input concrete sample image data based on the target intelligent concrete detection network after the mapping conversion.

[0106] In this embodiment, when new concrete sample image data is input into the target concrete intelligent detection network after the mapping conversion, the target concrete intelligent detection network will process according to the adjusted structure and weight information. For example, a new concrete sample image is input, and the network starts from the input layer, and performs feature extraction, analysis and comprehensive judgment on the image data according to the weight information of the neurons, and gradually obtains the detection results of the characteristics such as concrete strength and porosity. This optimized target concrete intelligent detection network can detect concrete samples more accurately, improve the reliability and efficiency of concrete quality detection, and meet the needs of concrete engineering in practical applications.

[0107] In a possible implementation, step S160 includes:

[0108] Step S161, extracting neuron weight information from the target concrete intelligent detection network, wherein the target concrete intelligent detection network comprises a plurality of network layers, each network layer is composed of a plurality of neurons, and the connection between neurons has corresponding weight information, and the weight information reflects the degree of recognition of different features and data relationships by the target concrete intelligent detection network during the learning process.

[0109] Step S162, structural analysis is performed on the extracted neuron weight information to determine the distribution information of the neuron weight information in the target concrete intelligent detection network, wherein the distribution information includes the difference in the size range of weights between different network layers and the positive and negative distribution of weights. Also, statistical analysis is performed on the neuron weight information to determine the statistics of the mean, median, and standard deviation of the neuron weight information, wherein the statistics and the distribution information constitute the analysis result of the neuron weight information.

[0110] In this embodiment, after extracting neuron weight information from the target concrete intelligent detection network, the extracted neuron weight information must first be subjected to structural analysis. Different network layers in the target concrete intelligent detection network differ in function and data processing, which is reflected in the distribution of neuron weight information. For example, in the network layer close to the input layer, the numerical range of the weight may be small, which is mainly responsible for the preliminary feature extraction of the original image data. These weights are more related to the basic features of the image (such as edges, colors, etc.); while in the network layer close to the output layer, the numerical range of the weight may be large, because these layers need to combine the previously processed features to obtain the final detection results such as concrete strength and porosity. At the same time, the positive and negative distribution of weights is also meaningful. Positive weights may indicate a positive correlation between a certain feature and the target detection result, while negative weights may indicate a negative correlation. In addition to structural analysis, statistical analysis of neuron weight information is also required to determine its mean, median, standard deviation and other statistics. These statistics and distribution information together constitute the analysis results of neuron weight information.

[0111] Step S163, determining an initial range of a discrete numerical interval according to the analysis result, wherein the range of the discrete numerical interval is determined based on multiples of a standard deviation of the neuron weight information with the mean value of the neuron weight information as the center.

[0112] For example, if the initial discrete value interval is determined with the mean as the center and a fluctuation of 2 standard deviations, then this interval will contain most of the weight values. This is based on statistical principles. The weight values ​​within this range are highly representative and can also reflect the network's cognition of different features and data relationships to a certain extent.

[0113] Step S164, fine-tuning the initial range of the discrete numerical interval according to the task detection index defined for the target concrete intelligent detection network and the boundary conditions of the discrete numerical interval, to generate a final discrete numerical interval.

[0114] In this embodiment, the initial range of the discrete numerical interval determined only based on statistical principles may not fully meet the specific needs of the target concrete intelligent detection network, so it is necessary to fine-tune the initial range of the discrete numerical interval based on the task detection index defined for the target concrete intelligent detection network and the boundary conditions of the discrete numerical interval, so as to generate the final discrete numerical interval. The task detection index is set according to the specific requirements of concrete detection, such as the accuracy requirements for concrete strength detection, the error range for porosity detection, etc. If a task detection index requires a higher detection accuracy for a specific feature, and the weight corresponding to the feature is near the boundary of the initial discrete numerical interval, it may be necessary to adjust the interval range to ensure that the weight can more accurately reflect its role in detection after the mapping conversion.

[0115] Step S165, according to the analysis result and the discrete numerical interval, a mapping strategy is determined, and according to the determined mapping strategy, each neuron weight information in the target concrete intelligent detection network is mapped and converted, wherein if the mapping strategy is a linear mapping strategy, the corresponding value of each neuron weight information in the discrete numerical interval is calculated according to the proportional relationship between each neuron weight information and the continuous numerical domain range. Alternatively, under a nonlinear mapping strategy, according to the segmented range where each neuron weight information is located, the conversion is performed according to the corresponding mapping rules, and in the mapping conversion process, the mapping relationship between the original value of each neuron weight information and the mapped conversion value is recorded.

[0116] If the mapping strategy is a linear mapping strategy, then the corresponding value of each neuron weight information in the discrete numerical interval is calculated according to the proportional relationship between each neuron weight information and the continuous numerical domain range. For example, assuming that the range of the continuous numerical domain is [a, b], the discrete numerical interval is [c, d], and a neuron weight information is x, then its corresponding value y in the discrete numerical interval can be calculated by the linear relationship (y - c) / (d - c) = (x - a) / (b - a). This linear mapping strategy is suitable for the case where the weight information has a relatively simple linear relationship between the continuous numerical domain and the discrete numerical interval. Under the nonlinear mapping strategy, the weight information of each neuron is converted according to the corresponding mapping rules according to the segment range. For example, there may be different mapping rules for weight information in different numerical ranges (such as less than a threshold, between two thresholds, greater than another threshold, etc.), which may involve the application of nonlinear functions such as logarithmic functions and exponential functions. In the mapping conversion process, the mapping relationship between the original value of each neuron weight information and the mapped conversion value is recorded, which helps to trace and evaluate the effect of the mapping conversion in the subsequent analysis and optimization process.

[0117] Step S166, adjusting the structure of the target concrete intelligent detection network according to the neuron weight information after mapping conversion, and using a predefined verification data set to perform functional verification on the adjusted target concrete intelligent detection network, and fine-tuning the target concrete intelligent detection network according to the functional verification result, and finally obtaining the target concrete intelligent detection network after mapping conversion.

[0118] In this embodiment, some neuron connections in the target concrete intelligent detection network can be re-evaluated and adjusted, because the change of weight information may affect the information transmission and processing flow of the entire network. Then, a pre-defined verification data set is used to perform functional verification on the adjusted target concrete intelligent detection network. The verification data set is a set of concrete sample image data independent of the training data set, which contains concrete samples in various different situations and can comprehensively test the performance of the network. The target concrete intelligent detection network is fine-tuned according to the functional verification results. If it is found during the verification process that some test results do not meet expectations, such as the error of strength detection exceeds the specified range or the accuracy of porosity detection is lower than the requirement, the network needs to be adjusted. This adjustment may include further fine-tuning of some neuron weights, adjusting the structure of the network (such as adding or reducing some neuron connections), and finally obtaining the target concrete intelligent detection network after mapping conversion.

[0119] Figure 2 The hardware structure of the machine learning-based intelligent concrete detection system 100 for implementing the above-mentioned machine learning-based intelligent concrete detection method provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the machine learning-based intelligent concrete detection system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0120] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the machine-learning-based intelligent concrete detection system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0121] In the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the intelligent concrete detection method based on machine learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0122] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned machine learning-based concrete intelligent detection system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0123] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned intelligent concrete detection method based on machine learning is implemented.

[0124] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A method for intelligent concrete detection based on machine learning, characterized in that: The method comprises: Decomposing the priori machine learning network and the basic concrete intelligent detection network into a plurality of network module units respectively; constructing mapping connections respectively between each network module unit in the priori machine learning network and the basic concrete intelligent detection network; Acquire a first sample concrete sample image data sequence, traverse each network module unit of the basic concrete intelligent detection network, and perform module parameter learning on the traversed network module units one by one based on the first sample concrete sample image data sequence, the target network module units in the priori machine learning network and corresponding to the traversed network module units, until the module parameter learning of all network module units of the basic concrete intelligent detection network is completed, thereby generating a temporary concrete intelligent detection network; The a priori machine learning network and the temporary concrete intelligent detection network are subjected to module interaction processing to generate a plurality of fused network module units; the fused network module unit comprises a plurality of first interactive network module units extracted from the a priori machine learning network, and a plurality of second interactive network module units corresponding to the plurality of first interactive network module units extracted from the temporary concrete intelligent detection network, wherein the module interaction processing refers to extracting a plurality of first interactive network module units with different functions and characteristics from the a priori machine learning network, and extracting a plurality of second interactive network module units corresponding to the first interactive network module units from the temporary concrete intelligent detection network, thereby generating a plurality of fused network module units; Acquire a second sample concrete sample image data sequence, traverse each of the multiple fusion network module units, and based on the second sample concrete sample image data sequence and the multiple first interactive network module units in the traversed fusion network module units, perform migration parameter learning on the multiple second interactive network module units in the fusion network module units until each of the fusion network module units meets the training convergence requirements, thereby generating a target concrete intelligent detection network that completes module parameter learning.

2. The intelligent concrete detection method based on machine learning according to claim 1 is characterized in that: The method of acquiring a first sample concrete sample image data sequence, traversing each network module unit of the basic concrete intelligent detection network, and performing module parameter learning on the traversed network module units one by one based on the first sample concrete sample image data sequence, the target network module units in the priori machine learning network and corresponding to the traversed network module units, until module parameter learning of all network module units of the basic concrete intelligent detection network is completed, thereby generating a temporary concrete intelligent detection network, including: Acquire a first sample concrete sample image data sequence, load the first sample concrete sample image data sequence into the priori machine learning network and the target network module unit corresponding to the traversed network module unit, perform feature detection on the first sample concrete sample image data sequence using the target network module unit, and obtain a first feature detection result generated by the target network module unit; Each network module unit of the basic concrete intelligent detection network is traversed, the neuron weight information of other network module units is locked, and based on the first sample concrete sample image data sequence and the first feature detection result of the target network module unit, module-by-module parameter learning is performed on the neuron weight information of the traversed network module units until the module-by-module parameter learning of all network module units of the basic concrete intelligent detection network is completed to generate a temporary concrete intelligent detection network.

3. The intelligent concrete detection method based on machine learning according to claim 2 is characterized in that: Based on the first sample concrete sample image data sequence and the first feature detection result of the target network module unit, the neuron weight information of the traversed network module unit is subjected to module-by-module parameter learning, including: The first sample concrete sample image data sequence, the first feature detection result of the target network module unit, and the predefined first mapping relationship array are loaded into the traversed network module unit to obtain the second feature detection result generated by the traversed network module unit, wherein the first sample concrete sample image data sequence includes various raw data information related to concrete; Performing a matching operation on the second feature detection result of the traversed network module unit and a predefined second mapping relationship array to obtain a first matching result; Determining a module parameter learning error between the first matching result and a first feature detection result of the target network module unit; If the module parameter learning error meets the training convergence requirement, the module parameter learning of the traversed network module unit is completed one by one.

4. The intelligent concrete detection method based on machine learning according to claim 2 is characterized in that: The step of using the target network module unit to perform feature detection on the first example concrete sample image data sequence to obtain a first feature detection result generated by the target network module unit includes: Inputting the first sample concrete sample image data sequence into an input layer of a target network module unit, wherein the input layer of the target network module unit is composed of a plurality of neurons, and each neuron corresponds to a portion of data information in the first sample concrete sample image data sequence received; In the input layer, a linear transformation is performed on the sample concrete sample image data in the first sample concrete sample image data sequence through neurons, wherein each neuron performs a weighted sum operation on the sample concrete sample image data according to a weight and a bias parameter corresponding to the neuron and then transmits the data to a hidden layer of the target network module unit, wherein the hidden layer is composed of a plurality of hidden neurons, and the hidden neurons perform a nonlinear transformation on the data transmitted from the input layer, and the hidden neurons are interconnected and perform information transmission, and the output of each hidden neuron is used as the input of the next layer of neurons, and after being processed by the multiple layers of hidden neurons in the hidden layer, an initial feature set is obtained; The information gain method is used to evaluate the feature importance of the initial feature set, and the importance of each feature in the initial feature set to the sample concrete sample image data is generated as a feature importance evaluation result; wherein, for each feature in the initial feature set, the difference in information entropy of the sample concrete sample image data when the feature is included and when the feature is not included is calculated, and the difference is the information gain of the feature, and the larger the information gain, the more important the feature is for distinguishing different sample concrete sample image data; The initial feature set is screened according to the result of the feature importance evaluation, and the screened features are combined using a rule-based feature combination method to generate a combined feature set; The combined feature set is input again into the hidden layer of the target network module unit for deep mining. During the deep mining process, residual connections are used between the hidden layers, wherein the residual connections allow the information of the combined feature set to be directly transferred from the previous hidden layer to the subsequent hidden layer. After deep mining processing of multiple hidden layers, a deep mining feature set is obtained; Inputting the deep mining feature set into the output layer of the target network module unit, and linearly combining the deep mining feature set through the weights and bias parameters of the neurons in the output layer, wherein the output of each neuron in the output layer represents the possibility of the feature category corresponding to the neuron under the currently input deep mining feature set; The Softmax classification algorithm is used to classify the outputs of the neurons in the output layer, so as to convert the output value of each neuron in the output layer into a probability distribution between various categories, thereby generating the first feature detection result.

5. The intelligent concrete detection method based on machine learning according to any one of claims 1 to 4, characterized in that: The step of acquiring a second sample concrete sample image data sequence, traversing each of the multiple fusion network module units, and performing migration parameter learning on the multiple second interaction network module units in the fusion network module unit based on the second sample concrete sample image data sequence and the multiple first interaction network module units in the traversed fusion network module units, comprises: Acquire a second sample concrete sample image data sequence, load the second sample concrete sample image data sequence into a first network module unit among a plurality of first interactive network module units in the traversed fusion network module units, and generate a first transfer learning result generated by each network module unit in the plurality of first interactive network module units; Based on the second example concrete sample image data sequence and the first transfer learning result generated by each of the plurality of first interactive network module units, transfer parameter learning is performed on each of the plurality of second interactive network module units.

6. The intelligent concrete detection method based on machine learning according to claim 5 is characterized in that: The method of performing migration parameter learning on each of the plurality of second interactive network module units based on the second sample concrete sample image data sequence and the first migration learning result generated by each of the plurality of first interactive network module units comprises: The second sample concrete sample image data sequence is loaded into the first network module unit of the plurality of second interactive network module units in the traversed fusion network module units, and the first mapping relationship array and the first transfer learning result of each network module unit in the plurality of first interactive network module units are loaded into each network module unit in the plurality of second interactive network module units respectively, so as to generate the second transfer learning result generated by each network module unit in the plurality of second interactive network module units; Performing a matching operation on the second migration learning result of the last network module unit in the plurality of second interactive network module units and the second mapping relationship array to obtain a second matching result; Determine a first transfer learning result of a last network module unit in the plurality of first interactive network module units and a transfer learning error between the second matching result; If the transfer learning error meets the transfer learning convergence requirement, the transfer parameter learning of the traversed fusion network module unit is completed.

7. The intelligent concrete detection method based on machine learning according to any one of claims 1 to 4, characterized in that: The priori machine learning network and the basic concrete intelligent detection network are decomposed into a plurality of network module units, including: According to the network layered architecture, at least every two adjacent network modules of the priori machine learning network are respectively decomposed into a network module unit to generate a plurality of network module units corresponding to the priori machine learning network, and each network module of the foundation concrete intelligent detection network is respectively decomposed into a network module unit to generate a plurality of network module units corresponding to the foundation concrete intelligent detection network; The method also includes: constructing the priori machine learning network and the corresponding relationship between each network module unit in the foundation concrete intelligent detection network according to the network layered architecture.

8. The intelligent concrete detection method based on machine learning according to any one of claims 1 to 4, characterized in that: The method further comprises: Obtaining neuron weight information of the target concrete intelligent detection network that has completed module parameter learning; Performing a mapping conversion operation on the neuron weight information of the target concrete intelligent detection network, mapping the neuron weight information of the target concrete intelligent detection network from a continuous numerical domain to a discrete numerical interval, and obtaining a mapped and converted target concrete intelligent detection network; Based on the target concrete intelligent detection network after the mapping conversion, concrete intelligent detection is performed on any input concrete sample image data.

9. The intelligent concrete detection method based on machine learning according to claim 8 is characterized in that: The step of performing a mapping conversion operation on the neuron weight information of the target concrete intelligent detection network, mapping the neuron weight information of the target concrete intelligent detection network from a continuous numerical domain to a discrete numerical interval, and obtaining the mapped and converted target concrete intelligent detection network includes: Extracting neuron weight information from the target concrete intelligent detection network, wherein the target concrete intelligent detection network comprises a plurality of network layers, each network layer is composed of a plurality of neurons, and the connections between neurons have corresponding weight information, and the weight information reflects the degree of recognition of different features and data relationships by the target concrete intelligent detection network during the learning process; Performing structural analysis on the extracted neuron weight information to determine distribution information of the neuron weight information in the target concrete intelligent detection network, wherein the distribution information includes the difference in the size range of weights between different network layers and the positive and negative distribution of weights; and performing statistical analysis on the neuron weight information to determine the statistics of the mean, median, and standard deviation of the neuron weight information, wherein the statistics and the distribution information constitute the analysis result of the neuron weight information; According to the analysis result, determining an initial range of a discrete numerical interval, wherein the range of the discrete numerical interval is determined based on a multiple of a standard deviation of the neuron weight information, with the mean value of the neuron weight information as the center; According to the task detection index defined for the target concrete intelligent detection network and the boundary conditions of the discrete numerical interval, the initial range of the discrete numerical interval is fine-tuned to generate a final discrete numerical interval; According to the analysis result and the discrete numerical interval, a mapping strategy is determined, and according to the determined mapping strategy, each neuron weight information in the target concrete intelligent detection network is mapped and converted, wherein if the mapping strategy is a linear mapping strategy, the corresponding value of each neuron weight information in the discrete numerical interval is calculated according to the proportional relationship between each neuron weight information and the continuous numerical domain range; or, under a nonlinear mapping strategy, according to the segmented range where each neuron weight information is located, the conversion is performed according to the corresponding mapping rules, and in the mapping conversion process, the mapping relationship between the original value of each neuron weight information and the mapped conversion value is recorded; According to the neuron weight information after mapping conversion, the structure of the target concrete intelligent detection network is adjusted, and the function of the adjusted target concrete intelligent detection network is verified using a predefined verification data set. According to the function verification result, the target concrete intelligent detection network is fine-tuned to finally obtain the target concrete intelligent detection network after mapping conversion.

10. A machine learning-based intelligent concrete detection system, characterized in that: The machine learning-based intelligent concrete detection system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the machine learning-based intelligent concrete detection method according to any one of claims 1 to 9.

Images