Crane structure crack intelligent evaluation method based on adaptive neural network

Abstract

The invention discloses a crane structure crack intelligent evaluation method based on an adaptive neural network, and relates to the technical field of crack evaluation. Comprising the following steps: dividing different parts of the crane; acquiring a plurality of pieces of historical data information of a source domain, and collecting a plurality of pieces of sample information of different areas of the crane; the method comprises the following steps: establishing an original domain prediction model through historical data information of an original domain, freezing a universal layer of the original domain prediction model, adding region information as input information, and performing secondary training on the original domain prediction model through multiple pieces of sample information of different regions; substituting to-be-evaluated information into the preliminary prediction model, and calculating preliminary evaluation results of cracks in different areas; and calculating a verification factor and verifying the preliminary evaluation result. According to the method, crack evaluation samples in other fields are introduced into the crack evaluation of the crane, so that the characteristics are migrated, the number of the samples is increased, the trained model is more representative, and the crack evaluation of the crane can be more accurately carried out.

Description

Technical Field

[0001] This invention relates to the field of crack assessment technology, specifically to an intelligent crack assessment method for crane structures based on adaptive neural networks. Background Technology

[0002] During long-term service, cranes are prone to fatigue cracks in their metal structures due to frequent exposure to heavy loads, complex stresses, and environmental corrosion. If these cracks are not identified and repaired in time, they may propagate and lead to fracture, resulting in equipment damage or even personal injury. Therefore, crack assessment of cranes is of paramount importance: by assessing cracks and taking appropriate repair measures, catastrophic accidents can be effectively prevented, equipment lifespan extended, maintenance costs reduced, and operational safety ensured, meeting the core requirements of high reliability and safety in industrial production.

[0003] Crack assessment in cranes requires specific assessment methods, such as the fatigue crack propagation life prediction method for metal structures of metallurgical cranes disclosed in patent publication number CN111832125A, which includes the following steps: (1) constructing a state-space assessment model based on the Paris formula; (2) constructing an extended Kalman filter model based on objective data to obtain the predicted crack propagation life f1; (3) constructing a particle filter model based on objective data to obtain the predicted crack propagation life f2; (4) combining extended Kalman filter and particle filter to obtain the final predicted crack propagation life f. The prediction method of this invention takes into account the physical characteristics and life change characteristics of the component itself, and can also adjust the physical model according to different health conditions. It can achieve the advantages of health prediction in a single algorithm, and also reduce the error caused by a single algorithm.

[0004] Crane crack assessment can enable appropriate maintenance measures to be taken in advance, thus improving production safety. However, the assessment requires training a large number of samples to predict the model. The methods mentioned above and existing assessment methods only use the crack data of the crane itself as samples. The sample size is insufficient, making it difficult to train an accurate assessment model, resulting in low reference value of the assessment results. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent assessment method for cracks in crane structures based on adaptive neural networks, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent assessment method for cracks in crane structures based on adaptive neural networks, comprising: Area division: Divide the crane into different parts and number them; Data collection: Acquire multiple historical data from the source domain, including crack information, feature information, and evaluation information, and collect multiple sample information from different areas of the crane, including area information, crack information, feature information, and evaluation information; Training the original domain model and transferring it: Using crack information and feature information as input and evaluation information as output, the original domain prediction model is established through historical data information of the original domain. The general layer of the original domain prediction model is frozen, and regional information is added as input information. The original domain prediction model is trained a second time through multiple sample information from different regions to obtain the preliminary prediction model corresponding to each region. Preliminary assessment: Substitute the information to be assessed into the preliminary prediction model to calculate the preliminary assessment results of cracks in different regions; Calculate the validation factor: Calculate the validation factor based on sample information from different regions and according to the validation factor calculation method; Verification of evaluation results: The preliminary evaluation results are tested and verified by combining the evaluation correction method with the verification factor.

[0007] Preferably, the training of the original domain model and transfer specifically includes: The source domain prediction model is selected as a CNN model; Based on historical data, MSE is used as the loss function, and the CNN model is trained using the gradient descent algorithm. Freeze the convolutional layers of the CNN model, replace the last fully connected layer with a new fully connected layer. The input dimension of the new fully connected layer is determined based on the output of the previous layer, and the output dimension is 1. At the same time, add input channels for region information to the input layer. The sample data information in each region is randomly divided into a training set and a validation set according to a preset ratio, with each region as the unit. Using the training set of one region, MSE is used as the loss function, and the CNN model is trained using the gradient descent algorithm to obtain a preliminary prediction model for the target region.

[0008] Preferably, the verification factor calculation method includes: Obtain the sample information from the last test, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The validation factor is calculated according to the validation factor calculation formula, specifically as follows: ; in Indicates the verification factor. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k.

[0009] Preferably, the verification factor calculation method includes: Obtain information from multiple samples, select one sample, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The preliminary validation factor is calculated according to the formula: ; in A preliminary validation factor representing the target sample information. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k; Calculate the preliminary validation factor for each sample; Each region assessment within each sample is normalized, and then a preliminary composite score is calculated for each sample using the following formula: ; in This represents the initial composite score of the sample. This represents the total number of regions in the sample. This represents the normalized value of the assessment for the nth region; The preliminary composite score is mapped to the range of the preliminary validation factors to obtain a comprehensive equal score, specifically as follows: ; in This represents the overall score. This represents the initial composite score of the sample. This represents the minimum value of the initial verification factor. This represents the maximum value of the initial verification factor. This represents the maximum value of the preliminary overall score. This represents the maximum value of the preliminary overall score; Calculate the average of all composite scores as the validation factor.

[0010] Preferably, the verification factor calculation method includes: Methods for calculating validation factors include: Obtain information from multiple samples, select one sample, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The preliminary validation factor is calculated according to the formula: ; in A preliminary validation factor representing the target sample information. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k; Calculate the preliminary validation factor for each sample; The total number of regions in each sample is counted, and the preliminary validation factors are weighted according to the proportion of the number of regions and then summed to obtain the validation factors.

[0011] Preferably, the verification factor calculation method includes: Methods for calculating validation factors include: Obtain information from multiple samples, select one sample, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The preliminary validation factor is calculated according to the formula: ; in A preliminary validation factor representing the target sample information. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k; Calculate the preliminary validation factor for each sample; Count the total number of regions in each sample and calculate the proportion of regions between each sample. Obtain the crack length in different regions, and count the total crack length in all regions of the sample information. Then calculate the proportion of the total crack length between each sample information. Calculate the average of the proportion of the number of regions and the proportion of the total crack length in each sample information. Then, sum the values ​​of the average values ​​to obtain the initial verification factor.

[0012] Preferably, the evaluation and correction method includes: Based on the preliminary assessment results, areas with the same preliminary assessment results are grouped together, and the proportion of each group is calculated. The test factor is calculated according to the formula, specifically as follows: ; in Indicates the test factor. This indicates the total number of assessment grades in the preliminary assessment results. This indicates the percentage of regions within the group with an assessment level of m; If the change in the test factor is within 13.6% compared to the validation factor, the preliminary evaluation result is verified and becomes the final evaluation result. If the variation of the test factor from the validation factor exceeds the validation threshold, the preliminary assessment result fails validation. The test factor is then recalculated for different regions until validation is passed.

[0013] Preferably, the method for calculating the verification threshold includes: The total number of assessments prior to this one will be tallied. The verification threshold is calculated using the following formula: ; in This represents the maximum value of the validation factor. This represents the minimum value of the validation factor. This indicates the verification threshold.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By introducing crack assessment samples from other fields into the crack assessment of cranes, feature transfer is achieved, the number of samples is expanded, and the trained model becomes more representative, enabling it to assess crane cracks more accurately.

[0015] At the same time, different parts of a crane experience different stresses and have different mechanical structures. They are also subject to various external factors during operation. Dividing the different parts before evaluation can improve the accuracy of the assessment.

[0016] Furthermore, the assessment results can be verified in conjunction with previous assessment data, and the assessment results of different regions can be screened to identify those that fail the verification, thereby further improving the accuracy of the assessment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent evaluation method of the present invention; Figure 2 This is a flowchart illustrating the verification factor calculation method in this invention; Figure 3 This is a schematic diagram illustrating the crack level classification of cranes in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: The number of samples for crane crack assessment is relatively small, making it difficult to guarantee the accuracy and universality of the assessment model. When training the model, transforming crack assessment samples from other fields to achieve feature transfer and expand the number of samples can improve the accuracy of the assessment model.

[0020] like Figures 1-3 As shown, the present invention provides a technical solution: an intelligent assessment method for cracks in crane structures based on adaptive neural networks, comprising: Regional division: The crane is divided into different parts and numbered. Different parts of the crane are subjected to different stresses, have different mechanical structures, and are subject to different external factors during operation. Therefore, dividing the crane into different parts before evaluation can improve the accuracy of the evaluation. Data collection: Acquire multiple historical data from the source domain, including crack information, feature information, and evaluation information, and collect multiple sample information from different areas of the crane, including area information, crack information, feature information, and evaluation information; Training the original domain model and transferring it: Taking crack information and feature information as input and evaluation information as output, an original domain prediction model is established through historical data information of the original domain. The general layer of the original domain prediction model is frozen, and regional information is added as input information. The original domain prediction model is trained a second time through multiple sample information from different regions to obtain the preliminary prediction model corresponding to each region. The training model adopts the CNN neural network model. This model is an existing technology. Due to its unique structural design, it can efficiently process image data, automatically extract hierarchical features, and has strong generalization ability, making it an ideal tool for image recognition and evaluation. Preliminary assessment: Substitute the information to be assessed into the preliminary prediction model to calculate the preliminary assessment results of cracks in different regions. The preliminary assessment results are used to evaluate the cracks in different regions, which can reflect the crack situation in different regions. Calculate the validation factor: Calculate the validation factor based on information from multiple samples in different regions and according to the validation factor calculation method; Verification of evaluation results: The preliminary evaluation results are detected and verified by combining the evaluation correction method with the verification factor. Different parts of the crane need to work together during operation, so there is a certain potential correlation between the cracks generated in each part. The evaluation results are verified by the potential correlation between cracks in different regions.

[0021] It is important to note that when collecting crack information, it is necessary to collect data such as crack length, width, and orientation angle. Characteristic information includes crack type (network or stripe, etc., requiring digital coding). The more types of information collected, the higher the accuracy of the assessment. The assessment information is collected by technical personnel and categorized into different levels, such as no maintenance required, online maintenance, downtime maintenance, repair, and component replacement. Different levels need to be digitally coded; for example, crack severity should correspond to the aforementioned maintenance levels, which can be coded as 0, 1, 2, 3, and 4 respectively. Figure 3 As shown, cracks are classified (level 0 is not demonstrated, indicating that the crack length and other data are below level 1). The specific classification criteria are based on actual needs. The more detailed the classification, the more accurate the evaluation.

[0022] Training the original domain model and transferring the knowledge specifically includes: The source domain prediction model is selected as a CNN model; Based on historical data, MSE is used as the loss function, and the CNN model is trained using the gradient descent algorithm. Freeze the convolutional layers of the CNN model, replace the last fully connected layer with a new fully connected layer. The input dimension of the new fully connected layer is determined based on the output of the previous layer, and the output dimension is 1. At the same time, add input channels for region information to the input layer. The sample data information in each region is randomly divided into a training set and a validation set according to a preset ratio, with each region as the unit. Using the training set of one region, MSE is used as the loss function, and the CNN model is trained using the gradient descent algorithm to obtain a preliminary prediction model for the target region.

[0023] It should be noted that, for ease of understanding, the following simulated data is used: Network structure: The first convolutional layer has 64 filters, with a kernel size of 5×5 and ReLU activation.

[0024] Pooling layer 1: max pooling, kernel size 2×2, step size 2.

[0025] The second convolutional layer has 128 filters, a 3×3 kernel size, and ReLU activation.

[0026] Second pooling layer: max pooling, kernel size 2×2, step size 2.

[0027] Fully connected layer: 256 neurons, ReLU activated.

[0028] Output layer: 1 neuron (regression task), no activation function (directly outputs the predicted value).

[0029] Training parameters: Optimizer: Adam, initial learning rate 0.0005, decays to 0.9x every 30 epochs.

[0030] Loss function: Huber loss.

[0031] Batch size: 64.

[0032] Training epochs: 150.

[0033] Migration phase operations: The source domain data is divided into training and validation sets in a 7:3 ratio.

[0034] Target domain transfer learning phase: Network adjustments: Freeze layer: Freeze only the first convolutional layer (preserving low-level features such as edge detection), and unfreeze the second convolutional layer to adapt to the new data distribution.

[0035] Add a new input channel: information in the splicing area of ​​the input layer is merged with the original RGB channels to form a multi-channel input.

[0036] Replacement layer: Remove the original fully connected layer and add two new fully connected network layers: Layer 1: 128 neurons, ReLU activated.

[0037] Layer 2 (output layer): 1 neuron, linearly activated.

[0038] Training optimization: Optimizer: Adam, learning rate 0.00005.

[0039] Loss function: Weighted Huber loss, with weights dynamically allocated based on the standard deviation of the output of samples in the target domain.

[0040] Early Stopping: Stop when the validation set loss does not decrease for 8 consecutive epochs.

[0041] Data partitioning: The target domain data is divided into training and validation sets in a 7:3 ratio.

[0042] Through the above training, CNN models (preliminary prediction models) for different regions can be obtained. (The training and transfer of CNN models are existing technologies. The specific principles and processes will not be elaborated here. The above parameter settings are just one way of adjusting parameters. There are no actual restrictions on the parameter adjustment methods. Other parameter adjustment methods can also be used.)

[0043] CNN models, through activation functions such as ReLU, enable the model to fit arbitrarily complex nonlinear mappings, making them suitable for modeling complex patterns in data. They are easier and more accurate than multiple linear regression models when dealing with large volumes of data and nonlinear data. In addition, other data models can be used when selecting a model, and this embodiment does not restrict the types of models that can be selected.

[0044] like Figure 2 As shown, the verification factor calculation method includes: Obtain the sample information from the last test, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The validation factor is calculated according to the validation factor calculation formula, specifically as follows: ; in Indicates the verification factor. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k.

[0045] It should be noted that, for ease of understanding, the following simulated data is used: Assuming there are 5 areas in total, the area assessment results are as follows (the data number in parentheses is the assessment level): Online maintenance (1), Online maintenance (1), Downtime maintenance (2), Downtime maintenance (2), Repair (3); It can be seen that there are three assessment levels in total, and the regional proportions of the three assessment levels are 0.4, 0.4 and 0.2 respectively; According to the formula The verification factor was calculated as follows: ≈−(0.4×(−1.3219)+0.4×(−1.3219)+0.2×(−2.3219))=1.522.

[0046] The evaluation and correction methods include: Based on the preliminary assessment results, areas with the same preliminary assessment results are grouped together, and the proportion of each group is calculated. The test factor is calculated according to the formula, specifically as follows: ; in Indicates the test factor. This indicates the total number of assessment grades in the preliminary assessment results. This indicates the percentage of regions within the group with an assessment level of m; If the change in the test factor is within 13.6% compared to the validation factor, the preliminary evaluation result is verified and becomes the final evaluation result. If the variation of the test factor from the validation factor exceeds 13.6%, the preliminary assessment result fails validation. The test factor is then recalculated after removing assessment results from different regions until validation is successful.

[0047] It should be noted that, for ease of understanding, the following simulated data is used: Assuming there are a total of 5 areas, the preliminary assessment results are as follows (the data number in parentheses is the assessment level): Online maintenance (1), Online maintenance (3), Downtime maintenance (1), Downtime maintenance (1), Repair (2); And assume the verification threshold is 5%.

[0048] It can be seen that there are three assessment levels in total, and the regional proportions of the three assessment levels are 0.5, 0.25 and 0.25 respectively; According to the formula The test factor was calculated as follows: ≈−(0.5×(−1)+0.25×(−2)+0.25×(−2))=0.5+0.5+0.5=1.5.

[0049] Compared with the validation factor of 1.522, the change is approximately 1.4% < 5% (1.522-1.5) ÷ 1.522. Therefore, the preliminary evaluation result is validated and is the final evaluation result.

[0050] In practical use, if the verification fails, the preliminary assessment results of each region are calculated in turn, and the test factor is calculated again. The test factor with the smallest change compared to the verification factor is selected and the results are calculated in turn. If the verification passes, the corresponding region to be removed is assessed incorrectly, and the remaining regions are assessed correctly. If the verification fails, the number of regions to be removed is increased and the calculation continues until the verification passes or all regions are removed.

[0051] The methods for calculating the verification threshold include: The total number of assessments prior to this one will be tallied. The verification threshold is calculated using the following formula: ; in This represents the maximum value of the validation factor. This represents the minimum value of the validation factor. This indicates the verification threshold.

[0052] It should be noted that, for ease of understanding, the following simulated data is used: Assuming that three evaluations have been conducted previously, and the validation factors for the first three evaluations were 1.6, 1.4, and 1.3 respectively, the validation threshold can be calculated to be 7.3% according to the formula.

[0053] Example 2: As the number of evaluations increases, the amount of sample information also increases. In Example 1, the validation factor is calculated based solely on the previous evaluation results, and then the preliminary evaluation results are corrected. This makes it difficult to guarantee the universality of the validation factor, which may lead to inaccurate calculation of the validation factor and thus affect the evaluation results. Therefore, this example provides another method for calculating the validation factor to improve the accuracy of the calculation.

[0054] like Figure 2 As shown, the verification factor calculation method includes: The method for calculating the verification factor includes: Obtain information from multiple samples, select one sample, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The validation factor is calculated according to the validation factor calculation formula, specifically as follows: ; in A preliminary validation factor representing the target sample information. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k; Calculate the preliminary validation factor for each sample; Each region assessment within each sample is normalized, and then a preliminary composite score is calculated for each sample using the following formula: ; in This represents the initial composite score of the sample. This represents the total number of regions in the sample. This represents the normalized value of the assessment for the nth region; The preliminary composite score is mapped to the range of the preliminary validation factors to obtain a comprehensive equal score, specifically as follows: ; in This represents the overall score. This represents the initial composite score of the sample. This represents the minimum value of the initial verification factor. This represents the maximum value of the initial verification factor. This represents the maximum value of the preliminary overall score. This represents the maximum value of the preliminary overall score; Calculate the average of all composite scores as the validation factor.

[0055] It should be noted that, for ease of understanding, the following simulated data is used: The preliminary validation factor for each sample was calculated in the manner described in Example 1 (the calculation process is the same as described above and will not be demonstrated here). Each sample has three regions, as shown in Table 1 below: Table 1: Sample Information Table

[0056] The normalized data are shown in Table 2 (normalization is based on existing technology; the calculation process will not be demonstrated here, and the results are given directly): Table 2: Normalized Sample Information Table

[0057] Then according to the formula The preliminary comprehensive scores of the three sample information can be calculated to be 0.5, 1, and 0.583, respectively.

[0058] according to Mapping the preliminary composite score to the range of the preliminary validation factor, we obtain composite scores of 1, 1.5, and 1.083 for the three samples, with an average of approximately 1.19. Therefore, the validation factor is 1.19. By calculating the validation factor using information from multiple samples, compared to Example 1, the accuracy of the validation factor can be further improved, the impact of errors from a single evaluation on the validation factor can be reduced, and the effect of adjusting the validation factor can be improved.

[0059] Example 3: In Example 2, the number of regions evaluated is the same for multiple evaluations. However, for parts of the crane that are subjected to less load, it is not necessary to evaluate them every time. This can improve the efficiency of the evaluation and reduce the amount of calculation. In this case, Example 2 is difficult to perform a proper normalization operation, and the accuracy of the calculated verification factor is greatly affected. Therefore, this example proposes another method for calculating the verification factor in the extrusion of Example 2 to improve the accuracy of the verification factor in this special case.

[0060] like Figure 2 As shown, the verification factor calculation method includes: Methods for calculating validation factors include: Obtain information from multiple samples, select one sample, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The preliminary validation factor is calculated according to the formula: ; in A preliminary validation factor representing the target sample information. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k; Calculate the preliminary validation factor for each sample; The total number of regions in each sample is counted, and the preliminary validation factors are weighted according to the proportion of the number of regions and then summed to obtain the validation factors.

[0061] It should be noted that, for ease of understanding, the following simulated data is used: Assuming that the preliminary validation factor for each sample is calculated in the manner described in Example 1 (the calculation process is the same as above and will not be demonstrated here), and that the number of regions for each sample is different (leaving it blank indicates that the region does not exist), the following table 3 is obtained: Table 3: Information Table of Unequal Samples

[0062] The number of regions in samples 4, 5, and 6 are 2, 3, and 5, respectively, so the proportions of regions are 0.2, 0.3, and 0.5.

[0063] The validation factor is calculated as 0.2×1.2+0.3×1.3+0.5×1.4=0.24+0.39+0.7=1.33.

[0064] Increasing the contribution of samples with a large number of regions to the validation factor and suppressing the contribution of samples with a small number of regions to the validation factor can reduce the impact of differential data generated when the number of regions is small on the validation factor, thereby improving the accuracy of the validation factor when the number of regions in multiple samples is inconsistent.

[0065] Example 4: In Example 3, the initial verification factor is weighted by the proportion of the number of regions, thereby changing the contribution of each sample to the verification factor. However, different regions have different sizes and loads, and different usage conditions will cause different crack conditions. Not only does the difference in the number of regions have different effects on the verification factor, but the difference in factors such as the size, load and material of the regions will also cause different crack lengths, which will have different effects on the verification factor. Based on this, this example provides another method for calculating the verification factor based on Example 3, in order to improve the accuracy of the verification factor.

[0066] like Figure 2 As shown, the method for calculating the verification factor includes: Methods for calculating validation factors include: Obtain information from multiple samples, select one sample, and based on the evaluation results, group the regions with different evaluation information into a group and calculate the proportion of each group. The preliminary validation factor is calculated according to the formula: ; in A preliminary validation factor representing the target sample information. This indicates the total number of assessment levels in the assessment information. This indicates the percentage of regions within the group with an assessment level of k; Calculate the preliminary validation factor for each sample; Count the total number of regions in each sample and calculate the proportion of regions between each sample. Obtain the crack length in different regions, and calculate the total crack length of all regions in the sample information. Then calculate the proportion of the total crack length between each sample information. Calculate the average of the proportion of the number of regions and the proportion of the total crack length in each sample information. Then, sum the values ​​of the average values ​​to obtain the initial verification factor.

[0067] It should be noted that, for ease of understanding, the following simulated data is used: Continuing with the data from Example 3, the total crack lengths of multiple sample data are shown in Table 4 below: Table 4: Crack Length Table

[0068] According to the calculation results in Example 3, the proportions of the number of regions in Sample 4, Sample 5 and Sample 6 are 0.2, 0.3 and 0.5, respectively.

[0069] Based on Table 4, the percentage of total crack length can be calculated: the percentages of total crack length for samples 4, 5, and 6 are 0.2, 0.5, and 0.3, respectively.

[0070] Therefore, the average proportions of samples 4, 5, and 6 are 0.15, 0.4, and 0.4, respectively.

[0071] The validation factor is calculated as 0.2×1.2+0.4×1.3+0.4×1.4=0.24+0.52+0.56=1.32.

[0072] Compared to simply calculating the number of regions in Example 3, this example combines the number of regions and the total crack length to calculate the contribution of sample information to the verification factor. This approach can take into account different sizes and loads in different regions, further improving the accuracy of the verification factor.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for intelligent evaluation of cracks in crane structures based on adaptive neural networks, characterized by: The method comprises the following steps: Region division: different parts of the crane are divided and numbered; Data collection: a plurality of historical data information of a source domain is obtained, including crack information, feature information and evaluation information, and a plurality of sample information of different regions of the crane is collected, including region information, crack information, feature information and evaluation information; Training of the original domain model and migration: an original domain prediction model is established by taking the crack information and the feature information as input and the evaluation information as output, the general layer of the original domain prediction model is frozen, the region information is added as input information, and the original domain prediction model is trained again by using the plurality of sample information of different regions to obtain a preliminary prediction model corresponding to each region; Preliminary evaluation: the to-be-evaluated information is substituted into the preliminary prediction model to calculate the preliminary evaluation result of the crack in each region; Verification factor calculation: the verification factor is calculated according to the sample information of different regions and according to a verification factor calculation method; Verification of the evaluation result: the preliminary evaluation result is detected and verified by using an evaluation correction method and the verification factor.

2. The intelligent crane structure crack evaluation method based on adaptive neural network according to claim 1, characterized in that: The training of the original domain model and migration specifically comprises the following steps: The source domain prediction model is selected as a CNN model; The CNN model is trained by using a gradient descent algorithm according to the historical data information and taking the MSE as a loss function; The convolution layer of the CNN model is frozen, the last fully connected layer is replaced by a new fully connected layer, the input dimension of the new fully connected layer is determined according to the output of the previous layer, the output dimension is 1, and the input channel of the region information is added to the input layer; The sample data information in each region is randomly divided into a training set and a verification set according to a preset proportion; The CNN model is trained by using a gradient descent algorithm and taking the MSE as a loss function through the training set of one region to obtain a preliminary prediction model of the target region. 3.The crane structure crack intelligent evaluation method based on adaptive neural network according to claim 1, characterized in that: The verification factor calculation method comprises the following steps: The sample information detected last time is obtained, different regions of different evaluation information are divided into a group according to the evaluation result, and the proportion of each group of regions is calculated; The verification factor is calculated according to a verification factor calculation formula, specifically as follows: ； wherein represents a verification factor, represents the total number of evaluation levels in the evaluation information, represents the area proportion of the region in the group of evaluation level k.

4. The intelligent crane structure crack evaluation method based on adaptive neural network according to claim 1, characterized in that: The verification factor calculation method comprises the following steps: A plurality of sample information is obtained, one sample information is selected, different regions of different evaluation information are divided into a group according to the evaluation result, and the proportion of each group of regions is calculated; A preliminary verification factor is calculated according to a preliminary verification factor calculation formula, specifically as follows: ； wherein a preliminary verification factor representing target sample information, a total number of evaluation grades in the evaluation information, a proportion of the area in the group of evaluation grades k. The preliminary verification factor of each sample information is calculated; Each region evaluation in each sample information is normalized, and then a preliminary comprehensive score of each sample information is calculated according to the formula, specifically as follows: ； wherein represents the preliminary composite score of the sample, represents the total number of regions in the sample, represents the normalized value of the nth region assessment; The preliminary comprehensive score is mapped to the range of the preliminary verification factor to obtain a comprehensive equal division, specifically as follows: ； wherein represents the overall score, represents the preliminary overall score of the sample, represents the minimum value of the preliminary validation factor, represents the maximum value of the preliminary validation factor, represents the maximum value of the preliminary overall score, represents the maximum value of the preliminary overall score; The average value of all comprehensive scores is calculated as the verification factor.

5. The intelligent crane structure crack evaluation method based on adaptive neural network according to claim 1, characterized in that: The verification factor calculation method comprises the following steps: The verification factor calculation method comprises the following steps: A plurality of sample information is obtained, one sample information is selected, different regions of different evaluation information are divided into a group according to the evaluation result, and the proportion of each group of regions is calculated; A preliminary verification factor is calculated according to a preliminary verification factor calculation formula, specifically as follows: ； wherein a preliminary verification factor representing target sample information, a total number of evaluation grades in the evaluation information, a proportion of the area in the group of evaluation grades k; The preliminary verification factor of each sample information is calculated; The total number of regions in each sample information is counted, and the preliminary verification factors are weighted and summed in proportion to the number of regions to obtain the verification factor.

6. The intelligent crane structure crack evaluation method based on adaptive neural network according to claim 1, characterized in that: The verification factor calculation method comprises: The verification factor calculation method comprises: A plurality of sample information is obtained, one sample information is selected, regions of different evaluation information are divided into a group according to the evaluation result, and the proportion of each group of regions is calculated; According to the preliminary verification factor calculation formula, the preliminary verification factor is calculated, specifically: ； wherein represents a preliminary verification factor of target sample information, represents the total number of evaluation grades in the evaluation information, represents the area proportion in the group of evaluation grade k; The preliminary verification factor of each sample information is calculated; The total number of regions in each sample information is counted, and the proportion of the number of regions between each sample information is calculated; The crack length of different regions is obtained, and the total length of all cracks in the sample information is counted, and the proportion of the total length of cracks between each sample information is calculated; The average value of the proportion of the number of regions and the proportion of the total length of cracks in each sample information is calculated, the preliminary verification factor is weighted and summed by the average value, and the verification factor is obtained.

7. The method according to claim 1, characterized in that: The evaluation correction method comprises: According to the preliminary evaluation result, the regions with the same preliminary evaluation result are divided into a group, and the proportion of each group of regions is calculated; According to the test factor calculation formula, the test factor is calculated, specifically: ； wherein represents the test factor, represents the total number of evaluation grades in the preliminary evaluation result, represents the area proportion in the group with evaluation grade m; The test factor is compared with the verification factor, and if the change amplitude of the test factor is within 13.6% of the verification factor, the preliminary evaluation result is verified, that is, the final evaluation result; If the change amplitude of the test factor exceeds the verification threshold, the preliminary evaluation result is not verified, and the test factor is repeatedly calculated under the condition that the preliminary evaluation result of different regions is removed until the verification is passed. 8.The crane structure crack intelligent evaluation method based on adaptive neural network according to claim 1, characterized in that: The calculation method of the verification threshold comprises: The total number of evaluations before this time is counted; The verification threshold is calculated by the threshold calculation formula, specifically: ； wherein represents the maximum value of the validation factor, represents the minimum value of the validation factor, represents the validation threshold.

Images