A container damage detection method and system based on deep learning

Abstract

The present application relates to the technical field of container damage detection, and in particular to a method and system for container damage detection based on deep learning. The method includes: establishing multiple container categories, and establishing an image analysis model based on all container categories; setting an image acquisition strategy for the container to be detected, and obtaining an image data packet of the container to be detected according to the image acquisition strategy; generating a preprocessing result of the image data packet according to the image analysis model; generating a first-level detection strategy for the container to be detected according to the preprocessing result, and generating a damage risk value for the container to be detected; constructing multiple container categories based on different container equipment parameters, and using deep learning technology to construct sub-image models of each container category, so as to adapt to different types of container damage in the actual production process of different ports, and perform preliminary damage detection by collecting image data of the container to be detected, thereby improving the efficiency of container damage detection and reducing the detection and operation and maintenance costs.

Description

Technical Field

[0001] The present application relates to the technical field of container damage detection, and in particular to a container damage detection method and system based on deep learning. Background Art

[0002] With the rapid development of global trade, port throughput continues to grow, and the requirements for port efficiency are also increasing. As a key unit for port loading and unloading operations, the degree of damage of containers has received increasing attention. Some containers have suffered damage such as breakage, holes, and rust as they have been used for a long time.

[0003] The degree of damage to containers is related to the safety of cargo and container ships, so it is very important to study efficient detection methods for container damage detection. Traditional container damage detection methods, such as lidar detection, often have high requirements for gate hardware equipment and slow detection speed, increasing container management and operating costs. Summary of the invention

[0004] The purpose of this application is: to solve the above-mentioned technical problems, this application provides a container damage detection method and system based on deep learning, aiming to improve the efficiency and accuracy of container damage detection and reduce the overall operation and maintenance cost of containers.

[0005] In some embodiments of the present application, multiple container categories are constructed based on different container equipment parameters, and sub-image models of each container category are constructed using deep learning technology to adapt to different types of container damage in the actual production process of different ports. Preliminary damage detection is performed by collecting image data of the container to be inspected, thereby improving the efficiency of container damage detection.

[0006] In some embodiments of the present application, by analyzing the image data packets of the container to be inspected, a preliminary judgment is made on the container to be inspected, and a secondary inspection is performed on the risk sub-area therein. During the secondary inspection, multi-source data (laser data, infrared data, ultrasonic data, etc.) is collected for auxiliary judgment, so as to timely warn and diagnose potential residual risks, improve the diagnostic accuracy of container residues, and reduce the overall inspection and operation and maintenance costs.

[0007] In some embodiments of the present application, a container damage detection method based on deep learning is provided, comprising:

[0008] Establish multiple container categories and build image analysis models based on all container categories;

[0009] Setting an image acquisition strategy for the container to be inspected, and acquiring an image data packet of the container to be inspected according to the image acquisition strategy;

[0010] Generate preprocessing results of image data packets according to the image analysis model;

[0011] Generate a primary inspection strategy for the container to be inspected according to the preprocessing result, and generate a damage risk value for the container to be inspected;

[0012] Among them, when setting multiple container categories, including:

[0013] Establish container category sequence A, A=(a 1 ,a 2 …a i …a n ), where a i is the i-th container category; n is the number of container categories.

[0014] In some embodiments of the present application, when an image analysis model is established according to all container categories, it includes:

[0015] According to the container type sequence A, set a i is the target container category;

[0016] Generate a historical data package of the target container category;

[0017] Generate a detection evaluation value c of the target container category based on historical data packets;

[0018] The area of ​​the segmented sub-region of the target container category is set according to the detection evaluation value c, and a plurality of segmented sub-regions of the target container category are set;

[0019] Establish the segmentation sub-region sequence B of the target container category, B=(b 1 , b 2 …b i …b m ), where b i is the i-th segmentation sub-region in the target container category; m is the number of segmentation sub-regions in the target container category;

[0020] Generate training data packets for each segmented sub-region, and establish a sub-image model of the target container category based on all training data packets;

[0021] Generate an expected risk value for each segmented sub-region based on historical data packets, and set the image acquisition amount for each segmented sub-region based on all expected risk values;

[0022] Set the sub-collection strategy of the target container category according to the image collection amount of each segmented sub-area;

[0023] Set the sub-image model and sub-collection strategy for each container category in turn;

[0024] An image analysis model is established based on all sub-image models and all sub-acquisition strategies.

[0025] In some embodiments of the present application, the historical data packet generates a detection evaluation value c of the target container category, including:

[0026] c=e1*Q1* (α 1i *s i )+e2*Q2* (α 2i *w i );

[0027] Among them, e1 is the preset first weight coefficient; e2 is the preset second weight coefficient; Q1 is the preset first fixed coefficient; Q2 is the preset second fixed coefficient; θ1 is the number of damage evaluation indicators; α 1i is the influencing factor of the i-th damage assessment index; s i is the reference value of the i-th damage evaluation index generated based on the historical data packet of the target container category; θ2 is the number of equipment evaluation indicators; α 2i is the influencing factor of the i-th equipment evaluation index; w i is the reference value of the i-th equipment evaluation index in the target container category.

[0028] In some embodiments of the present application, generating an expected risk value for each segmented sub-region includes:

[0029] Set b in sequence according to the number of segmented sub-regions B i Segment the target into sub-regions;

[0030] Generate the expected risk value d of the target segmentation sub-region;

[0031] d= (µ i *j i )

[0032] Among them, θ3 is the number of regional evaluation indicators; µ i is the influencing factor of the evaluation index of the i-th region; j i is the reference value of the evaluation index of the i-th region in the target segmentation subregion;

[0033] Generate the expected risk value of each segmented sub-region in turn;

[0034] Establish the expected risk value series D, D=(d 1 ,d 2 …d i …d m ), where d iis the expected risk value of the i-th segmented sub-region in the target container category; m is the number of segmented sub-regions in the target container category.

[0035] In some embodiments of the present application, setting an image acquisition strategy for a container to be inspected includes:

[0036] Obtain characteristic parameters of the container to be detected;

[0037] Generate similarity evaluation values ​​between the container to be inspected and each container category;

[0038] Establish similar evaluation value series P, P=(p 1 , p 2 …p i …p n ), where p i is the similarity evaluation value between the container to be tested and the i-th container category; n is the number of container categories;

[0039] pi=[ β r *(v r -v' ir ) 2 ];

[0040] Among them, u is the number of characteristic evaluation indicators; β r is the influencing factor of the rth characteristic evaluation index; v r is the reference value of the rth characteristic evaluation index in the container to be tested; v' ir is the reference value of the rth characteristic evaluation index in the i-th container category;

[0041] Set the maximum value p in the similarity evaluation value sequence P max The sub-collection strategy of the corresponding container category is the image collection strategy of the container to be detected.

[0042] In some embodiments of the present application, a primary detection strategy for a container to be detected is generated according to the preprocessing result, including:

[0043] According to the image acquisition strategy of the container to be detected, the detection sub-area sequence B1 of the container to be detected is established, B1=(b 11 ,b 12 …b 1i …b 1n1 ), where b 1i is the i-th detection sub-area of ​​the container to be detected; n1 is the number of detection sub-areas of the container to be detected;

[0044] Generate a primary damage value for each detection sub-area according to the preprocessing result;

[0045] Establish a first-level damage value sequence F, F = (f 1 , f 2 … f i … f n1 ), where fi is the first-level damage value of the i-th detection sub-region; n1 is the number of detection sub-regions of the container to be detected;

[0046] Preset a first damage value threshold F1 and a second damage value threshold F2, and F1 < F2;

[0047] If f i < F1, set the i-th detection sub-region as a safe sub-region;

[0048] If F1 < f i < F2, set the i-th detection sub-region as a risk sub-region;

[0049] If f i > F2, set the i-th detection sub-region as a damaged sub-region;

[0050] Set the first-level detection strategy of the container to be detected according to all risk sub-regions.

[0051] In some embodiments of the present application, the first-level detection strategy includes:

[0052] Establish a risk sub-region sequence B2, B2 = (b 21 , b 22 … b 2i … b 2n2 ), where b 2i is the i-th risk sub-region of the container to be detected; n2 is the number of risk sub-regions of the container to be detected;

[0053] Set b 2i as the target risk sub-region in turn according to the risk sub-region sequence B2;

[0054] Obtain the feedback data packet of the target risk sub-region and generate the second-level damage value g of the target risk sub-region;

[0055] g = µ i * t i ;

[0056] where x is the number of data categories to be collected in the target risk sub-region; µ i is the influence factor of the i-th type of data; t i is the expected damage value generated based on the collected i-th type of data;

[0057] Generate the second-level damage values of each risk sub-region in turn.

[0058] In some embodiments of the present application, generating a damage risk value of a container to be inspected includes:

[0059] Generate a damage risk value h based on the secondary damage value of each risk sub-area and all damaged sub-areas;

[0060] h=e3*Q3* Y(i)*(f 2i -f')]+e4*Q4* η i *z i ;

[0061] Among them, e3 is the preset third weight coefficient; e4 is the preset fourth weight coefficient; Q3 is the preset third fixed coefficient; Q4 is the preset fourth fixed coefficient; f 2i is the secondary damage value of the ith risk sub-area; n2 is the number of risk sub-areas for containers to be tested; f2i is the secondary damage value of the ith risk sub-area for containers to be tested; f' is the secondary damage value threshold; Y(i) is the selection coefficient; if (f 2i -f')>0,Y(i)=1; if (f 2i -f')<0,Y(i)=0; θ4 is the number of damaged area indicators; η i is the influencing factor of the i-th damaged area index; z i is the reference value of the index of the i-th damaged area in the container to be tested;

[0062] Preset damage risk value threshold H1;

[0063] If h>H1, a damage warning instruction for the container to be inspected is generated.

[0064] In some embodiments of the present application, a container damage detection system based on deep learning is provided, including:

[0065] The central control unit is used to establish multiple container categories and establish an image analysis model based on all container categories;

[0066] A detection unit, including a plurality of detection submodules, wherein the detection unit is used to set an image acquisition strategy for a container to be detected;

[0067] The detection unit is also used to obtain an image data packet of the container to be detected according to an image acquisition strategy;

[0068] The central control unit comprises:

[0069] A first processing module, used for generating a preprocessing result of an image data packet according to an image analysis model;

[0070] The second processing module is used to generate a primary detection strategy for the container to be detected according to the preprocessing result, and generate a damage risk value of the container to be detected;

[0071] An early warning module is used to determine whether to generate a damage early warning instruction according to the damage risk value;

[0072] The third processing module is used to establish a container category sequence A, A=(a 1 ,a 2 …a i …a n ), where a i is the i-th container category; n is the number of container categories.

[0073] In some embodiments of the present application, the central control unit further includes:

[0074] A fourth processing module is used to establish an image analysis model according to all container categories;

[0075] The fourth processing module is further used for:

[0076] According to the container type sequence A, set a i is the target container category;

[0077] Generate a historical data package of the target container category;

[0078] Generate a detection evaluation value c of the target container category based on historical data packets;

[0079] The area of ​​the segmented sub-region of the target container category is set according to the detection evaluation value c, and a plurality of segmented sub-regions of the target container category are set;

[0080] Establish the segmentation sub-region sequence B of the target container category, B=(b 1 , b 2 …b i …b m ), where b i is the i-th segmentation sub-region in the target container category; m is the number of segmentation sub-regions in the target container category;

[0081] Generate training data packets for each segmented sub-region, and establish a sub-image model of the target container category based on all training data packets;

[0082] Generate an expected risk value for each segmented sub-region based on historical data packets, and set the image acquisition amount for each segmented sub-region based on all expected risk values;

[0083] Set the sub-collection strategy of the target container category according to the image collection amount of each segmented sub-area;

[0084] Set the sub-image model and sub-collection strategy for each container category in turn;

[0085] An image analysis model is established based on all sub-image models and all sub-acquisition strategies.

[0086] Compared with the prior art, the container damage detection method and system based on deep learning in the embodiment of the present application have the following beneficial effects:

[0087] Based on different container equipment parameters, multiple container categories are constructed, and deep learning technology is used to build sub-image models for each container category, so as to adapt to different types of container damage in the actual production process of different ports. Preliminary damage detection is performed by collecting image data of the containers to be inspected, thereby improving the efficiency of container damage detection.

[0088] By analyzing the image data packets of the containers to be inspected, a preliminary judgment is made on the containers to be inspected, and a secondary inspection is carried out on the risk sub-areas therein. During the secondary inspection, multi-source data (laser data, infrared data, ultrasonic data, etc.) is collected for auxiliary judgment, so as to timely warn and diagnose potential residual risks, improve the diagnostic accuracy of container residues, and reduce the overall inspection and operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of a container damage detection method based on deep learning in a preferred embodiment of the present application. DETAILED DESCRIPTION

[0090] The specific implementation methods of the present application are further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but are not intended to limit the scope of the present application.

[0091] In the description of the present application, it should be understood that the terms "center", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0092] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0093] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0094] like Figure 1 As shown, a container damage detection method based on deep learning in a preferred embodiment of the present application includes:

[0095] S101: Establish multiple container categories, and establish an image analysis model based on all container categories;

[0096] S102: setting an image acquisition strategy for the container to be inspected, and acquiring an image data packet of the container to be inspected according to the image acquisition strategy;

[0097] S103: generating a preprocessing result of an image data packet according to the image analysis model;

[0098] S104: generating a primary inspection strategy for the container to be inspected according to the preprocessing result, and generating a damage risk value for the container to be inspected;

[0099] Among them, when setting multiple container categories, including:

[0100] Establish container category sequence A, A=(a 1 ,a 2 …a i …a n ), where a i is the i-th container category; n is the number of container categories.

[0101] Specifically, when establishing an image analysis model based on all container categories, it includes:

[0102] According to the container type sequence A, set a i is the target container category;

[0103] Generate a historical data package of the target container category;

[0104] Generate a detection evaluation value c of the target container category based on historical data packets;

[0105] The area of ​​the segmented sub-region of the target container category is set according to the detection evaluation value c, and a plurality of segmented sub-regions of the target container category are set;

[0106] Establish the segmentation sub-region sequence B of the target container category, B=(b 1 , b 2 …b i …b m ), where b i is the i-th segmentation sub-region in the target container category; m is the number of segmentation sub-regions in the target container category;

[0107] Generate training data packets for each segmented sub-region, and establish a sub-image model of the target container category based on all training data packets;

[0108] Generate an expected risk value for each segmented sub-region based on historical data packets, and set the image acquisition amount for each segmented sub-region based on all expected risk values;

[0109] Set the sub-collection strategy of the target container category according to the image collection amount of each segmented sub-area;

[0110] Set the sub-image model and sub-collection strategy for each container category in turn;

[0111] An image analysis model is established based on all sub-image models and all sub-acquisition strategies.

[0112] Specifically, characteristic evaluation indicators of multiple containers are set, including but not limited to parameters such as container model, service life, shipping time, etc., and multiple container categories are established through random combination results of different value ranges of each characteristic evaluation indicator.

[0113] Specifically, by shooting images of the damage of each segmented sub-area of ​​the target container category under different light levels, weather conditions, and angles, a training data packet for each segmented sub-area is generated.

[0114] Specifically, the training data package includes various situations such as severe damage, minor damage, severe holes, minor holes, missing seals, and rust. The number of damaged images of each type remains consistent, and Labelimg is used to annotate the collected images, and the annotated image data is processed into a format that can be trained by the ultralytics framework. The data set is divided into training set, validation set, and test set in a ratio of 7:2:1.

[0115] Specifically, based on deep learning technology, the initial model of each segmented sub-region is trained, and a sub-comparison model of each segmented sub-region is established according to the iteration results, and a sub-image model of the target container category is generated according to all the sub-comparison models.

[0116] It can be understood that in the above embodiment, multiple container categories are constructed based on different container equipment parameters, and deep learning technology is used to construct sub-image models of each container category, so as to adapt to different types of container damage in the actual production process of different ports.

[0117] In a preferred embodiment of the present application, the historical data packet generates a detection evaluation value c of the target container category, including:

[0118] c=e1*Q1* (α 1i *s i )+e2*Q2* (α 2i *w i );

[0119] Among them, e1 is the preset first weight coefficient; e2 is the preset second weight coefficient; Q1 is the preset first fixed coefficient; Q2 is the preset second fixed coefficient; θ1 is the number of damage evaluation indicators; α 1i is the influencing factor of the i-th damage assessment index; s i is the reference value of the i-th damage evaluation index generated based on the historical data packet of the target container category; θ2 is the number of equipment evaluation indicators; α 2i is the influencing factor of the i-th equipment evaluation index; w i is the reference value of the i-th equipment evaluation index in the target container category.

[0120] Specifically, all parameters in the model are normalized by presetting the first fixed coefficient and the second fixed coefficient, so that each parameter in the model is in the same value range.

[0121] Specifically, the survival evaluation index includes but is not limited to multiple parameters such as the damage frequency of the target container category and the probability of cargo loss after survival.

[0122] Specifically, the equipment evaluation indicators include, but are not limited to, the container area within the container category, the frequency of containers of the target container category entering the port and other parameters.

[0123] Specifically, the larger the detection evaluation value is, the higher the frequency at which containers of the corresponding target container category need to be inspected, and the lower the possibility of damage.

[0124] Specifically, the larger the detection evaluation value, the larger the corresponding segmentation sub-region area. By dynamically adjusting the area of ​​the segmentation sub-region, the image acquisition amount can be reduced and the image acquisition efficiency can be improved while ensuring the accuracy of damage detection, thereby improving the overall damage detection efficiency and reducing operation and maintenance costs.

[0125] Specifically, the expected risk value of each segmented sub-region is generated, including:

[0126] According to the segmentation sub-region sequence B, set bi as the target segmentation sub-region in sequence;

[0127] Generate the expected risk value d of the target segmentation sub-region;

[0128] d= (µ i *j i )

[0129] Among them, θ3 is the number of regional evaluation indicators; µ i is the influencing factor of the evaluation index of the i-th region; j i is the reference value of the evaluation index of the i-th region in the target segmentation subregion;

[0130] Generate the expected risk value of each segmented sub-region in turn;

[0131] Establish the expected risk value series D, D=(d 1 ,d 2 …d i …d m ), where d i is the expected risk value of the i-th segmented sub-region in the target container category; m is the number of segmented sub-regions in the target container category.

[0132] Specifically, regional evaluation indicators include but are not limited to the probability of damage occurring in the region, the type of damage occurring, the degree of interference with the overall container, the significance of the damage, and other parameters. The larger the expected risk value, the greater the probability that damage may occur in the current segmented sub-region, and the lower the significance of the damage, the greater the interference with the container.

[0133] Specifically, the higher the significance, the higher the accuracy of identifying the damage problem in the segmented sub-region through image analysis.

[0134] Specifically, the larger the expected risk value, the larger the corresponding image acquisition amount, thereby ensuring accurate analysis of the damage status of each segmented sub-region.

[0135] It can be understood that in the above embodiment, by dynamically adjusting the image acquisition strategy of each container category, the image acquisition efficiency is improved on the basis of ensuring the damage detection accuracy, thereby improving the overall damage detection efficiency and detection accuracy and reducing the operation and maintenance costs.

[0136] In a preferred embodiment of the present application, the image acquisition strategy of the container to be inspected is set, including:

[0137] Obtain characteristic parameters of the container to be detected;

[0138] Generate similarity evaluation values ​​between the container to be inspected and each container category;

[0139] Establish similar evaluation value series P, P=(p 1 , p 2 …p i …p n ), where p i is the similarity evaluation value between the container to be tested and the i-th container category; n is the number of container categories;

[0140] pi=[ β r *(v r -v' ir ) 2 ];

[0141] Among them, u is the number of characteristic evaluation indicators; β r is the influencing factor of the rth characteristic evaluation index; v r is the reference value of the rth characteristic evaluation index in the container to be tested; v' ir is the reference value of the rth characteristic evaluation index in the i-th container category;

[0142] Set the maximum value p in the similarity evaluation value sequence P max The sub-collection strategy of the corresponding container category is the image collection strategy of the container to be detected.

[0143] Specifically, characteristic evaluation indicators include but are not limited to parameters such as container model, service life, and shipping time.

[0144] Specifically, the larger the similarity evaluation value is, the more suitable the image acquisition strategy and sub-image model within the corresponding container category are for the current container to be detected.

[0145] Specifically, a primary detection strategy for the container to be detected is generated based on the preprocessing results, including:

[0146] According to the image acquisition strategy of the container to be detected, the detection sub-area sequence B1 of the container to be detected is established, B1=(b 11 ,b 12 …b 1i …b 1n1 ), where b 1i is the i-th detection sub-area of ​​the container to be detected; n1 is the number of detection sub-areas of the container to be detected;

[0147] Generate a primary damage value for each detection sub-area according to the preprocessing result;

[0148] Establish the first-level residual value series F, F=(f 1 ,f2 …f i …f n1 ), where f i is the first-level damage value of the i-th detection sub-region; n1 is the number of detection sub-regions of the container to be detected;

[0149] Preset the first damage value threshold F1 and the second damage value threshold F2, and F1 < F2;

[0150] If f i < F1, set the i-th detection sub-region as a safe sub-region;

[0151] If F1 < f i < F2, set the i-th detection sub-region as a risk sub-region;

[0152] If f i > F2, set the i-th detection sub-region as a damaged sub-region;

[0153] Set the first-level detection strategy of the container to be detected according to all risk sub-regions.

[0154] Specifically, the preprocessing result refers to performing sub-image analysis on the image data packet of the container to be detected collected by calling the sub-image analysis model corresponding to the container category through the image acquisition strategy, and identifying the damage parameters in each detection sub-region.

[0155] Specifically, the larger the first-level damage value, the more serious the damage degree in the current segmented sub-region. By preprocessing the image data packet, the remaining categories in each damaged sub-region can be determined.

[0156] Specifically, a safe sub-region refers to a container region where there is no damage currently. A risk sub-region refers to a region where there may be a risk of damage through image acquisition analysis, but the remaining category and degree cannot be accurately judged. A damaged sub-region refers to a region where the damage category can be accurately judged through image analysis and the damage degree is relatively large.

[0157] It can be understood that in the above embodiments, by generating the similarity evaluation value between the container to be detected and each container category, the image acquisition strategy of the container to be detected is quickly generated, and the preliminary damage detection is performed by collecting the image data of the container to be detected, improving the damage detection efficiency of the container.

[0158] In the preferred embodiment of this application, the first-level detection strategy includes:

[0159] Establish a risk sub-region sequence B2, B2 = (b 21 , b 22 …b 2i …b 2n2 ), where b2i is the i-th risk sub-area of ​​the container to be inspected; n2 is the number of risk sub-areas of the container to be inspected;

[0160] According to the risk sub-area sequence B2, set b 2i is the target risk sub-area;

[0161] Obtaining a feedback data packet of the target risk sub-area and generating a secondary residual value g of the target risk sub-area;

[0162] g= µ i *t i ;

[0163] Where x is the number of data categories that need to be collected in the target risk sub-area; µ i is the impact factor of the i-th type of data; t i is the expected residual value generated based on the collected data of type i;

[0164] Generate the secondary residual value of each risk sub-area in turn.

[0165] Specifically, the data collected includes, but is not limited to, laser feedback data, ultrasonic feedback data, infrared data and other data. By analyzing and processing various data, the corresponding expected damage value is generated. By analyzing multi-source data, the potential damage risk of the container can be warned in a timely manner.

[0166] Specifically, the larger the secondary damage value is, the more serious the damage is in the current risk sub-area.

[0167] Specifically, the damage categories within each risk sub-area can be generated through comprehensive analysis of multi-source data.

[0168] Specifically, the damage risk value of the container to be inspected is generated, including:

[0169] Generate a damage risk value h based on the secondary damage value of each risk sub-area and all damaged sub-areas;

[0170] h=e3*Q3* Y(i)*(f 2i -f')]+e4*Q4* η i *z i ;

[0171] Among them, e3 is the preset third weight coefficient; e4 is the preset fourth weight coefficient; Q3 is the preset third fixed coefficient; Q4 is the preset fourth fixed coefficient; f 2iis the secondary damage value of the ith risk sub-area; n2 is the number of risk sub-areas for containers to be tested; f2i is the secondary damage value of the ith risk sub-area for containers to be tested; f' is the secondary damage value threshold; Y(i) is the selection coefficient; if (f 2i -f')>0,Y(i)=1; if (f 2i -f')<0,Y(i)=0; θ4 is the number of damaged area indicators; η i is the influencing factor of the i-th damaged area index; z i is the reference value of the index of the i-th damaged area in the container to be tested;

[0172] Preset damage risk value threshold H1;

[0173] If h>H1, a damage warning instruction for the container to be inspected is generated.

[0174] Specifically, all parameters in the model are normalized by presetting the third fixed coefficient and the fourth fixed coefficient, so that each parameter in the model is in the same value range.

[0175] Specifically, the greater the damage risk value, the greater the possibility of cargo loss in the container.

[0176] Specifically, the residual risk value threshold can be set based on historical parameters.

[0177] Specifically, according to the damage warning instruction, a damage category and damage degree table of each detection sub-area can be generated and a corresponding maintenance plan can be set.

[0178] It can be understood that in the above embodiment, by analyzing the image data packets of the container to be inspected, a preliminary judgment is made on the container to be inspected, and a secondary inspection is performed on the risk sub-area therein, and multi-source data (laser data, infrared data, ultrasonic data, etc.) is collected during the secondary inspection for auxiliary judgment, so as to timely warn and diagnose potential residual risks, improve the diagnostic accuracy of container residues, and reduce the overall inspection and operation and maintenance costs.

[0179] Based on another preferred embodiment of a container damage detection method based on deep learning in any of the above preferred embodiments, this preferred embodiment provides a container damage detection system based on deep learning, including:

[0180] The central control unit is used to establish multiple container categories and establish an image analysis model based on all container categories;

[0181] A detection unit, including a plurality of detection submodules, the detection unit is used to set an image acquisition strategy for a container to be detected;

[0182] The detection unit is also used to obtain an image data packet of the container to be detected according to the image acquisition strategy;

[0183] The central control unit includes:

[0184] A first processing module, used for generating a preprocessing result of an image data packet according to an image analysis model;

[0185] The second processing module is used to generate a primary detection strategy for the container to be detected according to the preprocessing result, and generate a damage risk value of the container to be detected;

[0186] An early warning module is used to determine whether to generate a damage early warning instruction according to the damage risk value;

[0187] The third processing module is used to establish a container category sequence A, A=(a 1 ,a 2 …a i …a n ), where a i is the i-th container category; n is the number of container categories.

[0188] Specifically, the detection unit is preferably an unmanned aerial vehicle device, which is equipped with multiple data acquisition devices such as image acquisition equipment, laser data acquisition equipment, infrared equipment, ultrasonic acquisition equipment, etc.

[0189] Specifically, the central control unit also includes:

[0190] A fourth processing module is used to establish an image analysis model according to all container categories;

[0191] The fourth processing module is also used for:

[0192] According to the container type sequence A, set a i is the target container category;

[0193] Generate a historical data package of the target container category;

[0194] Generate a detection evaluation value c of the target container category based on historical data packets;

[0195] The area of ​​the segmented sub-region of the target container category is set according to the detection evaluation value c, and a plurality of segmented sub-regions of the target container category are set;

[0196] Establish the segmentation sub-region sequence B of the target container category, B=(b 1 , b 2 …b i …b m ), where b i is the i-th segmentation sub-region in the target container category; m is the number of segmentation sub-regions in the target container category;

[0197] Generate training data packets for each segmented sub-region, and establish a sub-image model of the target container category based on all training data packets;

[0198] Generate an expected risk value for each segmented sub-region based on historical data packets, and set the image acquisition amount for each segmented sub-region based on all expected risk values;

[0199] Set the sub-collection strategy of the target container category according to the image collection amount of each segmented sub-area;

[0200] Set the sub-image model and sub-collection strategy for each container category in turn;

[0201] An image analysis model is established based on all sub-image models and all sub-acquisition strategies.

[0202] According to the first concept of the present application, multiple container categories are constructed based on different container equipment parameters, and sub-image models of each container category are constructed using deep learning technology to adapt to different types of container damage in the actual production process of different ports. Preliminary damage detection is performed by collecting image data of the containers to be inspected, thereby improving the efficiency of container damage detection.

[0203] According to the second concept of the present application, by analyzing the image data packets of the container to be inspected, a preliminary judgment is made on the container to be inspected, and a secondary inspection is performed on the risk sub-area therein. During the secondary inspection, multi-source data (laser data, infrared data, ultrasonic data, etc.) is collected for auxiliary judgment, so as to timely warn and diagnose potential residual risks, improve the diagnostic accuracy of container residues, and reduce the overall inspection and operation and maintenance costs.

[0204] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present application. These improvements and substitutions should also be regarded as the scope of protection of the present application.

Claims

1. A container damage detection method based on deep learning, characterized in that: include: Establish multiple container categories and build image analysis models based on all container categories; Setting an image acquisition strategy for the container to be inspected, and acquiring an image data packet of the container to be inspected according to the image acquisition strategy; Generate preprocessing results of image data packets according to the image analysis model; Generate a primary inspection strategy for the container to be inspected according to the preprocessing result, and generate a damage risk value for the container to be inspected; Among them, when setting multiple container categories, including: Establish container category sequence A, A=(a1,a2…a i …a n ), where a i is the i-th container category; n is the number of container categories; The primary detection strategy includes: Generate a primary damage value for each detection sub-area according to the preprocessing result; Generate the secondary residual value of each risk sub-area in turn; Generate a damage risk value h based on the secondary damage value of each risk sub-area and all damaged sub-areas; When building an image analysis model based on all container categories, it includes: According to the container type sequence A, set a i is the target container category; Generate a historical data package of the target container category; Generate a detection evaluation value c of the target container category based on historical data packets; The area of ​​the segmented sub-region of the target container category is set according to the detection evaluation value c, and a plurality of segmented sub-regions of the target container category are set; Establish a sequence B of segmented sub-regions of the target container category, B=(b1, b2…b i …b m ), where b i is the i-th segmentation sub-region in the target container category; m is the number of segmentation sub-regions in the target container category; Generate training data packets for each segmented sub-region, and establish a sub-image model of the target container category based on all training data packets; Generate an expected risk value for each segmented sub-region based on historical data packets, and set the image acquisition amount for each segmented sub-region based on all expected risk values; Set the sub-collection strategy of the target container category according to the image collection amount of each segmented sub-area; Set the sub-image model and sub-collection strategy for each container category in turn; An image analysis model is established based on all sub-image models and all sub-acquisition strategies.

2. The container damage detection method based on deep learning according to claim 1, characterized in that: The historical data packet generates the detection evaluation value c of the target container category, including: c=e1*Q1* (a 1i *s i )+e2*Q2* (a 2i *w i ); Among them, e1 is the preset first weight coefficient; e2 is the preset second weight coefficient; Q1 is the preset first fixed coefficient; Q2 is the preset second fixed coefficient; θ1 is the number of damage evaluation indicators; α 1i is the influencing factor of the i-th damage assessment index; s i is the reference value of the i-th damage evaluation index generated based on the historical data packet of the target container category; θ2 is the number of equipment evaluation indicators; α 2i is the influencing factor of the i-th equipment evaluation index; w i is the reference value of the i-th equipment evaluation index in the target container category.

3. The container damage detection method based on deep learning according to claim 2, characterized in that: Generate expected risk values ​​for each segmented sub-region, including: Set b in sequence according to the number of segmented sub-regions B i Segment the target into sub-regions; Generate the expected risk value d of the target segmentation sub-region; d= (µ i *j i ) Among them, θ3 is the number of regional evaluation indicators; µ i is the influencing factor of the evaluation index of the i-th region; j i is the reference value of the evaluation index of the i-th region in the target segmentation subregion; Generate the expected risk value of each segmented sub-region in turn; Establish the expected risk value series D, D = (d1, d2…d i …d m ), where d i is the expected risk value of the i-th segmented sub-region in the target container category; m is the number of segmented sub-regions in the target container category.

4. The container damage detection method based on deep learning according to claim 1, characterized in that: Set the image acquisition strategy for the container to be inspected, including: Obtain characteristic parameters of the container to be detected; Generate similarity evaluation values ​​between the container to be inspected and each container category; Establish a similar evaluation value sequence P, P = (p1, p2…p i …p n ), where p i is the similarity evaluation value between the container to be tested and the i-th container category; n is the number of container categories; pi=[ β r *(v) r -v' ir ) 2 ]; Among them, u is the number of characteristic evaluation indicators; β r is the influencing factor of the rth characteristic evaluation index; v r is the reference value of the rth characteristic evaluation index in the container to be tested; v' ir is the reference value of the rth characteristic evaluation index in the i-th container category; Set the maximum value p in the similarity evaluation value sequence P max The sub-collection strategy of the corresponding container category is the image collection strategy of the container to be detected.

5. The container damage detection method based on deep learning according to claim 4, characterized in that: Generate the first-level detection strategy of the container to be detected based on the preprocessing results, including: According to the image acquisition strategy of the container to be detected, the detection sub-area sequence B1 of the container to be detected is established, B1=(b 11 ,b 12 …b 1i …b 1n1 ), where b 1i is the i-th detection sub-area of ​​the container to be detected; n1 is the number of detection sub-areas of the container to be detected; Generate a primary damage value for each detection sub-area according to the preprocessing result; Establish a first-level residual value sequence F, F=(f1,f2…f i …f n1 ), where f i is the primary damage value of the i-th detection sub-area; n1 is the number of detection sub-areas of the container to be detected; A first residual value threshold F1 and a second residual value threshold F2 are preset, and F1 <F2; If f i <F1, set the i-th detected sub-region as a safe sub-region; If F1 < f i <F2, set the i-th detection sub-region as a risk sub-region; If f i >F2, set the i-th detected sub-region as a damaged sub-region; A primary inspection strategy for containers to be inspected is set based on all risk sub-areas.

6. The container damage detection method based on deep learning according to claim 5, characterized in that: The primary detection strategy includes: Establish risk sub-area sequence B2, B2=(b 21 ,b 22 …b 2i …b 2n2 ), where b 2i is the i-th risk sub-area of ​​the container to be inspected; n2 is the number of risk sub-areas of the container to be inspected; According to the risk sub-area sequence B2, set b 2i is the target risk sub-area; Obtaining a feedback data packet of the target risk sub-area and generating a secondary residual value g of the target risk sub-area; g= µ i *t i ; Where x is the number of data categories that need to be collected in the target risk sub-area; µ i is the impact factor of the i-th type of data; t i is the expected residual value generated based on the collected data of type i; Generate the secondary residual value of each risk sub-area in turn.

7. The container damage detection method based on deep learning according to claim 6, characterized in that: Generate the damage risk value of the container to be inspected, including: Generate a damage risk value h based on the secondary damage value of each risk sub-area and all damaged sub-areas; h=e3*Q3* Y(i)*(f 2i -f')]+e4*Q4* η i *z i ; Among them, e3 is the preset third weight coefficient; e4 is the preset fourth weight coefficient; Q3 is the preset third fixed coefficient; Q4 is the preset fourth fixed coefficient; f 2i is the secondary damage value of the ith risk sub-area; n2 is the number of risk sub-areas for containers to be tested; f2i is the secondary damage value of the ith risk sub-area for containers to be tested; f' is the secondary damage value threshold; Y(i) is the selection coefficient; if (f 2i -f')>0,Y(i)=1; if (f 2i -f')<0,Y(i)=0; θ4 is the number of damaged area indicators; η i is the influencing factor of the i-th damaged area index; z i is the reference value of the index of the i-th damaged area in the container to be tested; Preset damage risk value threshold H1; If h>H1, a damage warning instruction for the container to be inspected is generated.

8. A container damage detection system based on deep learning, using the container damage detection method based on deep learning described in any one of claims 1 to 7, characterized in that: include: The central control unit is used to establish multiple container categories and establish an image analysis model based on all container categories; A detection unit, including a plurality of detection submodules, wherein the detection unit is used to set an image acquisition strategy for a container to be detected; The detection unit is also used to obtain an image data packet of the container to be detected according to an image acquisition strategy; The central control unit comprises: A first processing module, used for generating a preprocessing result of an image data packet according to an image analysis model; The second processing module is used to generate a primary detection strategy for the container to be detected according to the preprocessing result, and generate a damage risk value of the container to be detected; An early warning module is used to determine whether to generate a damage early warning instruction according to the damage risk value; The third processing module is used to establish a container category sequence A, A=(a1, a2…a i …a n ), where a i is the i-th container category; n is the number of container categories.

9. The container damage detection system based on deep learning according to claim 8, characterized in that: The central control unit also includes: A fourth processing module is used to establish an image analysis model according to all container categories; The fourth processing module is further used for: According to the container type sequence A, set a i is the target container category; Generate a historical data package of the target container category; Generate a detection evaluation value c of the target container category based on historical data packets; The area of ​​the segmented sub-region of the target container category is set according to the detection evaluation value c, and a plurality of segmented sub-regions of the target container category are set; Establish a sequence B of segmented sub-regions of the target container category, B=(b1, b2…b i …b m ), where b i is the i-th segmentation sub-region in the target container category; m is the number of segmentation sub-regions in the target container category; Generate training data packets for each segmented sub-region, and establish a sub-image model of the target container category based on all training data packets; Generate an expected risk value for each segmented sub-region based on historical data packets, and set the image acquisition amount for each segmented sub-region based on all expected risk values; Set the sub-collection strategy of the target container category according to the image collection amount of each segmented sub-area; Set the sub-image model and sub-collection strategy for each container category in turn; An image analysis model is established based on all sub-image models and all sub-acquisition strategies.

Images