Method and system for detecting defects of steel belt elevator with machine room based on infrared nondestructive imaging

Abstract

The invention provides a defect detection method and system for a steel belt elevator with a machine room based on infrared nondestructive imaging, and relates to the technical field of equipment detection.The infrared thermal image sequence formed after a steel belt is excited by a pulse heat source is obtained through infrared camera devices arranged on the two sides of a traction machine, the pulse heat source is used for performing pulse irradiation on the running steel strip so as to form temperature difference on the surface and the subsurface layer of the steel strip; the infrared thermal image sequence is input into a defect target detection model for processing, a defect positioning result corresponding to the infrared thermal image sequence is output, and the defect target detection model comprises a feature extraction backbone network, a feature fusion neck network and a target detection head which are connected in sequence. According to the invention, the defect detectable rate of the steel belt can be improved.

Description

Technical Field

[0001] This invention relates to the field of equipment testing technology, and in particular to a method and system for detecting defects in steel-framed elevators with organic rooms based on infrared non-destructive imaging. Background Technology

[0002] Steel belt elevators are an improvement on traditional traction elevators, with the core feature being the use of flat steel belts instead of traditional steel wire ropes as the load-bearing component for traction drive. Steel belt elevators still employ the traction drive principle: the car and counterweight are connected by a steel belt, which wraps around the traction sheave, and the lifting motion is achieved through the friction between the sheave and the steel belt. The counterweight is typically designed to have the empty weight of the car plus 40%-50% of the rated load to achieve energy saving and load balancing. The steel belt is composed of multiple parallel steel wires embedded in a polyurethane or rubber outer layer, forming a flat strip. A frequency converter precisely controls the speed and direction of the traction machine, and multiple protection devices such as speed governors and safety brakes ensure smooth start-stop and precise floor stopping.

[0003] Chinese patent CN116256368A discloses a method for determining the scrapping of elevator composite steel belts based on internal and external defect detection. This method uses cameras installed on both sides of the steel belt on the elevator traction sheave to acquire images of the outer surface of both sides of the steel belt. The outer surface images are preprocessed to detect defects. Magnetic flux leakage detection technology is used to detect internal defects in the steel belt within the imaged area. If internal defects are found in the imaged area, the steel belt meets the scrapping standard. If not, the imaged area is further examined for external defects. If external defects exist and the total area of ​​external weighted defects reaches 10% of the imaged area, the steel belt meets the scrapping standard. Otherwise, the method continues to detect defects on the outer surface and inside of other areas captured as the elevator traction sheave rotates. However, this method relies on magnetic flux leakage and is only sensitive to discontinuous defects in ferromagnetic materials. The signal from deeply buried defects and small broken wires attenuates rapidly with depth, resulting in insufficient defect detection rate for the steel belt. Therefore, it is essential to provide a method and system for detecting defects in steel-belt elevators with machine rooms based on infrared non-destructive imaging to improve the detectability rate of defects in the steel belt. Summary of the Invention

[0004] In view of this, the present invention proposes a defect detection method and system for steel-belt elevators with organic rooms based on infrared non-destructive imaging.

[0005] This invention provides a defect detection method for steel-framed elevators with organic rooms based on infrared non-destructive imaging, the method comprising:

[0006] Infrared thermal image sequences formed by the steel strip after being excited by a pulsed heat source are obtained by infrared camera devices installed on both sides of the traction machine. The pulsed heat source irradiates the running steel strip with pulses to create a temperature difference between the surface and subsurface of the steel strip.

[0007] The infrared thermal image sequence is input into the defect target detection model for processing to output the defect localization result corresponding to the infrared thermal image sequence. The defect target detection model includes a feature extraction backbone network, a feature fusion neck network, and a target detection head connected in sequence.

[0008] Based on the above technical solutions, preferably, the feature extraction backbone network includes a feature extraction module and a multimodal attention space pyramid module. The feature extraction module learns multi-scale semantic and texture features from the infrared thermal image sequence from shallow to deep, and forms a basic feature level and a corresponding transition backbone feature map. The multimodal attention space pyramid module enhances and filters the transition backbone feature map to obtain the backbone feature map and highlight the defect responses in the backbone feature map.

[0009] Based on the above technical solutions, preferably, the feature extraction module includes an input stem, multiple parallel feature extraction sub-columns, a multi-level cross-column reversible fusion unit, and an output aggregation end, wherein...

[0010] The input stem is connected to all feature extraction sub-columns. The input stem is used to receive the input infrared thermal image sequence and embed the features of the infrared thermal image sequence through the first convolutional layer.

[0011] Each feature extraction sub-column, which is set up in parallel, includes multiple levels of downsampling units and Tiny units with 3×3 convolution as the core, arranged in order of depth.

[0012] The multi-level cross-column reversible fusion unit is set between the corresponding levels of adjacent feature extraction sub-columns. The multi-level cross-column reversible fusion unit is used to align the features of each sub-column at the same level and then achieve information interaction by summing them element by element.

[0013] The output aggregation terminal is connected to all feature extraction sub-columns. The output aggregation terminal is used to fuse the end or intermediate layer features of each feature extraction sub-column and output a transition backbone feature map with multi-scale features.

[0014] More preferably, the multimodal attention space pyramid module includes a frequency domain branching unit, a channel attention branching unit, a hybrid pooling branching unit, and a fusion and output unit connected in sequence, wherein,

[0015] The frequency domain branching unit is used to perform a two-dimensional discrete wavelet transform on the transition backbone feature map, so as to suppress noise in the transition backbone feature map through low-frequency components and enhance the target edges of the transition backbone feature map through high-frequency components, so as to generate a multi-resolution feature map.

[0016] The channel attention branch unit is used to sequentially perform global average pooling and global max pooling on the multi-resolution feature map to generate channel attention and dynamic channel weights corresponding to the channel attention, and to weightedly fuse the channel attention with spatial frequency attention to obtain the weighted multi-resolution feature map.

[0017] The hybrid pooling branch unit is used to perform max pooling and average pooling on the weighted multi-resolution feature map to obtain a standard feature map.

[0018] The fusion and output unit is used to perform convolution optimization on the standard feature map according to the dynamic large kernel depth separable convolution to obtain the backbone feature map, wherein the kernel size of the dynamic large kernel depth separable convolution is adaptively selected according to the input resolution.

[0019] More preferably, the frequency domain branching unit uses discrete wavelet transform to perform a first-level decomposition of the features corresponding to the transition backbone feature map to obtain LL subband, LH subband, HL subband and HH subband. The LH subband, the HL subband and the HH subband are normalized and convolved to enhance the edge response. The LL subband is subjected to fast Fourier transform to obtain the spectral energy map, and a space-frequency attention mask is formed based on the spectral energy map.

[0020] More preferably, the feature fusion neck network includes a multi-scale input terminal, a lateral alignment unit, a sampling unit, a dynamically gated weighting unit, and a lightweight aggregation unit connected in sequence, wherein,

[0021] The multi-scale input terminal is used to receive in parallel multiple backbone feature maps with different resolutions output by the feature extraction backbone network;

[0022] The lateral alignment unit is used to unify the number of channels in the feature maps of each backbone.

[0023] The sampling unit is used to upsample the deep features in the backbone feature map to a medium-shallow resolution, or downsample the shallow features in the backbone feature map to a medium-deep resolution, so as to obtain multi-source features of the same scale.

[0024] The dynamic gated weighting unit is used to adaptively generate weights based on the global statistics of each backbone feature map, and to weight or selectively superimpose multi-source features of the same scale.

[0025] Lightweight aggregation units are used to perform feature fusion at multiple scales in an element-wise summation or channel-wise concatenation manner, and output multi-scale features for the detection head.

[0026] More preferably, the convolutions in the lightweight aggregation unit employ depthwise separable convolutions and share a convolution kernel, and reuse the same set of parameters in each scale path.

[0027] A second aspect of this application provides a defect detection system for steel-framed elevators with machine rooms based on infrared non-destructive imaging. The system includes an image acquisition module and a defect detection module.

[0028] The image acquisition module is used to acquire an infrared thermal image sequence formed by the steel strip after being excited by a pulse heat source through infrared camera devices set on both sides of the traction machine. The pulse heat source irradiates the running steel strip with pulses to create a temperature difference between the surface and subsurface of the steel strip.

[0029] The defect detection module is used to input the infrared thermal image sequence into the defect target detection model for processing, so as to output the defect localization result corresponding to the infrared thermal image sequence. The defect target detection model includes a feature extraction backbone network, a feature fusion neck network, and a target detection head connected in sequence.

[0030] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory.

[0031] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a method for detecting defects in a steel-framed elevator with an organic room based on infrared non-destructive imaging.

[0032] The present invention provides a method and system for defect detection in steel-framed elevators with organic rooms based on infrared non-destructive imaging, which has the following advantages over existing technologies:

[0033] (1) By using pulsed thermal excitation to form transient temperature difference on the surface and subsurface of the steel strip, the defect area produces significant thermal contrast due to thermal diffusion / heat capacity difference, thereby realizing early detection of internal hidden dangers. The difference in temperature decay / response over time can suppress background drift and ambient temperature fluctuation, improve defect contrast and stability, and thus significantly improve the defect detection rate in composite strip scenarios. The pulse excitation time is short and the heat load is low. Combined with continuous imaging and real-time processing, rapid screening can be completed while the elevator is running, avoiding the impact and cost of downtime detection on operation and maintenance. At the same time, the backbone network extracts high / low layer semantic features, the neck network performs feature pyramid fusion, and the detection head outputs results for features of different scales simultaneously, taking into account both micro-cracks / small debonding and large area defects. With thermal image sequence as input, the model learns dynamic features such as temperature time-varying curve and thermal diffusion rate, thereby better distinguishing between real defects and occasional noise or surface stains.

[0034] (2) By using the structure from the input stem to the output aggregation end, the characterization quality of infrared temporal defect detection is significantly improved without significantly increasing the parameters and computation. The input stem performs first-layer convolution embedding on the thermal image sequence to suppress sensor noise and background drift and retain the time evolution prior. Multiple parallel feature extraction sub-columns improve the detectability of small targets and low-contrast subsurface defects and have high throughput. Multi-level cross-column reversible fusion units set at the same level in adjacent sub-columns achieve efficient information complementarity by element-wise addition after scale alignment. Finally, higher detection rate / lower false negatives are achieved under complex textures and environmental disturbances, while meeting the requirements of online real-time detection and edge deployment. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating a defect detection method for steel-belt elevators with organic rooms based on infrared non-destructive imaging provided by this invention;

[0037] Figure 2 This is a schematic diagram of the hardware device installation provided by the present invention;

[0038] Figure 3 This is a schematic diagram of the structure of the defect target detection model provided by the present invention;

[0039] Figure 4 This is a schematic diagram of the feature extraction module provided by the present invention;

[0040] Figure 5 This is a schematic diagram of the structure of the multimodal attention space pyramid module provided by the present invention;

[0041] Figure 6 This is a schematic diagram of the structure of the real-time task scheduling system provided by the present invention;

[0042] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0043] Explanation of reference numerals in the attached diagram: 1. Defect detection system for steel-framed elevators with machine room; 11. Image acquisition module; 12. Defect detection module; 2. Electronic equipment; 21. Processor; 22. Communication bus; 23. User interface; 24. Network interface; 25. Memory. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0045] This invention discloses a defect detection method for steel-framed elevators with organic rooms based on infrared non-destructive imaging, with reference to... Figure 1 The steps of this method include S1 to S2.

[0046] Step S1: The infrared thermal image sequence formed by the steel strip after being excited by a pulse heat source is obtained by infrared camera devices set on both sides of the traction machine. The pulse heat source irradiates the running steel strip with pulses to form a temperature difference between the surface and subsurface of the steel strip.

[0047] Please see Figure 2 , Figure 2 This diagram illustrates the structure and installation of an infrared camera system. The infrared cameras are symmetrically arranged on both sides of the traction machine. Halogen lamps, positioned on the ceiling of the elevator machine room, act as pulse sources to irradiate the steel strip, accelerating its temperature change and facilitating subsequent processing. The infrared camera system mainly consists of an infrared camera, a guide rail, a pipeline protection cover, a cylinder module, and a camera base. The infrared camera, mounted on the base, captures the thermal image of the steel strip after halogen lamp irradiation. The base is mounted on the guide rail, and the cylinder module drives the infrared camera to move along the guide rail along a predetermined path. This combination allows for on-the-spot adjustment of the infrared camera's position. The pipeline protection cover protects the relevant wiring.

[0048] In infrared nondestructive testing, the temperature field distribution on a material surface is primarily influenced by internal heat conduction. Heat conduction within a material can occur in three directions: the x and y axes represent the surface direction, while the z-axis represents the depth direction. Therefore, the temperature change within the material over time can be expressed by the following equation:

[0049] ΔT = T(x,y,z,t)

[0050] Where ΔT represents the temperature change, x, y, z represent the coordinates on the x, y, z axes respectively, and t represents time.

[0051] The heat transfer process inside a material can be described by the following differential equation:

[0052]

[0053] Where λ represents the thermal conductivity of the material in different directions, ρ represents the density of the material, and c represents the specific heat capacity. Without considering lateral heat diffusion, the heat conduction model can be simplified to a one-dimensional problem, assuming that heat conduction only varies along the Z-axis:

[0054]

[0055] Where q(z,t) represents the heat absorbed or released at any location of the material at time t. When heat exchange occurs in an object, the heat flux q(z,t) can be expressed as:

[0056] q(z,t)=hΔT=h(T b -T h )

[0057] Among them, T b T represents the surface temperature of a material. h Here, represents the ambient temperature or fluid temperature, and h represents the convective heat transfer coefficient. In a one-dimensional heat conduction model, if lateral heat diffusion is neglected, the initial conditions can be given by the following equation:

[0058]

[0059] When defects exist within a material, the heat conduction process is affected, leading to abnormal temperatures in the defect area. Therefore, the temperature difference between the defect area and the normal area can be detected using an infrared thermal imager. Subsequently, by employing a steel strip infrared image defect target detection method based on MRLD-Yolo, steel strip defects in steel strip elevators can be detected.

[0060] Step S2: Input the infrared thermal image sequence into the defect target detection model for processing, so as to output the defect localization result corresponding to the infrared thermal image sequence. The defect target detection model includes a feature extraction backbone network, a feature fusion neck network, and a target detection head connected in sequence.

[0061] Infrared thermal image sequences are input into a defect target detection model for processing to output defect location results corresponding to the infrared thermal image sequences. The defect target detection model includes a feature extraction backbone network, a feature fusion neck network, and a target detection head connected in sequence.

[0062] Please see Figure 3 , Figure 3 This is a schematic diagram of the defect target detection model. The RevCol-Tiny feature extraction backbone network proposed in this invention can effectively reduce the complexity of feature extraction. RevCol-Tiny is mainly used to reconstruct and enhance the YOLOv5 backbone network. In the RevCol-Tiny feature extraction backbone network, a multimodal attention space pyramid module (MA-LASP) is used to avoid the influence of the input image size. The LTDMF of the neck is based on the traditional feature pyramid network architecture (FPN), which can generate multi-scale sequence features and use infrared thermal images captured by an infrared camera to obtain multi-scale infrared image information of the steel strip, capturing the correlation between features. The target detection head uses the YOLOv5 detection mechanism to perform fast and effective regression on small targets in the infrared image. Specifically, the backbone of the RevCol-Tiny feature extraction backbone network includes four feature extraction layers, denoted as feature extraction layer n, where n∈[1,4], which are important components for extracting discriminative features. In the hierarchical feature extraction process, each layer generates multi-scale feature maps. P2, P3, and P4 specifically represent the outputs of the last three stages of the RevCol-Tiny feature extraction backbone network, playing a crucial role in capturing discriminative information. P5 represents the final output of RevCol-Tiny, including feature maps enhanced by the MA-LASP module.

[0063] During feature fusion, intermediate feature maps, denoted as Z1, Z2, Z3, and Z4, are created. These feature maps represent the vertices of the intermediate layers of the feature pyramid fusion process, effectively fusing information from different scales. Finally, the fused multi-level feature maps are named S1, S2, and S3, improving the detection capability of infrared images. The generated multi-scale feature maps are then input into the YOLOv5 detection head, and after regression analysis, the defect region detection bounding boxes are accurately obtained.

[0064] like Figure 4 As shown, the feature extraction module includes an input stem, multiple parallel feature extraction sub-columns, a multi-level cross-column reversible fusion unit, and an output aggregation end.

[0065] The input stem is connected to all feature extraction sub-columns. The input stem is used to receive the input infrared thermal image sequence and perform feature embedding on the infrared thermal image sequence through the first convolutional layer.

[0066] Each feature extraction sub-column, which is set up in parallel, includes multiple levels of downsampling units and Tiny units with 3×3 convolution as the core, arranged in order of depth.

[0067] Multi-level cross-column reversible fusion units are set between corresponding levels of adjacent feature extraction sub-columns. The multi-level cross-column reversible fusion units are used to align the features of each sub-column at the same level and then achieve information interaction by summing them element by element. Moreover, this interaction is mathematically reversible so that intermediate activations do not need to be saved when backpropagating.

[0068] The output aggregation terminal is connected to all feature extraction sub-columns. The output aggregation terminal is used to fuse the end or intermediate layer features of each feature extraction sub-column and output a transition backbone feature map with multi-scale features.

[0069] In this embodiment, by using an input-to-output aggregation structure, the characterization quality, training / inference efficiency, and engineering deployability of infrared temporal defect detection are significantly improved without significantly increasing parameters and computation. The input-to-output aggregation structure performs first-layer convolutional embedding on the thermal image sequence to suppress sensor noise and background drift while preserving temporal evolution priors. Multiple parallel feature extraction sub-columns consist of multi-level downsampling and Tiny units with 3×3 convolutions at their core, constructing multi-scale, multi-receptive-field paths to accommodate both fine-grained edges and large-scale thermal fields, improving the detectability of small targets and low-contrast subsurface defects while achieving high throughput. The structure is configured in adjacent sub-columns... The multi-level cross-column reversible fusion unit in the hierarchy achieves efficient information complementarity by element-wise summation after scale / channel alignment. Compared with splicing fusion, it has zero parameter increment and low bandwidth consumption. When reversing, it reconstructs intermediate activations and significantly saves GPU memory, thus supporting deeper and wider networks or longer time series / higher resolution training. The output aggregation end performs unified fusion of the end or middle layer features of each sub-column to form a multi-scale "transition backbone feature map" that directly connects to the subsequent neck network, strengthens supervision and gradient short-circuiting, accelerates convergence and improves robustness. Finally, it achieves higher detection rate / lower false negatives under complex textures and environmental perturbations, while meeting the requirements of online real-time detection and edge deployment.

[0070] Furthermore, the number of parallel feature extraction sub-columns is three, forming three feature pathways: left, middle, and right. The feature extraction backbone network has at least four levels of reversible fusion units along its depth. Each level includes bidirectional summation nodes located between adjacent feature extraction sub-columns, used for cross-column information exchange after the Tiny unit computation of the current level. Downsampling units use convolutions with a stride of 2 or equivalent operators to reduce spatial resolution; the downsampling ratio of adjacent levels is 1 / 2 sequentially. The Tiny unit is a lightweight bottleneck structure, including at least one 3×3 convolutional layer and one 1×1 convolutional layer, which are added to the input through residual bypasses to form short connections. The 3×3 convolution in the Tiny unit is located in the main branch, and the 1×1 convolution is used for channel compression and fusion, with normalization and activation functions applied between them. The input stem includes a set of convolution-normalization-activation sequences used to map the input image to a feature map with the base number of channels. The cross-column reversible fusion unit performs element-wise addition on features from two or three feature extraction sub-columns after aligning them along the channel dimension. Its forward / backward directions satisfy a reversible update relationship, allowing the previous state of any column to be reconstructed from the current state and the transformation functions of adjacent columns. The output aggregation end concatenates or weights features from different columns and levels, outputting feature maps at least three scales for subsequent neck network applications.

[0071] Within at least one level of the feature extraction sub-column, a CSP-style structure of splitting-transformation-concatenation is implemented: the input features are divided into two branches, one of which passes through several Bottleneck units sequentially and is then concatenated with the bypass branch along the channel dimension, followed by fusion via 1×1 convolution. The Bottleneck unit comprises a main branch consisting of two levels of convolution operators and a residual bypass of an identity or 1×1 convolution. The output of the main branch and the bypass are added at the end of the unit to obtain the unit output. The number of channels increases with depth at each level, and channel alignment is achieved via 1×1 convolution before cross-column reversible fusion.

[0072] In one example, to accelerate the acquisition of multi-scale feature information of steel strip defects in infrared images and reduce information loss, the RevCol method was used to reorganize the YOLOv5 feature extraction backbone network, resulting in the RevCol-Tiny feature extraction backbone network.

[0073] Tiny units effectively capture multi-scale feature information related to small infrared targets. The use of reversible transformations helps preserve important extracted information while mitigating information conflicts. Specifically, the RevCol-Tiny feature extraction backbone network combines embedding information from different layers and utilizes an unwrapping mechanism to reduce information loss that may occur when extracting features from deeper layers in a single-column network architecture. Furthermore, the RevCol-Tiny feature extraction backbone network also helps fuse deep and shallow semantic information, thereby creating a more comprehensive and richer representation of the input data. The reversible transformation can be expressed as:

[0074] x o =K o (x o-1 x o-p+1 )+εx o-p (Positive)

[0075] x o-p =ε -1 [x o -K o (x o-1 x o-p+1 (Reverse)

[0076] Where o specifies that each subnetwork is at layer o, p represents the number of subnetworks (p≥2), and K o This represents an n×Tiny unit in the o-th layer, where ε represents a fixed set of parameters, set to 0, x o This represents the feature map of the o-th layer.

[0077] The overall implementation process of the RevCol-Tiny feature extraction backbone network is summarized as follows:

[0078] (1) To ensure consistency with the downsampling dimension of YOLOv5, a convolutional layer with a kernel of 1×1 and a stride of 1 is used, and padding is omitted.

[0079] (2) According to the implementation method of RevCol, the downsampled feature map is processed by a feature extraction network consisting of three sub-networks, each of which contains an invertible transformation.

[0080] (3) Figure 4 As shown, the Tiny unit is used to extract features from the current and previous subnetworks.

[0081] (4) The initial sub-network is equivalent to the original feature extraction network used in YOLOv5. In the second and third sub-networks, a fusion module is employed to fuse the features extracted by the Tiny unit from the previous sub-network with the features in the current sub-network. The entire fusion process is described as follows: First, we upsample the low-resolution feature map to improve its resolution. Then, we downsample the high-resolution feature map to match its resolution with the upsampled low-resolution feature map. Finally, we use a "Sum" operation, i.e., adding corresponding elements, to merge the two adjusted feature maps, resulting in a composite feature map containing both resolutions.

[0082] (5) Information from the corresponding layer in the previous sub-network is embedded into it through reversible transformation.

[0083] (6) To mitigate the information collapse problem, RevCol incorporates an intermediate supervision method in the second sub-network. This method allows the network to receive feedback and gradients not only at the final output but also at each intermediate layer.

[0084] like Figure 5 As shown, the multimodal attention space pyramid module includes frequency domain branching units, channel attention branching units, hybrid pooling branching units, and fusion and output units connected in sequence.

[0085] The frequency domain branch unit is used to perform two-dimensional discrete wavelet transform on the transition backbone feature map, so as to suppress noise in the transition backbone feature map through low-frequency components and enhance the target edge of the transition backbone feature map through high-frequency components, so as to generate multi-resolution feature map.

[0086] The channel attention branch unit is used to sequentially perform global average pooling and global max pooling on the multi-resolution feature map to generate channel attention and dynamic channel weights corresponding to the channel attention. The channel attention and spatial frequency attention are weighted and fused together to obtain the weighted multi-resolution feature map.

[0087] The hybrid pooling branch unit is used to perform max pooling and average pooling on the weighted multi-resolution feature map to obtain the standard feature map.

[0088] The fusion and output unit is used to perform convolution optimization on the standard feature map based on the dynamic large kernel depth separable convolution to obtain the backbone feature map, wherein the kernel size of the dynamic large kernel depth separable convolution is adaptively selected according to the input resolution.

[0089] The frequency domain branch unit uses discrete wavelet transform to perform first-level decomposition on the features corresponding to the transition backbone feature map to obtain LL subband, LH subband, HL subband and HH subband. Among them, the LH subband, HL subband and HH subband are normalized and convolved to enhance the edge response. The LL subband is subjected to fast Fourier transform to obtain the spectral energy map, and a space-frequency attention mask is formed based on the spectral energy map.

[0090] Furthermore, based on the LASP module, to address the issues of noise interference and edge blurring in infrared images of steel strips against complex elevator shaft backgrounds, a multimodal attention mechanism is introduced. By fusing spatial-frequency domain analysis and channel-spatial attention, a multimodal attention spatial pyramid module (MA-LASP) is proposed, significantly improving the ability to distinguish between targets and backgrounds and reducing the false detection rate. Its structure is as follows: Figure 4 As shown.

[0091] The multimodal attention space pyramid module processes feature maps in the following order, consisting of four steps:

[0092] Step 1, Wavelet Domain Feature Decomposition: Perform a two-dimensional discrete wavelet transform (DWT) on the input feature map X, decomposing it into four sub-bands. The low-frequency sub-band (LL) is used to capture smooth background information (resolution reduced to 100%). High-frequency subbands (LH, HL, HH): correspond to edge details in the horizontal, vertical, and diagonal directions, respectively (with the same resolution). ).

[0093] Normalized high-frequency subbands to enhance target edges:

[0094] X high =Normalize(X LH ,X HL ,X HH )

[0095] Among them, X high X represents the normalized set of high-frequency subbands. LH X represents the high-frequency subband with horizontal details after wavelet decomposition. HL X represents the high-frequency subband with vertical details after wavelet decomposition. HH This represents the high-frequency subband with diagonal details after wavelet decomposition, and Normalize() represents the operation function that normalizes the input high-frequency subband.

[0096] By concatenating the low-frequency and high-frequency components along the channel dimension, a multi-resolution feature map is obtained:

[0097]

[0098] Among them, X fusedX represents the multi-resolution features of low-frequency and high-frequency fusion. LL The low-frequency subband obtained by wavelet decomposition is represented by H and W, which represent the spatial height and width of the input features, respectively. R represents the real number space, C represents the number of channels of the input features, and Concat() represents the concatenation operation function along the channel dimension.

[0099] It preserves multi-scale information, suppresses noise with low-frequency components, enhances target edges with high-frequency components, and adapts the downsampling characteristics of wavelet transform to the sparse distribution of small targets.

[0100] Step 2, multimodal attention mechanism, for splicing features X fused Perform global average pooling (GAP) and global max pooling (GMP) respectively to generate channel descriptors:

[0101] GAP(X fused )∈R 4C GMP(X) fused )∈R 4C

[0102] Among them, GAP(X) fused ) represents the channel descriptor after global average pooling, GMP(X) fused The channel descriptor after global max pooling.

[0103] Dynamic channel weights are generated using a fully connected layer (FC):

[0104] A channel =σ(FC(GAP(X) fused )))+FC(GMP(X fused ))

[0105] Where σ represents the Sigmoid function, A channel This represents the dynamic channel weights, and FC() represents a fully connected layer.

[0106] Dynamically weighted according to channel importance to suppress redundant background channels.

[0107] Spatial-frequency attention, for high-frequency components X high Applying 3×3 convolution to generate spatially sensitive masks:

[0108] M spatial =Conv 3×3 (X high )

[0109] Among them, M spatial Represents a spatial attention map, Conv 3×3 This represents a convolution with a kernel size of 3×3, where X... highThis indicates the splicing of normalized high-frequency subbands.

[0110] Frequency attention, for low-frequency component X LL Perform a Fast Fourier Transform (FFT) to calculate the frequency domain energy map:

[0111] E freq =FFT(X) LL )

[0112] A freq =Softmax(Conv1×1(|E freq |))

[0113] Among them, E freq The frequency domain energy map is represented by FFT(), which stands for Fast Fourier Transform. freq Represents frequency attention weights, Conv 1×1 This represents a convolution with a kernel size of 1×1, and Softmax() represents normalization on the channel dimension or a specified dimension.

[0114] Attention fusion combines channel attention with spatial frequency attention in a weighted manner.

[0115]

[0116] Among them, A final This represents the total attention after fusion. This indicates multiplication by channel.

[0117] Step 3: Hybrid pooling strategy, applying attention-weighted feature map X att =X fused ⊙A final Perform two types of pooling:

[0118] Max pooling preserves local extrema (the target pixel is usually a highlight):

[0119] X max =MaxPool(X att )

[0120] Where MaxPool() represents the max pooling operator, X max X represents the feature after max pooling. att This represents attention-weighted features.

[0121] Average pooling smooths background noise:

[0122] X avg =AvgPool(X att )

[0123] Where MaxPool() represents the average pooling operator, X avgX represents the feature after average pooling. att This represents attention-weighted features.

[0124] The splicing result is represented as follows:

[0125]

[0126] Among them, X Pool This represents the concatenated hybrid pooling features. Concat() represents channel-level concatenation, balancing target saliency (MaxPool) and background suppression (AvgPool).

[0127] Step 4, Dynamic Large Kernel Separating Convolution (Improved LSKA), the dynamic kernel selection adaptively selects the convolution kernel size according to the input resolution (for example, a 15×15 kernel is used for a 640×640 input, and a 9×9 kernel is used for a 320×320 input).

[0128] Dilated convolution optimization introduces a dilation rate d=2 in depthwise separable convolution to expand the receptive field:

[0129] X out =DepthwiseConv k×k,d＝2 (X Pool )+PointwiseConv 1×1 (X Pool )

[0130] Where DepthwiseConv represents the depthwise convolution part of a depthwise separable convolution, and PointwiseConv... 1×1 This represents pointwise convolution, where k represents the dynamic kernel size, d represents the dilation rate, and X represents the pointwise convolution. out Represents the output features of dynamic large kernel convolution.

[0131] Output feature map final output Dynamic kernels adapt to targets of different scales, and dilated convolutions reduce the number of parameters while capturing long-range dependencies.

[0132] Please continue reading. Figure 3 The feature fusion neck network comprises a multi-scale input, a lateral alignment unit, a sampling unit, a dynamically gated weighting unit, and a lightweight aggregation unit connected in sequence.

[0133] The multi-scale input is used to receive multiple backbone feature maps with different resolutions output from the feature extraction backbone network in parallel.

[0134] Lateral alignment units are used to unify the number of channels in the feature maps of each backbone.

[0135] The sampling unit is used to upsample deep features in the backbone feature map to medium-shallow resolution, or downsample shallow features in the backbone feature map to medium-deep resolution, so as to obtain multi-source features of the same scale.

[0136] The dynamic gated weighting unit is used to adaptively generate weights based on the global statistics of each backbone feature map, and to weight or selectively superimpose multi-source features of the same scale.

[0137] Lightweight aggregation units are used to perform feature fusion at multiple scales through element-wise summation or channel concatenation, and output multi-scale features for the detection head. The convolutions in the lightweight aggregation units are depthwise separable convolutions and share the same kernel, and reuse the same set of parameters in each scale path.

[0138] In one example, the LDMF module is responsible for dynamically fusing multi-scale features from different layers to enhance the representation of small objects. LDMF provides three input feature maps: shallow feature S, mid-level feature M, and deep feature D. The LDMF module aligns the spatial dimensions of the three input feature maps and concatenates them to form a unified multi-scale feature representation. First, the size of the shallow feature map S is adjusted using a combination of adaptive max pooling and adaptive average pooling to match the spatial dimension of M, i.e., L. resized =P max (S)+P avg (S), where P max (S) and P avg (S) represent adaptive max pooling and adaptive average pooling operations, respectively. This ensures the preservation of low-level spatial details while suppressing background noise. Next, nearest neighbor interpolation is used to upsample the depth features D to the spatial dimension of M. The upsampled D upsampled Represented as in This indicates an upsampling operation. This step ensures that deep semantic information is preserved. Finally, the adjusted feature map S... resized M and D upsampled Connect along the channel dimension to form a fused multi-scale feature map (SMD). in This indicates a connection operation. This process captures fine-grained spatial details and deep semantics, which is crucial for detecting defective targets in complex infrared images.

[0139] Because steel strip defects are small in infrared images, the traditional Intersection Greater Than Union (IoU) metric often encounters situations where the predicted bounding box does not overlap with the ground reality bounding box. In such cases, the IoU value becomes zero. This prevents the loss function from effectively guiding the network in the desired training direction during backpropagation. CIoU successfully overcomes this limitation by incorporating factors such as minimum bounding box and aspect ratio. Furthermore, due to the small size of defects in infrared images, width and height losses play a crucial role in determining the final result. CIoU also considers width and height losses, aspect ratio, and the distance between the centers of the predicted and ground reality bounding boxes. This improves the fit and accuracy. However, CIoU has limited adaptability to objects of different scales, so the Normalized Wasserstein Distance (NWD) loss function, which is insensitive to object size, is introduced. This comprehensively considers the difference between the predicted and ground reality values. The combined loss function ensures detection performance while also exhibiting stronger robustness. The loss function F can be calculated using the following formula:

[0140]

[0141] IoU=(K mn ∩K q ) / (K mn ∪K q )

[0142] Among them, F NWCIoU F represents the loss of the combo box regression, τ represents the weight coefficient, and F represents the weight coefficient. CIoU F represents the CIoU class loss. NWD IoU represents the normalized Wasserstein distance loss, and IoU represents the intersection-union ratio. mn =[x mn ,y mn ] T and j q =[x q ,y q ] T Representing the center coordinates of the actual and predicted bounding boxes respectively, d is the diagonal length of the smallest rectangle enclosing the predicted and actual bounding boxes, ||p gt ,p p || is the Euclidean distance, (w mn ,h mn ) and (w q ,h q K represents the width and height of the real bounding box and the prediction bounding box, respectively. mn and K q These represent the actual and predicted bounding boxes, respectively; IoU represents K. mn and K qThe overlap between them; τ is set to 0.5, θ represents the shape term weight, and E represents the aspect ratio deviation term.

[0143] In this embodiment, a transient temperature difference is created on the surface and subsurface of the steel strip through pulsed thermal excitation. The defect area exhibits significant thermal contrast due to differences in thermal diffusion / heat capacity, thereby enabling early detection of internal hidden dangers. The difference in temperature decay / response over time can suppress background drift and ambient temperature fluctuations, improving defect contrast and stability, thus significantly enhancing the defect detection rate in composite strip scenarios. The pulse excitation time is short and the thermal load is low. Combined with continuous imaging and real-time processing, rapid screening can be completed while the elevator is in operation, avoiding the impact and cost of downtime inspection on operation and maintenance. At the same time, the backbone network extracts high / low layer semantic features, the neck network performs feature pyramid fusion, and the detection head outputs results for features of different scales simultaneously, taking into account both microcracks / small debonding and large-area defects. Using thermal image sequences as input, the model learns dynamic features such as temperature time-varying curves and thermal diffusion rates, thereby better distinguishing between real defects and occasional noise or surface stains.

[0144] Based on the above method, this application discloses a defect detection system for steel-framed elevators with organic rooms based on infrared non-destructive imaging, referencing... Figure 6 The defect detection module for steel-framed elevators with machine rooms includes an image acquisition module 11 and a defect detection module 12.

[0145] The image acquisition module 11 is used to acquire an infrared thermal image sequence formed by the steel strip after being excited by a pulse heat source through infrared camera devices set on both sides of the traction machine. The pulse heat source irradiates the running steel strip with pulses to create a temperature difference between the surface and subsurface of the steel strip.

[0146] The defect detection module 12 is used to input the infrared thermal image sequence into the defect target detection model for processing, so as to output the defect location result corresponding to the infrared thermal image sequence. The defect target detection model includes a feature extraction backbone network, a feature fusion neck network, and a target detection head connected in sequence.

[0147] In one example, the feature extraction backbone network includes a feature extraction module and a multimodal attention space pyramid module. The feature extraction module learns multi-scale semantic and texture features from infrared thermal image sequences from shallow to deep, and forms a basic feature level and a corresponding transition backbone feature map. The multimodal attention space pyramid module enhances and filters the transition backbone feature map to obtain the backbone feature map and highlight the defective responses in the backbone feature map.

[0148] In one example, the feature extraction module includes an input stem, multiple parallel feature extraction sub-columns, a multi-level cross-column reversible fusion unit, and an output aggregation end, wherein...

[0149] The input stem is connected to all feature extraction sub-columns. The input stem is used to receive the input infrared thermal image sequence and perform feature embedding on the infrared thermal image sequence through the first convolutional layer.

[0150] Each feature extraction sub-column, which is set up in parallel, includes multiple levels of downsampling units and Tiny units with 3×3 convolution as the core, arranged in order of depth.

[0151] Multi-level cross-column reversible fusion units are set between corresponding levels of adjacent feature extraction sub-columns. The multi-level cross-column reversible fusion units are used to align the features of each sub-column at the same level and then achieve information interaction through element-by-element summation.

[0152] The output aggregation terminal is connected to all feature extraction sub-columns. The output aggregation terminal is used to fuse the end or intermediate layer features of each feature extraction sub-column and output a transition backbone feature map with multi-scale features.

[0153] In one example, the multimodal attention space pyramid module includes frequency domain branching units, channel attention branching units, hybrid pooling branching units, and fusion and output units connected in sequence.

[0154] The frequency domain branch unit is used to perform two-dimensional discrete wavelet transform on the transition backbone feature map, so as to suppress noise in the transition backbone feature map through low-frequency components and enhance the target edge of the transition backbone feature map through high-frequency components, so as to generate multi-resolution feature map.

[0155] The channel attention branch unit is used to sequentially perform global average pooling and global max pooling on the multi-resolution feature map to generate channel attention and dynamic channel weights corresponding to the channel attention. The channel attention and spatial frequency attention are weighted and fused together to obtain the weighted multi-resolution feature map.

[0156] The hybrid pooling branch unit is used to perform max pooling and average pooling on the weighted multi-resolution feature map to obtain the standard feature map.

[0157] The fusion and output unit is used to perform convolution optimization on the standard feature map based on the dynamic large kernel depth separable convolution to obtain the backbone feature map, wherein the kernel size of the dynamic large kernel depth separable convolution is adaptively selected according to the input resolution.

[0158] In one example, the frequency domain branch unit uses discrete wavelet transform to perform a first-level decomposition of the features corresponding to the transition backbone feature map to obtain LL subband, LH subband, HL subband and HH subband. Among them, the LH subband, HL subband and HH subband are normalized and convolved to enhance the edge response. The LL subband is subjected to fast Fourier transform to obtain the spectral energy map, and a space-frequency attention mask is formed based on the spectral energy map.

[0159] In one example, the feature fusion neck network includes a multi-scale input, a lateral alignment unit, a sampling unit, a dynamically gated weighting unit, and a lightweight aggregation unit connected in sequence.

[0160] The multi-scale input is used to receive multiple backbone feature maps with different resolutions output from the feature extraction backbone network in parallel.

[0161] Lateral alignment units are used to unify the number of channels in the feature maps of each backbone.

[0162] The sampling unit is used to upsample deep features in the backbone feature map to medium-shallow resolution, or downsample shallow features in the backbone feature map to medium-deep resolution, so as to obtain multi-source features of the same scale.

[0163] The dynamic gated weighting unit is used to adaptively generate weights based on the global statistics of each backbone feature map, and to weight or selectively superimpose multi-source features of the same scale.

[0164] Lightweight aggregation units are used to perform feature fusion at multiple scales in an element-wise summation or channel-wise concatenation manner, and output multi-scale features for the detection head.

[0165] In one example, the convolutions in the lightweight aggregation unit employ depthwise separable convolutions and share a kernel, and reuse the same set of parameters across all scale paths.

[0166] Please see Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 2 may include: at least one processor 21, at least one network interface 24, user interface 23, memory 25, and at least one communication bus 22.

[0167] The communication bus 22 is used to enable communication between these components.

[0168] The user interface 23 may include a display screen and a camera. Optionally, the user interface 23 may also include a standard wired interface and a wireless interface.

[0169] The network interface 24 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0170] The processor 21 may include one or more processing cores. The processor 21 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 25, and by calling data stored in the memory 25. Optionally, the processor 21 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 21.

[0171] The memory 25 may include random access memory (RAM) or read-only memory. Optionally, the memory 25 may include non-transitory computer-readable storage medium. The memory 25 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 25 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 25 may also be at least one storage device located remotely from the aforementioned processor 21. Figure 7 As shown, the memory 25, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a defect detection method for steel-framed elevators with organic rooms based on infrared non-destructive imaging.

[0172] exist Figure 7In the electronic device 2 shown, the user interface 23 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 21 can be used to call the application program stored in the memory 25, which is a defect detection method for steel strip elevators with organic rooms based on infrared non-destructive imaging. When executed by one or more processors, the electronic device performs one or more methods as described in the above embodiments.

[0173] A computer-readable storage medium storing instructions that, when executed by one or more processors, cause a computer to perform one or more methods as described in the embodiments above.

[0174] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0175] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0176] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0177] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0178] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0179] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0180] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truths. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure.

[0181] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting defects in organic room steel belt elevators based on infrared non-destructive imaging, characterized by, The method comprises: obtaining an infrared thermal image sequence formed after the steel strip is excited by a pulsed heat source through infrared camera devices arranged on both sides of the traction machine, wherein the pulsed heat source performs pulsed irradiation on the running steel strip to form a temperature difference on the surface and subsurface of the steel strip; inputting the infrared thermal image sequence into a defect target detection model for processing to output a defect positioning result corresponding to the infrared thermal image sequence, wherein the defect target detection model comprises a feature extraction backbone network, a feature fusion neck network and a target detection head connected in sequence.

2. The method for defect detection of organic room steel belt elevator based on infrared non-destructive imaging according to claim 1, characterized in that, The feature extraction backbone network comprises a feature extraction module and a multi-modal attention spatial pyramid module, the feature extraction module learns from the infrared thermal image sequence to form a basic feature level and a transition backbone feature map corresponding to the basic feature level, and the multi-modal attention spatial pyramid module enhances and filters the transition backbone feature map to obtain a backbone feature map and highlight the defect response in the backbone feature map.

3. The method for defect detection of organic room steel belt elevator based on infrared non-destructive imaging according to claim 2, characterized in that, The feature extraction module comprises an input stem, a plurality of parallel feature extraction sub-columns, a multi-stage cross-column reversible fusion unit and an output aggregation end, wherein the input stem is connected with all the feature extraction sub-columns, and the input stem is used to receive the input infrared thermal image sequence and perform feature embedding on the infrared thermal image sequence through a first convolutional layer; each feature extraction sub-column in the plurality of parallel feature extraction sub-columns comprises a plurality of multi-stage down-sampling units and a Tiny unit with 3x3 convolution as the core in sequence by depth; the multi-stage cross-column reversible fusion unit is arranged between the corresponding levels of adjacent feature extraction sub-columns, and is used to align the features of each sub-column at the same level and realize information interaction through element-wise addition; the output aggregation end is connected with all the feature extraction sub-columns, and the output aggregation end is used to fuse the end or intermediate layer features of each feature extraction sub-column and output a transition backbone feature map with multi-scale features.

4. The method for defect detection of organic room steel belt elevator based on infrared non-destructive imaging according to claim 2, characterized in that, The multi-modal attention spatial pyramid module comprises a frequency domain branch unit, a channel attention branch unit, a hybrid pooling branch unit and a fusion and output unit connected in sequence, wherein the frequency domain branch unit is used to perform two-dimensional discrete wavelet transform on the transition backbone feature map to suppress noise of the transition backbone feature map through low-frequency components and to strengthen the target edges of the transition backbone feature map through high-frequency components to generate a multi-resolution feature map; the channel attention branch unit is used to sequentially perform global average pooling and global maximum pooling on the multi-resolution feature map to generate a channel attention and a dynamic channel weight corresponding to the channel attention, to fuse the channel attention and the spatial frequency attention, and to obtain a weighted multi-resolution feature map; the hybrid pooling branch unit is used to perform maximum pooling and average pooling on the weighted multi-resolution feature map to obtain a standard feature map; and the fusion and output unit is used to fuse the standard feature map and the dynamic channel weight to obtain a channel attention feature map, and to fuse the channel attention feature map and the spatial frequency attention to obtain a final feature map. The fusion and output unit is configured to perform convolution optimization on the standard feature map according to a dynamic large kernel depth separable convolution to obtain a backbone feature map, wherein a convolution kernel size of the dynamic large kernel depth separable convolution is adaptively selected according to an input resolution.

5. The method for defect detection of organic room temperature steel belt elevator based on infrared non-destructive imaging according to claim 4, characterized in that, The frequency domain branch unit performs one-level decomposition on the features corresponding to the transition backbone feature map by using a discrete wavelet transform to obtain an LL subband, an LH subband, an HL subband, and an HH subband, wherein the LH subband, the HL subband, and the HH subband are normalized and convolution enhanced edge responses, a fast Fourier transform is performed on the LL subband to obtain a frequency spectrum energy map, and a spatial-frequency attention mask is formed according to the frequency spectrum energy map.

6. The method for defect detection of organic room temperature steel belt elevator based on infrared non-destructive imaging according to claim 2, wherein, The feature fusion neck network comprises a multi-scale input end, a lateral alignment unit, a sampling unit, a dynamic gate weighting unit, and a lightweight aggregation unit connected in sequence, wherein, The multi-scale input end is configured to receive, in parallel, the multiple backbone feature maps with different resolutions output by the feature extraction backbone network; The lateral alignment unit is configured to unify the channel numbers of the backbone feature maps; The sampling unit is configured to up-sample deep features in the backbone feature maps to a medium-shallow resolution, or down-sample shallow features in the backbone feature maps to a medium-deep resolution, to obtain multi-source features of the same scale; The dynamic gate weighting unit is configured to adaptively generate weights according to global statistics of the backbone feature maps, and to weight or selectively superimpose the multi-source features of the same scale; The lightweight aggregation unit is configured to complete feature fusion in multiple scales in a way of element-wise summation or channel splicing, and to output multi-scale features for a detection head.

7. The method for defect detection of organic room temperature steel belt elevator based on infrared non-destructive imaging according to claim 6, characterized in that, The convolution in the lightweight aggregation unit adopts depth separable convolution and shares convolution kernels, and the same set of parameters is reused in each scale path.

8. An organic room temperature steel belt elevator defect detection system based on infrared non-destructive imaging, characterized by, The organic room steel belt elevator defect detection system (1) comprises an image acquisition module (11) and a defect detection module (12), wherein, The image acquisition module (11) is configured to acquire an infrared thermal image sequence formed after the steel belt is excited by a pulse heat source through infrared camera devices arranged on both sides of a traction machine, wherein the pulse heat source irradiates the running steel belt to form a temperature difference on the surface and subsurface of the steel belt; The defect detection module (12) is configured to input the infrared thermal image sequence into a defect target detection model for processing to output a defect positioning result corresponding to the infrared thermal image sequence, wherein the defect target detection model comprises a feature extraction backbone network, a feature fusion neck network, and a target detection head connected in sequence.

9. An electronic device, comprising: The electronic device (2) comprises a processor (21), a memory (25), a user interface (23), and a network interface (24), the memory (25) is configured to store instructions, the user interface (23) and the network interface (24) are configured to communicate with other devices, and the processor (21) is configured to execute the instructions stored in the memory (25) to enable the electronic device (2) to perform the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, which is executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images