A method and device for detecting the production quality of fire-resistant windows

Abstract

The application discloses a fire-resistant window production quality detection method and device, belongs to the technical field of fire-resistant window quality detection, and aims to solve the technical problem that the current quality detection method for fire-resistant window products can damage the detection samples, cannot detect the products one by one, increases the risk of quality hidden danger, and causes certain product loss. The method comprises the following steps: collecting the spectrum diagram and the sound wave detection signal of a plurality of fire-resistant window samples, extracting the spectrum feature of the spectrum diagram to obtain a first quality feature; extracting the signal feature of the sound wave detection signal to obtain a second quality feature; fusing the first quality feature and the second quality feature to obtain the corresponding sample quality feature; determining a model training data set based on the sample quality feature and the corresponding fire resistance test result; training the fire-resistant window quality detection model through the model training data set; and collecting the spectrum diagram and the sound wave detection signal of each fire-resistant window product in real time and inputting the fire-resistant window quality detection model for detection.

Description

Technical Field

[0001] This invention relates to the field of fire-resistant window quality testing technology, and in particular to a method and apparatus for testing the production quality of fire-resistant windows. Background Technology

[0002] The development of fire-resistant windows has a relatively long history in my country, and there is a mature system for research and standards on the fire resistance performance of fire-resistant windows. However, fire-resistant windows are mainly used in interior partition walls such as firewalls and fire-resistant partitions. With the increasing number of high-rise buildings, in order to prevent the rapid spread of fires in high-rise buildings, research on the fire resistance of external partition walls and windows is increasing, and fire integrity requirements have been put forward for external walls and windows. The aim is to alleviate the pressure on emergency rescue capabilities in the event of a fire in a high-rise building in China by strengthening the passive fire defense capabilities of high-rise buildings themselves.

[0003] Because research on fire-resistant windows is still in its early stages, current quality testing of fire-resistant windows mainly relies on manual fire resistance tests, such as combustion tests and other physical methods. This method is not only time-consuming and labor-intensive, requiring significant manpower and testing resources, but it also often damages samples during the testing process, rendering them unsellable as normal products. This makes it impossible to conduct individual quality tests on every fire-resistant window product, forcing the use of sampling inspections. This limitation not only increases the risk of potential quality issues and makes it difficult to ensure the quality and safety of all products, but also increases product wastage, causing economic losses for manufacturers. Therefore, developing a non-destructive, comprehensive, accurate, and reliable method for testing the quality of fire-resistant windows is of paramount importance. Summary of the Invention

[0004] This invention provides a method and apparatus for testing the quality of fire-resistant windows, which addresses the following technical problem: current methods for testing the quality of fire-resistant window products damage the test samples, making it impossible to test each product individually, increasing the risk of potential quality issues, and causing some product loss.

[0005] The embodiments of the present invention adopt the following technical solutions:

[0006] On one hand, embodiments of the present invention provide a method for testing the production quality of fire-resistant windows, the method comprising: collecting spectral images and acoustic detection signals of several fire-resistant window samples, and conducting combustion tests to obtain fire resistance test results;

[0007] Extract the spectral features from the spectrum to obtain the first quality feature;

[0008] The signal features of the acoustic wave detection signal are extracted to obtain the second quality feature;

[0009] The first quality characteristic and the second quality characteristic of the same fire-resistant window sample are fused to obtain the corresponding sample quality characteristic;

[0010] Based on the sample quality characteristics and the corresponding fire resistance test results, the model training dataset is determined;

[0011] A fire-resistant window quality inspection model is constructed, and the fire-resistant window quality inspection model is trained using the model training dataset.

[0012] The spectral data and acoustic signals of each fire-resistant window product are collected in real time on the established fire-resistant window quality inspection line. These signals are then input into the fire-resistant window quality inspection model for testing, and the quality assessment results of each fire-resistant window product are obtained.

[0013] In one feasible implementation, spectral data and acoustic detection signals of several fire-resistant window samples are collected, and combustion tests are performed to obtain fire resistance test results, specifically including:

[0014] In a traditional device for testing the quality of fire-resistant windows through combustion testing, a spectral transceiver and an acoustic transceiver are added; the spectral transceiver and the acoustic transceiver can work simultaneously and are set in the stage before the combustion testing device.

[0015] Before each fire-resistant window sample is subjected to a combustion test, the spectral images of multiple key detection points of the fire-resistant window sample are acquired through the spectral transceiver device; wherein, the spectral transceiver device includes a continuous spectrum emission device and an absorption spectrum collection device.

[0016] Before each fire-resistant window sample is subjected to a combustion test, the acoustic transceiver device acquires acoustic detection signals from multiple key detection points of the fire-resistant window sample; wherein, the acoustic transceiver device includes an ultrasonic transmitter and an ultrasonic receiver.

[0017] After conducting a combustion test on each fire-resistant window sample and obtaining the fire resistance test results, the fire resistance test results were correlated and stored with the corresponding spectral diagrams and acoustic detection signals.

[0018] After collecting a preset number of sample data, data acquisition is stopped and the spectral transceiver and acoustic transceiver are disassembled.

[0019] In one feasible implementation, the spectral features of the spectrum are extracted to obtain a first quality feature, specifically including:

[0020] A basic analysis is performed on each spectrum to obtain basic spectral characteristics; wherein, the basic spectral characteristics include at least: redshift / blueshift characteristics, peak position characteristics, peak width characteristics, peak intensity characteristics, peak shape characteristics, and wavelength range characteristics;

[0021] By using the DB4 wavelet basis in the Daubeeehies wavelet function, n wavelet transforms are performed on each spectral image to obtain the discrete detail signal of the spectral image;

[0022] Wavelet decomposition is performed on the discrete detail signal to obtain wavelet component features;

[0023] The basic spectral features and the wavelet component features constitute the first quality feature.

[0024] In one feasible implementation, the signal features of the acoustic wave detection signal are extracted to obtain the second quality feature, specifically including:

[0025] Perform a Fourier transform on each acoustic wave detection signal to obtain the initial frequency domain features;

[0026] Perform an inverse Fourier transform on the initial frequency domain features to obtain the harmonic wavelet function;

[0027] The acoustic wave detection signal is decomposed into discrete signals by using the harmonic wavelet function.

[0028] The discrete signal is input into a pre-trained convolutional neural network for further feature extraction to obtain the second quality feature.

[0029] In one feasible implementation, the first quality characteristic and the second quality characteristic of the same fire-resistant window sample are fused to obtain the corresponding sample quality characteristics, specifically including:

[0030] Assign initial contribution weight values ​​to the first quality feature and the second quality feature respectively;

[0031] Based on the initial contribution weight value, the fusion ratio of the two feature channels is determined, and feature fusion is performed based on the fusion ratio to obtain the initial sample quality characteristics;

[0032] The fire-resistant window quality detection model is trained using the initial sample quality characteristics and the corresponding fire resistance test results. During the training process, the learning effect of the fire-resistant window quality detection model on the two parts of features and the contribution of the two parts of features to the recognition accuracy are analyzed.

[0033] Based on the learning effect and contribution, the initial contribution weight value is adjusted until the model's recognition accuracy is higher than a preset threshold. Then, based on the contribution weight value at this time, the first quality feature and the second quality feature of all fire-resistant window samples are fused to obtain the final sample quality feature.

[0034] In one feasible implementation, the model training dataset is determined based on the sample quality characteristics and the corresponding fire resistance test results, specifically including:

[0035] Extract the fire resistance test data and fire resistance score for each key test point from the fire resistance test results;

[0036] Based on the fire resistance rating, the key testing points are divided into positive testing points and negative testing points; the positive testing points are those that meet the fire resistance standard, and the negative testing points are those that do not meet the fire resistance standard.

[0037] After associating the sample quality characteristics of the same positive detection point with the fire resistance test data and the fire resistance score, they are stored in the dataset to form a positive sample dataset.

[0038] The sample quality characteristics of the same negative detection point are associated with the fire resistance test data and the fire resistance score, and then stored in the dataset to form a negative sample dataset.

[0039] In one feasible implementation, a fire-resistant window quality inspection model is constructed, and the model is trained using the model training dataset, specifically including:

[0040] The GhostNet network is combined with the YOLOv4 network, and an attention mechanism is introduced to construct the fire-resistant window quality detection model. The input data of the fire-resistant window quality detection model are spectral images and acoustic detection signals, and the output results are the quality assessment results of each key detection point.

[0041] The attention feature fusion network and feature recognition module in the fire-resistant window quality detection model were trained using positive and negative sample datasets, respectively, and the parameters were tuned based on the loss function until the model converged.

[0042] In one feasible implementation, the GhostNet network is combined with the YOLOv4 network, and an attention mechanism is introduced to construct the fire-resistant window quality detection model, specifically including:

[0043] A feature extraction module is constructed based on the Chost model in the GhostNet network and a preset feature extraction algorithm. The feature extraction module includes a first feature extraction module and a second feature extraction module, and the two feature extraction modules operate independently to extract the first quality feature and the second quality feature, respectively.

[0044] A feature fusion module is constructed based on a preset feature fusion algorithm; the output of the overall feature extraction module is the input of the feature fusion module.

[0045] By combining the CBAM attention mechanism, spatial pyramid pooling layer, and attention path aggregation network, an attention feature fusion network is constructed.

[0046] Based on the YOLOv4 model, a feature recognition module is constructed using the YOLO-Head module, and the Focal Loss loss function is incorporated into the feature recognition module.

[0047] The feature extraction module, feature fusion module, input module, attention feature fusion network, and feature recognition module are sequentially connected to form the fire-resistant window quality detection model.

[0048] In one feasible implementation, the spectral data and acoustic wave detection signals of each fire-resistant window product are collected in real time on the constructed fire-resistant window quality inspection production line. These signals are then input into the fire-resistant window quality inspection model for testing, resulting in a quality assessment result for each fire-resistant window product. Specifically, this includes:

[0049] A non-destructive fire-resistant window quality inspection line is constructed based on a spectral transceiver and an acoustic transceiver. A spectral transceiver and an acoustic transceiver are installed at predetermined positions on the fire-resistant window quality inspection line, and the spectral transceiver and the acoustic transceiver operate simultaneously.

[0050] The real-time spectral image of the fire-resistant window product under inspection is obtained through the spectral transceiver; the real-time acoustic detection signal of the fire-resistant window product under inspection is obtained through the acoustic transceiver.

[0051] The real-time spectral image and the real-time acoustic wave detection signal are input into the fire-resistant window quality detection model to obtain the output quality assessment result; wherein, the quality assessment result includes at least: fire resistance prediction data and fire resistance prediction score of each key detection point in the fire-resistant window product to be inspected;

[0052] Based on the quality assessment results, the fire-resistant window product to be inspected is transferred to the corresponding production line branch for quality inspection of the next fire-resistant window product to be inspected; the production line branch includes at least a qualified branch and a failed branch.

[0053] On the other hand, embodiments of the present invention also provide a fire-resistant window production quality inspection device, the device comprising:

[0054] The model training module is used to collect spectral images and acoustic detection signals of several fire-resistant window samples, and conduct combustion tests to obtain fire resistance test results; extract spectral features from the spectral images to obtain a first quality feature; extract signal features from the acoustic detection signals to obtain a second quality feature; fuse the first and second quality features of the same fire-resistant window sample to obtain the corresponding sample quality feature; determine the model training dataset based on the sample quality feature and the corresponding fire resistance test results; construct a fire-resistant window quality detection model, and train the fire-resistant window quality detection model using the model training dataset;

[0055] The quality assessment module is used to collect the spectral images and acoustic detection signals of each fire-resistant window product in real time on the established fire-resistant window quality inspection line, extract the corresponding sample quality characteristics, input them into the fire-resistant window quality detection model for detection, and obtain the quality assessment results of each fire-resistant window product.

[0056] Compared with the prior art, the fire-resistant window production quality testing method and apparatus provided in this embodiment of the invention have the following beneficial effects:

[0057] This invention constructs a non-destructive quality inspection method for fire-resistant windows by combining non-contact detection methods such as spectral detection and acoustic wave detection with a neural network model. It collects a large amount of test data in traditional quality inspection processes and extracts features related to fire resistance results. Then, an improved neural network model is used to deeply learn the relationship between these data features and the fire resistance performance of the windows. This allows for high-precision prediction of the fire resistance performance of windows based on physical properties such as spectral and acoustic wave characteristics during actual quality inspection. This invention eliminates the need for sampling and combustion testing of fire-resistant windows. Combined with the novel quality inspection line proposed in this invention, it enables one-by-one quality inspection of fire-resistant window products, completing the inspection of a single product in a very short time and automatically classifying them as qualified or unqualified products, greatly improving inspection efficiency. This invention not only increases the comprehensiveness of fire-resistant window quality inspection but also reduces inspection waste, demonstrating its high practical value. Attached Figure Description

[0058] 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0059] Figure 1 A flowchart of a quality inspection method for fire-resistant window production is provided in an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of a fire-resistant window production quality testing device provided in an embodiment of the present invention. Detailed Implementation

[0061] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0062] This invention provides a method for quality inspection in the production of fire-resistant windows, such as... Figure 1 As shown, the specific steps of the fire-resistant window production quality inspection method include steps S101-S106:

[0063] S101. Collect the spectral data and acoustic detection signals of several fire-resistant window samples, and conduct combustion tests to obtain the fire resistance test results.

[0064] Specifically, a spectral transceiver and an acoustic transceiver are added to the traditional device for testing the quality of fire-resistant windows through combustion testing. The spectral transceiver and the acoustic transceiver can operate simultaneously and are positioned before the combustion testing device.

[0065] Furthermore, before each fire-resistant window sample is subjected to combustion testing, spectral images of multiple key detection points on the fire-resistant window sample are acquired using a spectral transceiver. This spectral transceiver includes a continuous spectrum emission device and an absorption spectrum collection device.

[0066] Before each fire-resistant window sample is subjected to a combustion test, acoustic detection signals are acquired at multiple key detection points of the fire-resistant window sample using an acoustic transceiver device. This device includes an ultrasonic transmitter and an ultrasonic receiver.

[0067] The two data collection processes described above can be performed simultaneously without affecting each other.

[0068] Furthermore, after conducting combustion tests on each fire-resistant window sample and obtaining the fire resistance test results, the fire resistance test results were correlated and stored with the corresponding spectral graphs and acoustic detection signals. After collecting a preset number of sample data, data acquisition was stopped and the spectral transceiver and acoustic transceiver were removed.

[0069] As a feasible implementation method, in the existing quality inspection process for fire-resistant windows, spectral and acoustic tests are performed on the samples before the combustion test. A continuous spectrum is emitted to the sample using a continuous spectrum emitting device; the aluminum alloy material and other components of the sample absorb a portion of the spectrum, which is then collected by a corresponding absorption spectrum collecting device. Ultrasonic waves are also emitted using an ultrasonic emitting device; these waves penetrate the sample and are received by a corresponding ultrasonic receiving device. Since the continuous spectrum and ultrasonic waves undergo certain changes after passing through the sample, this invention analyzes the relationship between the characteristics of these changes and the fire resistance performance, enabling a neural network model to learn this correlation. This allows for the measurement of the fire resistance performance of the fire-resistant window using non-contact, non-destructive physical quantities.

[0070] Furthermore, this invention does not collect data through combustion experiments on a large number of fire-resistant windows. Instead, it collects data by installing the device within the normal fire-resistant window quality inspection process. This avoids excessive sample waste and does not hinder or affect the normal quality inspection process. After collecting sufficient data, the device can be removed to restore the original state.

[0071] S102. Extract the spectral features of the spectrum to obtain the first quality feature; extract the signal features of the acoustic wave detection signal to obtain the second quality feature.

[0072] Specifically, a basic analysis is performed on each spectrum to obtain basic spectral characteristics; these basic spectral characteristics include at least: redshift / blueshift characteristics, peak position characteristics, peak width characteristics, peak intensity characteristics, peak shape characteristics, and wavelength range characteristics.

[0073] Furthermore, using the DB4 wavelet basis in the Daubeehies wavelet function, n wavelet transforms are performed on each spectrum to obtain discrete detail signals. Wavelet decomposition is then performed on these discrete detail signals to obtain wavelet component features. The basic spectral features and the wavelet component features are combined to form the first quality feature.

[0074] Furthermore, a Fourier transform is performed on each acoustic wave detection signal to obtain initial frequency domain features. An inverse Fourier transform is then performed on the initial frequency domain features to obtain harmonic wavelet functions. These harmonic wavelet functions are then used to decompose the acoustic wave detection signals into frequency bands, resulting in discrete signals. These discrete signals are then input into a pre-trained convolutional neural network for further feature extraction, yielding the second quality features.

[0075] S103. The first quality characteristic and the second quality characteristic of the same fire-resistant window sample are fused to obtain the corresponding sample quality characteristics.

[0076] Specifically, initial contribution weight values ​​are assigned to the first quality feature and the second quality feature, respectively. Based on the initial contribution weight values, the fusion ratio of the two feature channels is determined, and feature fusion is performed based on the fusion ratio to obtain the initial sample quality features.

[0077] Furthermore, a fire-resistant window quality detection model was trained using the initial sample quality characteristics and the corresponding fire resistance test results. During the training process, the learning effect of the fire-resistant window quality detection model on the two parts of features and the contribution of the two parts of features to the recognition accuracy were analyzed.

[0078] Furthermore, based on the learning effect and contribution, the initial contribution weight value is adjusted until the model's recognition accuracy is higher than the preset threshold. Then, based on the contribution weight value at this time, the first quality feature and the second quality feature of all fire-resistant window samples are fused to obtain the final sample quality feature.

[0079] As a feasible implementation method, the initial contribution weight values ​​can all be set to 1. During subsequent training, the contribution weights are dynamically adjusted by analyzing the contribution and correlation between the two features and the test results until the detection accuracy reaches a relatively high preset value. Then, the contribution weights of the two features at this point are obtained and applied to the feature fusion of all samples to obtain the final sample quality features. Thus, the first and second quality features are fused based on their contribution values ​​to the detection results, maximizing the effectiveness of both features.

[0080] S104. Based on the sample quality characteristics and the corresponding fire resistance test results, determine the model training dataset.

[0081] Specifically, the fire resistance test data and fire resistance score for each key testing point are extracted from the fire resistance test results. This fire resistance score is the rating given by the testers based on the fire resistance performance during the combustion test.

[0082] Furthermore, based on the fire resistance rating, key testing points are divided into positive and negative testing points. Positive testing points are those that meet the fire resistance standards, while negative testing points are those that do not.

[0083] Furthermore, the sample quality characteristics of the same positive detection point are correlated with the fire resistance test data and fire resistance score, and then stored in the dataset to form a positive sample dataset. Similarly, the sample quality characteristics of the same negative detection point are correlated with the fire resistance test data and fire resistance score, and then stored in the dataset to form a negative sample dataset.

[0084] S105. Construct a fire-resistant window quality inspection model and train the fire-resistant window quality inspection model using the model training dataset.

[0085] Specifically, a fire-resistant window quality inspection model is constructed by combining the GhostNet network with the YOLOv4 network and introducing an attention mechanism. The input data for this model consists of spectral data and acoustic detection signals, while the output is the quality assessment results for each key inspection point.

[0086] Furthermore, the attention feature fusion network and feature recognition module in the fire-resistant window quality detection model were trained using positive and negative sample datasets, respectively, and the parameters were tuned based on the loss function until the model converged.

[0087] As a feasible implementation method, the fire-resistant window quality detection model constructed in this invention is an improvement on the YOLOv4 network model, specifically including:

[0088] Based on the Chost model in the GhostNet network and a pre-defined feature extraction algorithm, a feature extraction module is constructed. This module comprises a first feature extraction module and a second feature extraction module, which operate independently to extract first-quality features and second-quality features, respectively. Then, based on a pre-defined feature fusion algorithm, a feature fusion module is constructed; the output of the overall feature extraction module serves as the input to the feature fusion module. Next, the CBAM attention mechanism, spatial pyramid pooling layer, and attention path aggregation network are combined to construct an attention-based feature fusion network. Furthermore, based on the YOLOv4 model's YOLO-Head module, a feature recognition module is constructed, incorporating the Focal Loss function.

[0089] Finally, the feature extraction module, feature fusion module, input module, attention feature fusion network, and feature recognition module are connected in sequence to form a fire-resistant window quality detection model.

[0090] S106. On the constructed fire-resistant window quality inspection line, the spectral diagrams and acoustic detection signals of each fire-resistant window product are collected in real time, and input into the fire-resistant window quality inspection model for testing to obtain the quality assessment results of each fire-resistant window product.

[0091] Specifically, a non-destructive fire-resistant window quality inspection line is built based on a spectral transceiver and an acoustic transceiver. The spectral transceiver and the acoustic transceiver are installed at preset positions on the fire-resistant window quality inspection line and are positioned opposite each other. The spectral transceiver and the acoustic transceiver work simultaneously.

[0092] Furthermore, the real-time spectral image of the fire-resistant window product under inspection is obtained through a spectral transceiver; and the real-time acoustic detection signal of the fire-resistant window product under inspection is obtained through an acoustic transceiver.

[0093] Furthermore, the real-time spectral image and the real-time acoustic wave detection signal are input into the fire-resistant window quality detection model to obtain the output quality assessment results; wherein, the quality assessment results include at least: the fire resistance prediction data and fire resistance prediction score of each key detection point in the fire-resistant window product to be inspected.

[0094] Furthermore, based on the quality assessment results, the fire-resistant window product to be inspected is transferred to the corresponding production line branch for quality inspection of the next fire-resistant window product to be inspected; the production line branch includes at least a qualified branch and a failed branch.

[0095] As a feasible implementation method, this invention designs a quality inspection line that matches the provided fire-resistant window production quality inspection method. Fire-resistant window products sequentially pass through a data acquisition device in the line. The collected data is then transmitted to a fire-resistant window quality inspection model for quality testing. This processing time is very short, only a few seconds or even milliseconds are needed to determine whether the fire-resistant window is qualified. If qualified, the fire-resistant window product is conveyed to a qualified branch line for further labeling, packaging, and other operations. If unqualified, it is conveyed to a unqualified branch line for further processing.

[0096] Because this invention tests the fire resistance of several key inspection points on fire-resistant windows, after a substandard product is detected, the type and location of the key inspection points showing the substandard results can be used to determine whether the product can be repaired or modified for normal use. If so, further repair can be carried out without the need for destruction and reconstruction, saving production resources. For example, if several inspection points on the glass of a fire-resistant window fail to meet fire resistance standards, the product can be repaired by replacing the glass. After repair, a second quality inspection can be conducted; if it passes, the product can be sold normally.

[0097] In addition, embodiments of the present invention also provide a fire-resistant window production quality testing device, such as... Figure 2 As shown, the fire-resistant window production quality inspection device 200 specifically includes:

[0098] Model training module 210 is used to collect spectral images and acoustic detection signals of several fire-resistant window samples, and conduct combustion tests to obtain fire resistance test results; extract spectral features from the spectral images to obtain a first quality feature; extract signal features from the acoustic detection signals to obtain a second quality feature; fuse the first quality feature and the second quality feature of the same fire-resistant window sample to obtain the corresponding sample quality feature; determine the model training dataset based on the sample quality feature and the corresponding fire resistance test results; construct a fire-resistant window quality detection model, and train the fire-resistant window quality detection model using the model training dataset;

[0099] The quality assessment module 220 is used to collect the spectral images and acoustic detection signals of each fire-resistant window product in real time on the established fire-resistant window quality inspection line, extract the corresponding sample quality characteristics, input them into the fire-resistant window quality detection model for detection, and obtain the quality assessment results of each fire-resistant window product.

[0100] As one feasible implementation, the device also includes a quality inspection line module 230, comprising a spectral transceiver, an acoustic transceiver, a main conveyor belt, and branch conveyor belts. This is used for acquiring spectral data and acoustic signals from fire-resistant window products, as well as for quality inspection.

[0101] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0102] The foregoing has described specific embodiments of the present invention. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0103] The above description is merely an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for quality inspection in the production of fire-resistant windows, characterized in that, The method includes: Spectroscopic images and acoustic detection signals of several fire-resistant window samples were collected, and combustion tests were conducted to obtain the fire resistance test results, specifically including: In a traditional device for quality testing of fire-resistant windows through combustion testing, a spectral transceiver and an acoustic transceiver are added. These two devices can operate simultaneously and are positioned before the combustion testing device. Before each fire-resistant window sample is tested for combustion, the spectral transceiver acquires the spectra of multiple key detection points on the sample. The spectral transceiver includes a continuous spectrum transmitter and an absorption spectrum collector. Before each fire-resistant window sample is tested for combustion, the acoustic transceiver acquires the acoustic detection signals of multiple key detection points on the sample. The acoustic transceiver includes an ultrasonic transmitter and an ultrasonic receiver. After each fire-resistant window sample undergoes combustion testing and the fire resistance test results are obtained, the fire resistance test results are associated and stored with the corresponding spectral maps and acoustic detection signals. After collecting a preset number of sample data, data acquisition is stopped, and the spectral transceiver and acoustic transceiver are disassembled. Extracting the spectral features of the spectrum to obtain the first quality feature specifically includes: A fundamental analysis is performed on each spectrum to obtain basic spectral features. These basic spectral features include at least: redshift / blueshift features, peak position features, peak width features, peak intensity features, peak shape features, and wavelength range features. Using the DB4 wavelet basis in the Daubeehies wavelet function, n wavelet transforms are performed on each spectrum to obtain discrete detail signals. Wavelet decomposition is performed on these discrete detail signals to obtain wavelet component features. The basic spectral features and the wavelet component features are combined to form the first quality feature. Extracting the signal features of the acoustic wave detection signal to obtain the second quality feature specifically includes: Perform a Fourier transform on each acoustic wave detection signal to obtain initial frequency domain features; perform an inverse Fourier transform on the initial frequency domain features to obtain a harmonic wavelet function; decompose the acoustic wave detection signal into frequency bands using the harmonic wavelet function to obtain a discrete signal; input the discrete signal into a pre-trained convolutional neural network for further feature extraction to obtain the second quality feature; The first and second quality characteristics of the same fire-resistant window sample are fused to obtain the corresponding sample quality characteristics, which specifically include: Initial contribution weight values ​​are assigned to the first quality feature and the second quality feature respectively; based on the initial contribution weight values, the fusion ratio of the two feature channels is determined, and feature fusion is performed based on the fusion ratio to obtain the initial sample quality features; a fire-resistant window quality detection model is trained using the initial sample quality features and the corresponding fire resistance test results, and the learning effect of the fire-resistant window quality detection model on the two features and the contribution of the two features to the recognition accuracy are analyzed during the training process; according to the learning effect and contribution, the initial contribution weight values ​​are adjusted until the recognition accuracy of the model is higher than a preset threshold, and then the first quality feature and the second quality feature of all fire-resistant window samples are fused according to the contribution weight values ​​at this time to obtain the final sample quality features; Based on the sample quality characteristics and the corresponding fire resistance test results, the model training dataset is determined, specifically including: Extract the fire resistance test data and fire resistance score for each key detection point from the fire resistance test results; based on the fire resistance score, divide the key detection points into positive detection points and negative detection points; positive detection points are those that meet the fire resistance standard, and negative detection points are those that do not meet the fire resistance standard; after associating the sample quality characteristics of the same positive detection point with the fire resistance test data and the fire resistance score, store them in a dataset to form a positive sample dataset; after associating the sample quality characteristics of the same negative detection point with the fire resistance test data and the fire resistance score, store them in a dataset to form a negative sample dataset; Constructing a fire-resistant window quality inspection model and training the model using the model training dataset specifically includes: The GhostNet network is combined with the YOLOv4 network, and an attention mechanism is introduced to construct the fire-resistant window quality detection model, specifically including: A feature extraction module is constructed based on the Chost model in the GhostNet network and a preset feature extraction algorithm. The feature extraction module includes a first feature extraction module and a second feature extraction module, and the two feature extraction modules operate independently to extract the first quality feature and the second quality feature, respectively. A feature fusion module is constructed based on a preset feature fusion algorithm; the output of the feature extraction module is the input of the feature fusion module; an attention feature fusion network is constructed by combining the CBAM attention mechanism, spatial pyramid pooling layer, and attention path aggregation network; a feature recognition module is constructed based on the YOLOv4 model's YOLO-Head module, and the Focal Loss loss function is incorporated into the feature recognition module; the feature extraction module, feature fusion module, input module, attention feature fusion network, and feature recognition module are sequentially connected to form the fire-resistant window quality detection model. The input data of the fire-resistant window quality detection model are spectral diagrams and acoustic detection signals, and the output results are the quality assessment results of each key detection point. The attention feature fusion network and feature recognition module in the fire-resistant window quality detection model were trained using positive and negative sample datasets, respectively, and the parameters were tuned based on the loss function until the model converged. The spectral data and acoustic signals of each fire-resistant window product are collected in real time on the established fire-resistant window quality inspection line. These signals are then input into the fire-resistant window quality inspection model for testing, resulting in quality assessment results for each fire-resistant window product. Specifically, these results include: A non-destructive fire-resistant window quality inspection line is constructed based on a spectral transceiver and an acoustic transceiver. A spectral transceiver and an acoustic transceiver are installed at predetermined positions on the fire-resistant window quality inspection line, and the spectral transceiver and the acoustic transceiver operate simultaneously. The real-time spectral image of the fire-resistant window product under inspection is obtained through the spectral transceiver; the real-time acoustic detection signal of the fire-resistant window product under inspection is obtained through the acoustic transceiver. The real-time spectral image and the real-time acoustic wave detection signal are input into the fire-resistant window quality detection model to obtain the output quality assessment result; wherein, the quality assessment result includes at least: fire resistance prediction data and fire resistance prediction score of each key detection point in the fire-resistant window product to be inspected; Based on the quality assessment results, the fire-resistant window product to be inspected is transferred to the corresponding production line branch for quality inspection of the next fire-resistant window product to be inspected; the production line branch includes at least a qualified branch and a failed branch. After a defective product is detected, the type and location of the key testing points that caused the defect are used to determine whether the product can be repaired or modified to be used normally. If so, further repairs are carried out, and the product is inspected again after repair. If it passes the inspection, it can be sold normally.

2. A fire-resistant window production quality inspection device, employing the fire-resistant window production quality inspection method as described in claim 1, characterized in that, The device includes: The model training module is used to collect spectral images and acoustic detection signals of several fire-resistant window samples, and conduct combustion tests to obtain fire resistance test results; extract spectral features from the spectral images to obtain a first quality feature; extract signal features from the acoustic detection signals to obtain a second quality feature; fuse the first and second quality features of the same fire-resistant window sample to obtain the corresponding sample quality feature; determine the model training dataset based on the sample quality feature and the corresponding fire resistance test results; construct a fire-resistant window quality detection model, and train the fire-resistant window quality detection model using the model training dataset; the quality assessment module is used to collect spectral images and acoustic detection signals of each fire-resistant window product in real time on the established fire-resistant window quality inspection line, extract the corresponding sample quality features, input them into the fire-resistant window quality detection model for detection, and obtain the quality assessment results of each fire-resistant window product.

Images