Material oxygen index prediction method and device based on machine vision and deep learning

Abstract

The application provides a material oxygen index prediction method and device based on machine vision and deep learning, which comprises sampling according to a material standard and preparing one sample; vertically fixing the sample in a combustion cylinder, the combustion cylinder having a single gas source with a fixed oxygen concentration flowing upward, igniting the top end of the sample; collecting video images of the test process; constructing an image detection model based on machine vision, inputting the video images into the trained image detection model, detecting the combustion position in the video images, obtaining a sample combustion length sequence and a sample combustion image sequence; constructing an oxygen index prediction model based on deep learning, inputting the sample combustion length sequence and the sample combustion image sequence into the oxygen index prediction model for reasoning, and obtaining a predicted value of the oxygen index. The application is suitable for rapid testing and screening of the material oxygen index in a non-laboratory environment, and greatly improves the efficiency and convenience of the material oxygen index screening.

Description

Technical Field

[0001] This invention belongs to the field of oxygen index measurement technology, and in particular relates to a method and device for predicting the oxygen index of materials based on machine vision and deep learning. Background Technology

[0002] The oxygen index (OI) is the minimum oxygen concentration required to sustain combustion of a material when a mixture of oxygen and nitrogen is introduced at 23°C ± 2°C. It is expressed as a volume fraction. A high OI indicates that the material is not easily combustible, while a low OI indicates that the material is easily combustible.

[0003] The principle of oxygen index determination is to fix a sample vertically in a transparent combustion tube containing an upward-flowing mixture of oxygen and nitrogen gas, ignite the top of the sample, observe the combustion characteristics of the sample, compare the continuous combustion time or combustion length of the sample with a given criterion, and estimate the minimum oxygen concentration that can sustain the combustion of the material through a series of tests at different oxygen concentrations.

[0004] The existing oxygen index determination is carried out in accordance with national / international standards, using calibrated specialized instruments and equipment to evaluate the combustion characteristics of materials. The procedures, calculation and expression of results, and calibration of equipment are all clearly defined.

[0005] However, existing national / international standard testing methods require two gas sources: oxygen and nitrogen / air (nitrogen is generally used), a gas mixer, and a gas concentration measurement and adjustment device (for measuring and adjusting oxygen concentration). Therefore, they are bulky and complex, making it difficult to miniaturize and port them for use in non-laboratory environments.

[0006] Meanwhile, the experiment requires a step-by-step selection of oxygen concentration using a "small sample increase-decrease method." Determining the initial oxygen concentration and changing its value necessitates repeated trials at different oxygen concentrations (repeatedly lowering or increasing the oxygen concentration based on combustion conditions, followed by re-observing the combustion process). Furthermore, the initial oxygen concentration must be selected based on results from similar materials or by observing the ignition of the sample in air. The initial oxygen concentration can be determined in any suitable step size, thus involving a degree of subjectivity and requiring the experience of the experimenter to determine the initial oxygen concentration quickly and accurately. Changing the oxygen concentration also requires multiple repeated trials, extensive data recording, and calculations, making the experimental process lengthy and cumbersome.

[0007] The above problems make it difficult to measure the oxygen index in non-laboratory environments (such as construction sites or fire accident sites). Relevant personnel have to take samples and send them to laboratories for testing, but laboratory testing cycles are often long and cannot meet the needs of rapid screening. Summary of the Invention

[0008] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes a method and device for predicting the oxygen index of materials based on machine vision and deep learning. The method extracts the combustion characteristics of materials under a fixed oxygen concentration through machine vision and inputs the combustion characteristics into a deep learning model to predict the oxygen index of materials.

[0009] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0010] A method for predicting the oxygen index of materials based on machine vision and deep learning, including

[0011] Sampling was performed according to the material standards, and one sample was prepared.

[0012] The sample is vertically fixed inside a combustion chamber with a single upward-flowing gas source of fixed oxygen concentration, which ignites the top of the sample.

[0013] Video images of the experimental process are acquired using an image acquisition and detection device;

[0014] A machine vision-based image detection model is constructed. Video images are input into the trained image detection model to detect the combustion location in the video images and obtain the sample combustion length sequence and sample combustion image sequence.

[0015] The sample combustion length sequence and sample combustion image sequence are used as input data to the constructed deep learning-based oxygen index prediction model for inference, and the predicted value of oxygen index is obtained.

[0016] Furthermore, the single gas source for the fixed oxygen concentration is a standard concentration gas premixed with oxygen and nitrogen in a gas cylinder.

[0017] Furthermore, the image detection model includes:

[0018] Input module, Backbone network, Neck module, Head module;

[0019] The training process of an image detection model includes:

[0020] S1: Use an image acquisition and detection device to acquire video images of the sample combustion process, label and convert the acquired video image data, and divide the processed dataset into training set and test set;

[0021] S2: Input the data from the training set into the input of the image detection model for preprocessing;

[0022] S3: Input the preprocessed image into the Backbone backbone network to obtain deep feature information and obtain feature maps of different depths;

[0023] S5: Input feature maps of different depths into the Neck module for upsampling and feature fusion to obtain tensor data of different scales;

[0024] S6: Select the correct detection box based on the predicted box and the ground box, and use the non-maximum suppression (NMS) method to remove redundant predicted boxes;

[0025] S7: Calculate the model's loss function based on the detection results, adjust the model parameters, and complete the model training when the loss function value fluctuates or no longer decreases.

[0026] Furthermore, the detection of the combustion location in the video image and the acquisition of the sample combustion length sequence and the sample combustion image sequence include:

[0027] Let t be the time number, ranging from t1 to t2. n t1 indicates the start of combustion; as the first sample, t n Indicates the last data collection;

[0028] When the image changes from one detection box to two detection boxes, the vertical coordinate of the bottom pixel of the flame detection box is recorded as the coordinate of the combustion start position c(t1), and the sample combustion length x(t1) = c(t1) - c(t1) = 0 is obtained.

[0029] During the experiment, the vertical coordinate of the pixel at the bottom of the flame detection frame was continuously detected according to a preset sampling frequency as the combustion position c(t), and the combustion length sequence of the sample was obtained as x(t) = c(t) - c(t1). When the preset number of samplings was reached, the vertical coordinate of the bottom of the detection frame was recorded as the combustion end position coordinate c(t). n The sample combustion length x(t) was obtained. n )=c(t n )-c(t1), then stop the detection;

[0030] During the data collection process, if a detection frame disappears, the ordinate of the bottom of the last detection frame before it disappears is recorded as the coordinate c(t) of the end of combustion position. e The sample combustion length x(t) was obtained. e )=c(t e )-c(t1); thereafter, if the sample does not reignite, then x(t)=x(t) e If the sample reignites, continue to detect the burning location in the video image using the aforementioned method.

[0031] While obtaining the sample combustion length sequence x(t), the video image at the same moment is captured as the sample combustion image sequence p(t).

[0032] Furthermore, methods for constructing oxygen index prediction models based on deep learning include:

[0033] (1) Obtain the true oxygen index of the sample by performing oxygen index test according to standard methods and using standard equipment, collect sample burning length sequence and sample burning image sequence, and construct oxygen index sample dataset.

[0034] (2) Construct an oxygen index prediction model, which includes an image data processing module and a time series data processing module;

[0035] The image data processing module is used to process spatial image data, extract the features of the data in the spatial domain, and convert the sample combustion image sequence into an image feature sequence, including a convolutional and deconvolutional neural network architecture with residual network and feature pyramid network.

[0036] The time series data processing module is used to process time series data, extract the features of the data in the time domain, splice the image feature sequence with the sample combustion length sequence and predict the oxygen index, including a decoder network architecture based on attention mechanism units and long short-term memory units.

[0037] (3) Determine the parameters of the image data processing module and the time-series data processing module;

[0038] (4) Train the oxygen index prediction model using the constructed oxygen index sample dataset, including:

[0039] A method for predicting future combustion images by inputting current combustion images of a sample involves pre-training the backbone network of the image data processing module and adjusting the model weights by calculating the loss function and backpropagating.

[0040] By fixing the weights of the backbone network of the image data processing module, the method of predicting the oxygen index of the sample by inputting the oxygen index sample dataset into the model is used to train the feature pyramid network in the image data processing module and the attention mechanism unit and long short-term memory unit in the time series data processing module. The model weights are adjusted by calculating the loss function and backpropagating.

[0041] The final model parameters are determined to obtain the final oxygen index prediction model.

[0042] Compared with existing technologies, the material oxygen index prediction method based on machine vision and deep learning described in this invention has the following advantages:

[0043] (1) By designing a single gas source with a fixed oxygen concentration, this invention eliminates the two gas sources, gas mixer, and gas concentration measurement and adjustment device in existing equipment, thereby solving the problems of large equipment size and complex structure, and making the equipment miniaturized and portable for use in non-laboratory environments.

[0044] (2) This invention avoids the tedious and repetitive steps of gradually selecting oxygen concentration based on the "small sample increase-decrease method" in the experiment, and does not require the experience of the test personnel to select the initial oxygen concentration and the step size of oxygen concentration change. It also avoids the recording and calculation of a large amount of test data, which greatly reduces the threshold for screening the oxygen index of materials.

[0045] (3) It facilitates rapid testing and screening of material oxygen index in non-laboratory environments, greatly improving the efficiency and convenience of material oxygen index screening.

[0046] This invention also provides a material oxygen index prediction device based on machine vision and deep learning, including an experimental device and an image acquisition and detection device.

[0047] The test apparatus includes a base, a combustion cylinder, a sample holder, a diffuser ring, a metal mesh, an igniter, and a gas source channel. The bottom end of the combustion cylinder is inserted into the base. The base is equipped with a metal mesh and a diffuser ring arranged vertically. The gas source channel is connected to the diffuser ring. The sample holder includes a support device and a clamp.

[0048] The image acquisition and detection device includes an optical camera and a data processing and storage device that communicates with the optical camera. The relative position and angle between the optical camera and the sample are fixed.

[0049] Furthermore, the data processing and storage device includes any one or more of a computer, a single-board computer, and a cloud server.

[0050] Furthermore, it also includes a light shield and a monitor. The light shield is used to cover the area around the test device to prevent external light sources from interfering with the optical camera's acquisition of the combustion flame. The monitor is connected to the optical camera and is used to observe the combustion scene.

[0051] Compared to existing technologies, the material oxygen index prediction device based on machine vision and deep learning described in this invention has the same beneficial effects as the aforementioned material oxygen index prediction method based on machine vision and deep learning, and will not be repeated here. Attached Figure Description

[0052] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0053] Figure 1 This is a schematic diagram of the structure of a material oxygen index prediction device based on machine vision and deep learning according to the present invention.

[0054] Figure 2 This is a schematic diagram of the detection frame of the present invention;

[0055] Figure 3 This is a schematic diagram of the oxygen index prediction model structure of the present invention;

[0056] Figure 4 This is a schematic diagram of the image data processing module of the present invention;

[0057] Figure 5 This is a schematic diagram of the attention mechanism unit of the present invention;

[0058] Figure 6 This is a schematic diagram of the long short-term memory unit of the present invention.

[0059] Explanation of reference numerals in the attached figures

[0060] 1-Base; 2-Combustion cylinder; 3-Sample; 4-Sample clamp; 5-Diffuser ring; 6-Metal mesh; 7-Igniter; 8-Gas source channel; 9-Optical camera; 10-Data processing and storage device. Detailed Implementation

[0061] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0062] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0063] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0064] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0065] This invention provides a method for predicting the oxygen index of materials based on machine vision and deep learning, including...

[0066] Step 1: Take samples according to the material standards and prepare one specimen;

[0067] Specifically, the present invention prepares samples according to the method in the national standard GB / T 2406.2-2009, including sampling, preparation, marking, and condition adjustment. The national standard requires the preparation of 15-30 samples, while the method of the present invention only requires the preparation of 1 sample, thus improving the detection efficiency.

[0068] Step 2: Vertically fix the sample inside a combustion chamber with a single upward-flowing gas source of fixed oxygen concentration, and ignite the top of the sample;

[0069] Specifically, this invention uses a single gas source with a fixed oxygen concentration. The single gas source with a fixed oxygen concentration is a standard concentration gas premixed with oxygen and nitrogen in a gas cylinder. The selection of the oxygen concentration ensures that the test sample can spontaneously combust and produce a flame.

[0070] Step 3: Use an image acquisition and detection device to acquire video images of the test process;

[0071] Specifically, this invention ignites the sample according to the method of national standard GB / T 2406.2-2009 (under a fixed oxygen concentration), records video images of the test process with an optical camera, and uses a camera with fixed parameters (such as pixels, focal length, field of view, etc.) to take horizontal pictures of the test process at a fixed relative distance, relative height, and relative angle (the center of the image is horizontally aligned with the center of the sample), thereby ensuring that the size of the sample in the image is in a fixed proportional relationship with the size of the sample in reality.

[0072] Step 4: Construct a machine vision-based image detection model. Input the video image into the trained image detection model to detect the combustion location in the video image and obtain the sample combustion length sequence and sample combustion image sequence.

[0073] Specifically, the present invention uses machine vision algorithms to process the above video images to obtain and store the combustion data of the test sample;

[0074] The machine vision algorithm is as follows:

[0075] The video images of the burning sample are input into a trained image detection model to detect the burning location in the video images. The burning length sequence x(t) of the sample is obtained according to a preset acquisition frequency and a preset number of acquisitions. At the same time as obtaining the burning length sequence x(t), images of the same moment in the video are acquired as the burning image sequence p(t) of the sample.

[0076] The image detection model includes: an input end, a backbone network, a neck module, and a head module;

[0077] The training process of the image detection model includes the following steps:

[0078] S1: Use an image acquisition and detection device to acquire video images of the sample combustion process, label and convert the acquired video image data, and divide the converted dataset into a training set and a test set.

[0079] S2: Input the data from the training set into the input of the image detection model for preprocessing;

[0080] S3: Input the preprocessed image into the Backbone backbone network to obtain deep feature information and obtain feature maps of different depths;

[0081] S5: Input feature maps of different depths into the Neck module for upsampling and feature fusion to obtain tensor data of different scales;

[0082] S6: Select the correct detection box based on the predicted box and the ground box, and use the non-maximum suppression (NMS) method to remove redundant predicted boxes;

[0083] S7: Calculate the model's loss function based on the detection results, adjust the model parameters, and complete the model training when the loss function value fluctuates or no longer decreases.

[0084] The combustion location in the video image is detected, and the sample combustion length sequence x(t) is obtained according to the preset acquisition frequency and preset number of acquisitions, including:

[0085] Let t be the time number, and its value range from t1 to t2. n t1 represents the start of combustion, i.e., the first collection, t n This represents the last data collection;

[0086] The instant when a detection box in an image changes from one detection box to two detection boxes (i.e., a flame in the image changes to two flames), such as Figure 2 As shown, the vertical coordinate of the pixel at the bottom of the flame detection frame is recorded as the coordinate of the combustion start position c(t1), and the combustion length of the sample x(t1) = c(t1) - c(t1) = 0 is obtained.

[0087] During the experiment, the vertical coordinate of the pixel at the bottom of the flame detection frame was continuously detected according to a preset sampling frequency as the combustion position c(t), resulting in the sample combustion length sequence x(t) = c(t) - c(t1). When the preset number of samplings was reached, the vertical coordinate of the bottom of the detection frame was recorded as the combustion end position coordinate c(t1). n The sample combustion length x(t) was obtained.n )=c(t n )-c(t1), then stop the detection.

[0088] During the data collection process, if the detection frame disappears (the flame goes out), the ordinate of the bottom of the last detection frame before it disappears is recorded as the coordinate c(t) of the end of combustion position. e The sample combustion length x(t) was obtained. e )=c(t e )-c(t1). After this, if the sample does not reignite, then x(t)=x(t) e If the sample reignites, continue to detect the burning location in the video image using the aforementioned method.

[0089] Step 5: Input the sample combustion length sequence and sample combustion image sequence as input data into the constructed deep learning-based oxygen index prediction model for inference, and obtain the predicted value of oxygen index.

[0090] Specifically, the method for constructing a deep learning-based oxygen index prediction model includes the following steps:

[0091] (1) First, the true oxygen index of the sample is obtained by conducting an oxygen index test according to the standard method and using standard equipment, and the combustion length sequence of the sample X(i) is collected. i (t1), ..., x i (t n )] and the sample combustion image sequence P(i)=[p i (t1), ..., p i (t n )], construct the oxygen index sample dataset S.

[0092] (2) Construct an oxygen index prediction model, which includes an image data processing module and a time-series data processing module, such as... Figure 3 As shown, the image data processing module is a convolutional and deconvolutional neural network architecture with residual networks and feature pyramid networks, such as... Figure 4 As shown, it is used to process image data and extract the features of the data in the spatial domain;

[0093] The time-series data processing module is based on an attention mechanism (AM, such as...) Figure 5 (as shown) units and Long Short-Term Memory (LSTM, such as) Figure 6 The decoder network architecture of the unit (shown) is used to process time-series data and extract the features of the data in the time domain;

[0094] After the image data processing module processes the sample combustion image sequence, the extracted image feature sequence is input into the time series data processing module. After processing by the attention mechanism unit, it is spliced ​​with the combustion length sequence, realizing the fusion of spatial morphological information and temporal evolution information during sample combustion, and improving the prediction performance of sample oxygen index under fixed oxygen concentration.

[0095] (3) Determine the parameters of the image data processing module and the time-series data processing module, including the size of the convolutional and deconvolutional layers, the size of the convolutional kernels, the type and size of the pooling layers, padding and stride, the number of layers and the number of long short-term memory units in the long short-term memory network, fully connected layers, activation functions, etc.

[0096] Specifically, in the backbone network of the image data processing module, the present invention uses ResNeXt-50 (32x4d) as the backbone network, with an input image size of 6×128×128 and an output image size of 4×128×128. The convolutional kernel size of the convolutional and deconvolutional layers is 3×3, the activation function is ReLU, the stride is 2, the padding is 1, and the pooling window size is 1×1.

[0097] Specifically, in the time-series data processing module, the present invention sets the number of layers of the Long Short-Term Memory (LSTM) network to 4, the number of LSM units to 64, and the activation function of the LSM units to the Sigmoid function.

[0098] (4) The oxygen index prediction model is trained as follows:

[0099] First, the backbone network of the image data processing module is pre-trained by predicting future combustion images from the current combustion images of the sample. Specifically, the backbone network is input into the current m-frame combustion image sequence of the sample to predict the subsequent n-frame combustion images, the loss function is calculated, and the model weights are adjusted via backpropagation.

[0100] Then, the weights of the backbone network of the image data processing module are fixed, and the feature pyramid network in the image data processing module, as well as the attention mechanism unit and long short-term memory unit in the time-series data processing module, are trained. That is, the oxygen index sample dataset S is input into the model, the oxygen index of the sample is predicted, the loss function is calculated, and the model weights are adjusted through backpropagation.

[0101] Specifically, the training process of this invention adopts stochastic gradient descent, uses the mean squared error (MSE) loss function and AdamW optimizer for network training, and uses 10-fold cross-validation to validate the model.

[0102] The oxygen index prediction model is used to test the sample and predict its oxygen index.

[0103] like Figure 1As shown, the present invention also provides a material oxygen index prediction device based on machine vision and deep learning, including an experimental device and an image acquisition and detection device.

[0104] The test apparatus includes a base 1, a combustion cylinder 2, a sample holder 4, a diffuser ring 5, a metal mesh 6, an igniter 7, and a gas source channel 8. The bottom end of the combustion cylinder 2 is inserted into the base 1. The metal mesh 6 and the diffuser ring 5 are arranged vertically inside the base 1. The gas source channel 8 is connected to the diffuser ring 5. The sample holder 4 includes a support device and a clamp.

[0105] The image acquisition and detection device includes an optical camera 9 and a data processing and storage device 10 that communicates with the optical camera 9. The relative position and angle between the optical camera 9 and the sample 3 are fixed.

[0106] Specifically, the data processing and storage device includes any one or more of a computer, a single-board computer, and a cloud server, which can acquire video images from an optical camera and run machine vision and deep learning algorithm programs through signal transmission.

[0107] Specifically, it also includes a light shield and a monitor. The light shield can cover the entire test device and has openings at the gas input, the top of the combustion cylinder, and the optical camera 9 to ensure normal test operation. The light shield can effectively prevent external light sources from interfering with the optical camera 9 in capturing the combustion flame. The optical camera 9 is connected to the monitor (either wired or wirelessly) so that the operator can see the combustion flame inside the light shield, which is convenient for operations such as igniting the sample.

[0108] 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 predicting the oxygen index of materials based on machine vision and deep learning, characterized in that: include Sampling was performed according to the material standards, and one sample was prepared. The sample is vertically fixed inside a combustion chamber with a single upward-flowing gas source of fixed oxygen concentration, which ignites the top of the sample. Video images of the experimental process are acquired using an image acquisition and detection device; A machine vision-based image detection model is constructed. Video images are input into the trained image detection model to detect the combustion location in the video images and obtain the sample combustion length sequence and sample combustion image sequence. The sample combustion length sequence and sample combustion image sequence are used as input data to the constructed deep learning-based oxygen index prediction model for inference, and the predicted value of oxygen index is obtained. The process of detecting the combustion location in the video image and obtaining the sample combustion length sequence and sample combustion image sequence includes: set up For time numbering, the value range is from arrive , This indicates the start of combustion, marking the first sample collection. Indicates the last data collection; The instant the detection box in the image changes from one to two, the vertical pixel coordinate of the bottom of the lower flame detection box is recorded as the coordinate of the combustion start position. The burning length of the sample was obtained. ; During the experiment, the vertical coordinates of the pixels at the bottom of the flame detection frame were continuously monitored at a preset sampling frequency to determine the combustion position. The combustion length sequence of the sample was obtained. When the preset number of collections is reached, the vertical coordinate of the bottom of the detection frame is recorded as the coordinate of the end of combustion position. The burning length of the sample was obtained. Then the testing was stopped; If a detection frame disappears during the data collection process, the ordinate of the bottom of the last detection frame before it disappears will be recorded as the coordinate of the end of combustion position. The burning length of the sample was obtained. ; If the sample does not reignite thereafter, then If the sample reignites, continue to detect the burning location in the video image using the aforementioned method. In obtaining the sample combustion length sequence Simultaneously, images from the same moment in the video were captured as a sequence of combustion images of the sample. .

2. The method for predicting the oxygen index of materials based on machine vision and deep learning according to claim 1, characterized in that: The single gas source for the fixed oxygen concentration is a standard concentration gas premixed with oxygen and nitrogen in a gas cylinder.

3. The method for predicting the oxygen index of materials based on machine vision and deep learning according to claim 1, characterized in that: The image detection model includes: Input module, Backbone network, Neck module, Head module; The training process of an image detection model includes: S1: Use an image acquisition and detection device to acquire video images of the sample combustion process, label and convert the acquired video image data, and divide the processed dataset into training set and test set; S2: Input the data from the training set into the input of the image detection model for preprocessing; S3: Input the preprocessed image into the Backbone backbone network to obtain deep feature information and obtain feature maps of different depths; S5: Input feature maps of different depths into the Neck module for upsampling and feature fusion to obtain tensor data of different scales; S6: Select the correct detection box based on the predicted box and the ground box, and use the non-maximum suppression (NMS) method to remove redundant predicted boxes; S7: Calculate the model's loss function based on the detection results, adjust the model parameters, and complete the model training when the loss function value changes or no longer decreases.

4. The method for predicting the oxygen index of materials based on machine vision and deep learning according to claim 1, characterized in that: Methods for constructing oxygen index prediction models based on deep learning include: (1) Obtain the true oxygen index of the sample by performing oxygen index test according to standard method and using standard equipment, collect sample burning length sequence and sample burning image sequence, and construct oxygen index sample dataset; (2) Construct an oxygen index prediction model, which includes an image data processing module and a time series data processing module; The image data processing module is used to process spatial image data, extract the features of the data in the spatial domain, and convert the sample combustion image sequence into an image feature sequence, including a convolutional and deconvolutional neural network architecture with residual network and feature pyramid network. The time series data processing module is used to process time series data, extract the features of the data in the time domain, splice the image feature sequence with the sample combustion length sequence and predict the oxygen index, including a decoder network architecture based on attention mechanism units and long short-term memory units. (3) Determine the parameters of the image data processing module and the time-series data processing module; (4) Train the oxygen index prediction model using the constructed oxygen index sample dataset, including: A method for predicting future combustion images by inputting current combustion images of a sample involves pre-training the backbone network of the image data processing module and adjusting the model weights by calculating the loss function and backpropagating. By fixing the weights of the backbone network of the image data processing module, the method of predicting the oxygen index of the sample by inputting the oxygen index sample dataset into the model is used to train the feature pyramid network in the image data processing module and the attention mechanism unit and long short-term memory unit in the time series data processing module. The model weights are adjusted by calculating the loss function and backpropagation. The final model parameters are determined to obtain the final oxygen index prediction model.

5. A material oxygen index prediction device based on machine vision and deep learning, characterized in that: Includes experimental equipment and image acquisition and detection equipment; The test apparatus includes a base, a combustion cylinder, a sample holder, a diffuser ring, a metal mesh, an igniter, and a gas source channel. The bottom end of the combustion cylinder is inserted into the base. The base is equipped with a metal mesh and a diffuser ring arranged vertically. The gas source channel is connected to the diffuser ring. The sample holder includes a support device and a clamp. The image acquisition and detection device includes an optical camera and a data processing and storage device that communicates with the optical camera. The relative position and angle between the optical camera and the sample are fixed. During the experiment, samples were taken according to the material standards, and one specimen was prepared. The sample is vertically fixed inside a combustion chamber with a single upward-flowing gas source of fixed oxygen concentration, which ignites the top of the sample. Video images of the experimental process are acquired using an image acquisition and detection device; The data processing and storage device is used for: A machine vision-based image detection model is constructed. Video images are input into the trained image detection model to detect the combustion location in the video images and obtain the sample combustion length sequence and sample combustion image sequence. The sample combustion length sequence and sample combustion image sequence are used as input data to the constructed deep learning-based oxygen index prediction model for inference, and the predicted value of oxygen index is obtained. The process of detecting the combustion location in the video image and obtaining the sample combustion length sequence and sample combustion image sequence includes: set up For time numbering, the value range is from arrive , This indicates the start of combustion, marking the first sample collection. Indicates the last data collection; The instant the detection box in the image changes from one to two, the vertical pixel coordinate of the bottom of the lower flame detection box is recorded as the coordinate of the combustion start position. The burning length of the sample was obtained. ; During the experiment, the vertical coordinates of the pixels at the bottom of the flame detection frame were continuously monitored at a preset sampling frequency to determine the combustion position. The combustion length sequence of the sample was obtained. When the preset number of collections is reached, the vertical coordinate of the bottom of the detection frame is recorded as the coordinate of the end of combustion position. The burning length of the sample was obtained. Then the testing was stopped; If a detection frame disappears during the data collection process, the ordinate of the bottom of the last detection frame before it disappears will be recorded as the coordinate of the end of combustion position. The burning length of the sample was obtained. ; If the sample does not reignite thereafter, then If the sample reignites, continue to detect the burning location in the video image using the aforementioned method. In obtaining the sample combustion length sequence Simultaneously, images from the same moment in the video were captured as a sequence of combustion images of the sample. .

6. The material oxygen index prediction device based on machine vision and deep learning according to claim 5, characterized in that: The data processing and storage device includes any one of a computer, a single-board computer, and a cloud server.

7. The material oxygen index prediction device based on machine vision and deep learning according to claim 5, characterized in that: It also includes a light shield and a monitor. The light shield is used to cover the area around the test device to prevent external light sources from interfering with the optical camera's capture of the combustion flame. The monitor is connected to the optical camera and is used to observe the combustion scene.

Images