Explosive impact sensitivity test identification method based on multi-parameter data driving

Abstract

The invention discloses an explosive impact sensitivity test identification method based on multi-parameter data driving. The explosive impact sensitivity test identification method comprises the following steps: acquiring multi-parameter coupling data in an explosive impact sensitivity test process; performing feature extraction on audio data in the multi-parameter coupling data, and converting the audio data into a spectrogram; amplifying the generated spectrogram based on a DCGAN model; performing explosive ignition intelligent identification model training based on the amplified spectrogram and gas data in the multi-parameter coupling data; and based on the trained explosive ignition intelligent identification model, inputting multi-parameter coupling data of the to-be-tested explosive to obtain a prediction result.

Description

Technical Field

[0001] The invention relates to the technical field of explosive safety and intelligent identification, and in particular to an explosive impact sensitivity test identification method driven by multi-parameter data. Background Art

[0002] In recent years, explosives have been increasingly used in military and civilian applications due to their high energy density, but safety accidents have also become more frequent, and casualties have become more severe. The occurrence of safety accidents with explosives is often related to inaccurate sensitivity evaluation and unclear understanding of the ignition mechanism. In particular, many explosion accidents are caused by impact of explosives. Therefore, impact sensitivity has become an important indicator for evaluating the safety of explosives, which is directly related to the safety performance of weapon systems and the personal safety of explosive operators.

[0003] The existing technology mainly has the following problems:

[0004] (1) Dependence on subjective methods: Impact sensitivity testing mainly uses subjective methods such as "listening to sounds, watching for flames, and observing traces". It lacks objective quantitative indicators, and the judgment results are heavily dependent on the subjective experience of the experimenter.

[0005] (2) Lack of data: Due to the relatively simple impact sensitivity test process, small site, high gas collection cost, and difficulty in taking photos during the test, there is a lack of systematic and comprehensive multi-parameter data on explosive ignition. The parameters of the explosive ignition process are numerous and mutually coupled, requiring systematic testing to obtain comprehensive data, but there is currently no relevant report.

[0006] (3) Lack of data fusion analysis methods: The ignition criteria of explosives involve multiple parameters, and it is necessary to extract key information from massive heterogeneous data and establish criteria, but there is currently no relevant research content.

[0007] (4) Insufficient research on intelligent identification: At present, the intelligent research on impact sensitivity mainly focuses on predicting the characteristic drop height, while there is a lack of intelligent identification methods for explosive ignition and results. Real-time intelligent identification of explosive ignition based on the established multi-parameter criteria is crucial to ensure the safety of explosives, but no relevant research has been seen.

[0008] Chinese patent 202110387136.2 invented a method for rapid quantitative prediction of impact sensitivity and friction sensitivity of energetic materials based on laser induced breakdown spectroscopy, which belongs to the technical field of prediction of mechanical sensitivity of energetic materials. The laser plasma emission spectrum of energetic materials with known impact sensitivity and friction sensitivity is obtained, and effective spectral preprocessing means such as effective data screening, continuous radiation background removal, and channel normalization are performed. Based on all spectral information such as characteristic atoms, characteristic ions, and characteristic molecular radiation, training, verification, and testing are performed to establish linear regression models of impact sensitivity and friction sensitivity respectively; the laser plasma emission spectrum of energetic materials with unknown impact sensitivity and friction sensitivity is obtained, and effective spectral preprocessing means such as effective data screening, continuous radiation background removal, and channel normalization are performed. Based on all spectral information such as characteristic atoms, characteristic ions, and characteristic molecular radiation, combined with the linear regression model of impact sensitivity or friction sensitivity, the predicted impact sensitivity value or friction sensitivity value of energetic materials with unknown sensitivity value is obtained. The method described in the present invention has very little sample consumption, fast analysis speed, accurate prediction results, and good application prospects. However, the cost is relatively high.

[0009] Chinese patent 202310141419.8 discloses an automated device for impact sensitivity testing of fireworks and firecrackers, which can reverse the reeling motor after detection to reel in the reeling rope, so that the impact hammer can move up and reset after the impact, thereby realizing the automatic recovery of the equipment, and can be operated cyclically, and can be tested repeatedly without human intervention, with a higher overall degree of automation and reduced human operation intensity. However, the judgment of the reaction results still requires manual judgment.

[0010] Chinese patent 202220451762.3 discloses an energetic material impact sensitivity test device, which combines an impact sensitivity meter with a temperature detector, a gas detector and an explosion probability calculation device, and can meet the precise test of the impact sensitivity of energetic materials. The test results are digitized and have high practical value. However, the device process is complicated, the cost is high, and the use value is low.

[0011] In summary, existing impact sensitivity test technologies and methods are relatively dependent on manual operation and personal experience, with high uncertainty and poor repeatability. Summary of the invention

[0012] The purpose of the present invention is to provide an explosive impact sensitivity test identification method based on multi-parameter data drive, obtain multi-parameter coupling data of the explosive impact sensitivity test process, and use a machine learning method to realize intelligent identification of the explosive impact sensitivity test.

[0013] In order to achieve the above object, the present invention adopts the following technical solutions:

[0014] A method for explosive impact sensitivity test identification based on multi-parameter data drive includes the following steps:

[0015] Obtain multi-parameter coupling data during explosive impact test;

[0016] Extract features from the audio data in the multi-parameter coupling data and convert it into a spectrum diagram;

[0017] Amplify the generated spectrogram based on the DCGAN model;

[0018] Based on the amplified spectrum and the gas data in the multi-parameter coupling data, the intelligent recognition model of explosive ignition is trained;

[0019] Based on the trained intelligent recognition model of explosive ignition, the multi-parameter coupling data of the explosive to be tested is input to obtain the prediction result.

[0020] In some embodiments, the multi-parameter coupling data includes: gas data and audio data.

[0021] In some embodiments, feature extraction is performed on audio data in multi-parameter coupling data and converted into a spectrogram, including: using the librosa library to extract the digitized time series and sampling rate in the audio data, and based on the librosa.stft() function, performing a short-time Fourier transform on the extracted time series to obtain the intensity of the frequency components of the audio signal at different times; and then representing it with an image, wherein the horizontal axis represents time, the vertical axis represents frequency, and the color represents intensity, i.e. loudness.

[0022] In some embodiments, the DCGAN model-based amplification of the generated spectrum graph includes:

[0023] Input the real data spectrum and train the discriminator through convolution to enable it to distinguish true from false;

[0024] After the discriminator has trained for a preset number of rounds, the generator starts working, generating images by generating random noise, and then giving them to the discriminator for training;

[0025] Multiple loop iterations make the images generated by the generator close to the real data.

[0026] In some embodiments, the DCGAN model includes: a discriminator and a generator; the discriminator part includes 1 input layer, 5 convolutional layers and 1 output layer, wherein the convolution kernel size of the convolutional layer is 4*4, the stride is 2, and each convolutional layer is connected to a Relu layer;

[0027] The generator consists of 1 input layer, 5 deconvolution layers and 1 output layer. Each deconvolution layer is activated by a Relu layer, and the output layer is activated by a tanh function. The real data spectrum is used as input to train the model to achieve the goal of generating data.

[0028] In some embodiments, the input real data spectrum diagram is used to train the discriminator through convolution so that it has the ability to distinguish between true and false, including: taking the picture as input, converting the picture into a matrix according to pixels, corresponding to the pixel values ​​of the three channels, and obtaining a feature map after convolution of each layer of the matrix through a convolution kernel. After multiple convolutions, a 512-dimensional feature map is obtained, and a number is obtained after processing through a fully connected layer. The number is processed by a sigmoid function and then output, and the output is close to 1 for true and close to 0 for false.

[0029] In some embodiments, after the discriminator has been trained for a preset number of rounds, the generator starts working, generates images by generating random noise, and then hands them over to the discriminator for training; including: the generator first randomly generates a noise vector with a dimension of 100*1*1, then changes the dimension and size of the feature map through deconvolution, and finally outputs a 3*128*128 feature map; then the generated image is put into the discriminator for training.

[0030] In some embodiments, the explosive ignition intelligent recognition model training based on the amplified spectrum diagram and the gas data in the multi-parameter coupling data includes:

[0031] The model is trained using the multi-parameter coupling data set after feature extraction. The model input is gas data and audio data, and the output is the confidence probability of the ignition state of the explosive.

[0032] The model adopts the CNN model, which includes 4 convolutional layers, 4 pooling layers, 2 fully connected layers and 1 output layer. The convolution kernel size of the first two layers is 5*5, the convolution kernel size of the last two layers is 3*3, the stride is 1, and the pooling layers are all maximum pooling, the pooling kernel size is 2*2, the stride is 2, and the size of the feature map is halved; finally, the output layer outputs the discrimination result.

[0033] In some embodiments, the CNN model uses an image as input, and the CNN model compresses the image into a size of 224*224 before training;

[0034] The input feature map is transformed into a feature map with increased dimensions after multiple layers of convolution. Finally, the multi-dimensional feature map is transformed into a 1024*1 form through a fully connected layer, and then through another fully connected layer to output two numbers (2*1).

[0035] Input gas data, input 11 values ​​(11*1), after passing through two layers of neural network, it becomes 1024*1 format; add it to the vector output to 1024*1 in CNN, it becomes 1*2048, and then connect it to a fully connected layer, it becomes two values ​​(2*1);

[0036] The two values ​​are then processed by the softmax function to become values ​​between 0 and 1, and then the classification with a large probability is selected as the judgment output.

[0037] The beneficial effects that may be brought about by the explosive impact sensitivity test identification method based on multi-parameter data drive disclosed in this application include but are not limited to:

[0038] 1. Systematic and comprehensive multi-parameter coupling data of the explosive ignition process were obtained, laying a data foundation for in-depth research on the ignition mechanism of explosives.

[0039] 2. A fusion analysis method for multi-source heterogeneous data was proposed, and the key multi-parameter information of explosive ignition was extracted from massive data.

[0040] 3. An intelligent recognition model for explosive ignition based on machine learning was established, which can realize real-time automatic judgment of the ignition status of explosives and safety warning.

[0041] 4. The present invention can be widely used in the production, storage and transportation of explosives, and is of great significance for improving the safety of explosives and reducing accident risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the DCGAN model flow chart;

[0043] Figure 2 A schematic diagram of amplifying the generated spectrum graph based on the DCGAN model;

[0044] Figure 3 A schematic diagram of training an intelligent identification model for explosive ignition based on the gas data in the amplified spectrum diagram and multi-parameter coupling data;

[0045] Figure 4 is a schematic diagram of the spectrum; DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present application more clear, the present application is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0047] On the contrary, the present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application as defined by the claims. Further, in order to make the public have a better understanding of the present application, some specific details are described in detail in the detailed description of the present application below. Those skilled in the art can fully understand the present application without the description of these details.

[0048] The following is a detailed description of an explosive impact sensitivity test identification method based on multi-parameter data driven according to an embodiment of the present application. It should be noted that the following embodiments are only used to explain the present application and do not constitute a limitation on the present application.

[0049] like Figure 1 As shown, a method for explosive impact sensitivity test identification based on multi-parameter data drive includes the following steps:

[0050] Step 1: Acquire multi-parameter coupling data during the explosive impact test; the multi-parameter coupling data includes: gas data and audio data. Use audio acquisition equipment and gas acquisition impact equipment to conduct tests on explosives. Through the impact sensitivity test, record the audio data and gas data during the reaction process.

[0051] When conducting the explosive impact sensitivity test, the gas collection equipment is used to collect the gas generated during the reaction and record the gas composition and its changes. The audio acquisition equipment includes an acoustic sensor, which is placed 2 meters away from the center of the impact sensitivity tester and 1.2 meters high. It is used to collect audio data during the test; and the data is transmitted to the computer through a connecting line for storage.

[0052] Step 2: Extract features from the audio data in the multi-parameter coupling data and convert it into a spectrogram; use the Librosa library to extract features from the audio and convert it into a spectrogram. The gas data itself is in digital form and does not require special processing, just organize it. The gas data is tabulated as follows. The numbers represent the percentage content of each component in the gas collected after the impact, which can be directly used as input data for training.

[0053]

[0054]

[0055] The audio conversion process begins with sampling. The collection device is a small metal microphone, which is connected to the collection device through a channel line and then to the computer through a data line. The signal collected by the collection device is in wav format, which is a digital audio format for storing sound waveforms, including sampling rate, bit depth and other information. This study first uses the librosa library to read the digitized time series (including time and corresponding amplitude) and sampling rate (how many times per second, expressed as frequency) in the wav format. Then use the librosa.stft() function to perform a short-time Fourier transform on the extracted time series to obtain the intensity of the frequency components of the audio signal at different times. Then use an image to represent it, with the horizontal axis representing time, the vertical axis representing frequency, and the color representing intensity, or loudness.

[0056] Convert audio to a spectrogram:

[0057] Spectrum diagram Figure 4 , the horizontal axis is time, the vertical axis is frequency, and the color represents the loudness of the sound. The left side of the figure is the waveform diagram of the audio, and the right side is the spectrum diagram.

[0058] Step 3: Amplify the generated spectrogram based on the DCGAN model;

[0059] In order to increase the amount of data, the DCGAN (Deep Convolutional Generative Adversarial Network) model is used to generate more audio data. The specific process includes: a. Input the real data spectrum and train the discriminator through convolution to enable it to distinguish true from false. b. After the discriminator is trained for 5 rounds, the generator starts to work and generates pictures by generating random noise, which are then handed over to the discriminator for training. c. Multiple cycles are repeated to make the pictures generated by the generator close to the real data, thereby increasing the amount of data and improving the accuracy of the model.

[0060] DCGAN is mainly divided into a discriminator and a generator. The discriminator includes 1 input layer, 5 convolutional layers, and 1 output layer. The convolution kernel size of the convolutional layer is 4*4, the stride is 2, and each convolutional layer is connected to a Relu layer. The generator consists of 1 input layer, 5 deconvolutional layers, and 1 output layer. Each deconvolutional layer is activated by a Relu layer, and the output layer is activated by a tanh function. Use real data as input to train the model to achieve the goal of generating data.

[0061] The DCGAN used defines a convolutional neural network as the discriminator. The specific process is as follows Figure 2, take the image as input, that is, a tensor in digital form (corresponding to the pixel values ​​of the three channels). The tensor means converting the image into a matrix based on pixels, but the color image has three channels, that is, one image becomes three matrices, which is called a tensor. Then convolve each layer of the matrix with a convolution kernel to get a feature map. After multiple convolutions, a 512-dimensional feature map is finally obtained. After processing through the fully connected layer, a number is obtained. After the number is processed by the sigmoid function, it is output. This output is regarded as a probability, close to 1 is true, and close to 0 is false.

[0062] Then a generator is defined, which first randomly generates a noise vector of dimension 100*1*1, then changes the dimension and size of the feature map through deconvolution, and finally outputs a feature map of 3*128*128. The generated map is then put into the discriminator for training.

[0063] Step 4: Train the intelligent recognition model for explosive ignition based on the amplified spectrum and the gas data in the multi-parameter coupling data; select a suitable machine learning algorithm, such as a convolutional neural network (CNN) or a VIT model, and use the multi-parameter data set after feature extraction for model training. The model input is multi-parameter dynamic data, and the output is the confidence probability of the explosive ignition state. Determine the confidence probability threshold based on expert experience, and establish the "whether to react-reaction probability-reaction magnitude" multi-parameter criterion for explosive ignition.

[0064] Both CNN and VIT models use images as input. Before training, the CNN and VIT models compress the images into a size of 224*224.

[0065] The CNN model includes 4 convolutional layers, 4 pooling layers, 2 fully connected layers and 1 output layer. The convolution kernel size of the first 2 layers is 5*5, the convolution kernel size of the last 2 layers is 3*3, the stride is 1, and the pooling layers are all maximum pooling, the pooling kernel size is 2*2, the stride is 2, and the feature map size is halved. Finally, the output layer outputs the discrimination result.

[0066] The overall CNN process is similar to the convolution process in DCGAN. The input image is converted into a feature map with increased dimensions after multiple layers of convolution. Finally, the multi-dimensional feature map is converted into a 1024*1 format through a fully connected layer. Then, it passes through another fully connected layer and outputs two numbers (2*1). The two numbers are then processed by the softmax function to become values ​​between 0 and 1, and then the classification with a large probability is selected as the judgment output.

[0067] The VIT model first has a PatchEmbedding layer, which divides the image into 16 blocks and converts each block into a one-dimensional vector. Then there is an Encoding layer, which labels the positions of the 16 vectors, followed by a dropout layer to prevent the model from overfitting, which randomly selects some parameters each time and does not participate in the training. Then, four Encoder Block layers are stacked, and the output layer is output after connecting a standardization layer. The Encoder Block layer is relatively complex, including a normalization layer, a multi-head attention module (used to calculate attention weights and process input tensors), and a dropout layer that discards some neurons to prevent overfitting. The result after the dropout layer is added to the original input to achieve residual connection, normalized, passed through the MLP (2 fully connected) layer, and then the dropout layer, and finally the residual connection.

[0068] The main steps of the VIT model are to take the image (3*224*224) as input, and first divide the image into 196 blocks of size 16*16 and 3 channels through convolution (convolution kernel size is 16*16). Then expand the 196 blocks of tensors and output them as a two-dimensional matrix, with each row corresponding to a patch. Then transpose the expanded two-dimensional matrix and use it as a digital input to the next layer. After encoding the expanded values, use the multi-head attention layer to calculate the correlation between each patch, and finally output the values ​​through the fully connected layer, and use softmax as the activation function to select the category with a large probability value for output.

[0069] Input an audio, and then the model will give its spectrum graph after calculation, and predict whether it explodes or not, and then it is considered that the probability of it exploding is 100%.

[0070] There is also a characteristic fusion model that adds gas, such as Figure 3 , based on the CNN model training spectrum, gas data is added, 11 values ​​(1*11) are input, and after two layers of neural networks, they become 1*1024, and merged with the 1*1024 vector output from CNN to become 1*2048, and then connected to a fully connected layer, becoming two values ​​(2*1), corresponding to two categories, explosion or no explosion, whichever number is larger, which category is output. Then use the softmax function to process it into a probability form, and select the category with a larger probability for output.

[0071] Step 5: Based on the trained intelligent identification model for explosive ignition, input the multi-parameter coupling data of the explosive to be tested to obtain the prediction results. Place different batches of explosive samples in the impact sensitivity device, apply the same impact conditions as in the data collection stage, collect multi-parameter data such as gas, audio, temperature, etc. in real time, and input them into the trained intelligent model. The model automatically calculates the confidence probability of the ignition state, identifies the state of sample ignition based on the established model criteria, and issues a safety warning signal of the corresponding level.

[0072] 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 and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying explosive impact sensitivity test based on multi-parameter data drive, characterized in that: The following steps are involved: Obtain multi-parameter coupling data during explosive impact test; Extract features from the audio data in the multi-parameter coupling data and convert it into a spectrum diagram; Amplify the generated spectrogram based on the DCGAN model; Based on the amplified spectrum and the gas data in the multi-parameter coupling data, the intelligent recognition model of explosive ignition is trained; Based on the trained intelligent recognition model of explosive ignition, the multi-parameter coupling data of the explosive to be tested is input to obtain the prediction result.

2. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 1 is characterized in that: The multi-parameter coupling data includes: gas data and audio data.

3. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 2 is characterized in that: Feature extraction is performed on the audio data in the multi-parameter coupling data and converted into a spectrum diagram, including: using the librosa library to extract the digitized time series and sampling rate in the audio data, and based on the librosa.stft() function, performing short-time Fourier transform on the extracted time series to obtain the intensity of the frequency components of the audio signal at different times; and then expressing it with an image, where the horizontal axis represents time, the vertical axis represents frequency, and the color represents intensity, i.e. loudness.

4. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 2 is characterized in that: The amplification of the generated spectrum graph based on the DCGAN model includes: Input the real data spectrum and train the discriminator through convolution to enable it to distinguish true from false; After the discriminator has trained for a preset number of rounds, the generator starts working, generating images by generating random noise, and then giving them to the discriminator for training; Multiple loop iterations make the images generated by the generator close to the real data.

5. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 4 is characterized in that: The DCGAN model includes: a discriminator and a generator; the discriminator part includes 1 input layer, 5 convolutional layers and 1 output layer, wherein the convolution kernel size of the convolutional layer is 4*4, the stride is 2, and each convolutional layer is connected to a Relu layer; The generator consists of 1 input layer, 5 deconvolution layers and 1 output layer. Each deconvolution layer is activated by a Relu layer, and the output layer is activated by a tanh function. The real data spectrum is used as input to train the model to achieve the goal of generating data.

6. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 5 is characterized in that: The real data spectrum diagram is input and the discriminator is trained by convolution to enable it to have the ability to distinguish true from false, including: taking the picture as input, converting the picture into a matrix according to pixels, corresponding to the pixel values ​​of the three channels, obtaining a feature map after convolution of each layer of the matrix through a convolution kernel, and obtaining a 512-dimensional feature map after multiple convolutions, obtaining a number after processing through a fully connected layer, and outputting the number after processing with a sigmoid function. The output is close to 1 for true and close to 0 for false.

7. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 5 is characterized in that: After the discriminator has trained for a preset number of rounds, the generator starts working, generating images by generating random noise, and then giving them to the discriminator for training; including: the generator first randomly generates a noise vector with a dimension of 100*1*1, then changes the dimension and size of the feature map through deconvolution, and finally outputs a 3*128*128 feature map; then the generated image is put into the discriminator for training.

8. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 4 is characterized in that: The method of training the explosive ignition intelligent recognition model based on the amplified spectrum diagram and the gas data in the multi-parameter coupling data comprises: The model is trained using the multi-parameter coupling data set after feature extraction. The model input is gas data and audio data, and the output is the confidence probability of the ignition state of the explosive. The model adopts the CNN model, which includes 4 convolutional layers, 4 pooling layers, 2 fully connected layers and 1 output layer. The convolution kernel size of the first two layers is 5*5, the convolution kernel size of the last two layers is 3*3, the stride is 1, and the pooling layers are all maximum pooling, the pooling kernel size is 2*2, the stride is 2, and the size of the feature map is halved; finally, the output layer outputs the discrimination result.

9. The explosive impact sensitivity test identification method based on multi-parameter data drive according to claim 8, characterized in that: The CNN model uses images as input. Before training, the CNN model compresses the images into a size of 224*224. The input feature map is transformed into a feature map with increased dimensions after multiple layers of convolution. Finally, the multi-dimensional feature map is transformed into a 1024*1 form through a fully connected layer, and then through another fully connected layer to output two numbers (2*1). Input gas data, input 11 values ​​(11*1), after passing through two layers of neural network, it becomes 1024*1 format; add it to the vector output to 1024*1 in CNN, it becomes 2048*1, and then connect it to a fully connected layer, it becomes two values ​​(2*1); The two values ​​are then processed by the softmax function to become values ​​between 0 and 1, and then the classification with a large probability is selected as the judgment output.

Images