Combustible metal dust spontaneous combustion early warning system used in industrial dust removal facility

Abstract

The invention provides a combustible metal dust spontaneous combustion early warning system used in an industrial dust removal facility, relates to the technical field of industrial safety monitoring, and aims to comprehensively monitor multiple parameters such as temperature, carbon monoxide concentration, hydrogen concentration, smoke concentration and the like and perform intelligent analysis by using a deep learning technology. A deep neural network model based on spatio-temporal feature fusion is adopted in the deep learning module, dynamic motion information is extracted by introducing an optical flow method, intelligent analysis and risk prediction of field images are achieved, the system can capture tiny changes in the initial stage of spontaneous combustion in time, the spontaneous combustion risk is accurately pre-judged through multi-parameter correlation analysis, and the risk prediction efficiency is improved. And alarm signals of different levels are automatically triggered according to the risk levels, so that the accuracy and timeliness of early warning are remarkably improved, and the occurrence rate of spontaneous combustion accidents of combustible metal dust is effectively reduced.

Description

Technical Field

[0001] The invention relates to the technical field of industrial safety monitoring, in particular to a combustible metal dust spontaneous combustion early warning system used in industrial dust removal facilities. Background Art

[0002] Combustible metal dusts such as aluminum powder and magnesium powder play a key role in aerospace, military industry, automobile manufacturing, paint and coatings due to their reactivity. However, these metal dusts have significant safety hazards during production, storage, transportation and use due to their flammable and explosive properties. In industrial environments, industrial dust removal facilities are an important part of ensuring production safety. They usually include dust collector bodies, pipes and fans. These systems often have complex internal environments, and metal dust is easy to accumulate. Especially under humid conditions, the risk of spontaneous combustion increases significantly, which may lead to fires or explosions, causing casualties and property losses.

[0003] Traditional combustible metal dust spontaneous combustion warning methods rely on manual inspections and sensor threshold monitoring, but this method has limitations. For example, the monitoring parameters are single and cannot fully capture the precursors of spontaneous combustion; sensor threshold monitoring is difficult to capture small changes in the early stages of spontaneous combustion; there is a lack of intelligent analysis capabilities, making it difficult to accurately predict the risk of spontaneous combustion; manual judgment is easily affected by subjective factors and is difficult to handle complex multi-parameter correlation analysis; manual monitoring is costly and inefficient, making it difficult to achieve 24-hour continuous monitoring; in large-scale production or storage sites, the labor cost is huge; warning information may not be transmitted in a timely manner, missing the best time to handle it; there is a lack of historical data analysis and trend prediction capabilities, making it impossible to optimize the warning strategy.

[0004] Therefore, there is an urgent need to develop a combustible metal dust spontaneous combustion early warning system that can be used in industrial dust removal facilities to comprehensively monitor, intelligently analyze and promptly warn of flammable metal dust, so as to improve the safety of industrial production. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to propose a combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities, including a sensor module, an industrial control unit, a host computer module, an image acquisition module, a deep learning module, a data storage module and an audio-visual alarm module;

[0006] The sensor module is connected to the industrial control unit, and the sensor module includes a temperature sensor, a carbon monoxide concentration sensor, a hydrogen concentration sensor and a smoke concentration sensor;

[0007] The industrial control unit includes a signal processing unit and a data transmission unit, the data transmission unit is used to receive the monitoring parameter time series transmitted by the sensor module, the monitoring parameter time series includes a temperature time series, a carbon monoxide concentration time series, a smoke concentration time series and a hydrogen concentration time series, the monitoring parameter time series includes monitoring parameters, and the monitoring parameters include temperature, carbon monoxide concentration, smoke concentration and hydrogen concentration; the signal processing unit is used to perform analog-to-digital conversion and filtering processing on the monitoring parameter time series to obtain a processed monitoring parameter time series, and the data transmission unit is also used to transmit the processed monitoring parameter time series to the host computer module;

[0008] The host computer module receives the processed monitoring parameter time series, and performs data standardization on the monitoring parameters in the processed monitoring parameter time series, and then performs time series analysis on the standardized monitoring parameter time series to obtain the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, and at the same time performs feature extraction on the standardized monitoring parameter time series to obtain the moving standard deviation, the moving average difference and the exponentially weighted moving average of the standardized monitoring parameter time series, and calculates the total risk score of the industrial dust removal facility according to the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, and the moving standard deviation, the moving average difference and the exponentially weighted moving average of the standardized monitoring parameter time series, and then determines the risk level according to the total risk score, and the risk level is low risk, medium risk or high risk;

[0009] The deep learning module is used to receive the on-site image sequence sent by the image acquisition module, and perform image processing on the on-site image sequence, perform optical flow calculation and motion feature extraction on the image sequence after image processing, and obtain multiple target images. For each target image, the target image is used as an input sample of the deep neural network model based on spatiotemporal feature fusion in the deep learning module, and the risk level corresponding to the target image is determined according to the upper computer module, and the risk level corresponding to the target image is used as the output sample of the deep neural network model based on spatiotemporal feature fusion in the deep learning module. The input sample and the output sample constitute a training sample, and multiple training samples constitute a training sample set. The deep neural network model based on spatiotemporal feature fusion is trained according to the training sample set to obtain a trained deep neural network model based on spatiotemporal feature fusion. The image to be detected is identified by the trained deep neural network model based on spatiotemporal feature fusion to obtain the risk level of the image to be detected, and the risk level of the image to be detected is transmitted to the sound and light alarm module;

[0010] The data storage module is used to store historical data of monitoring parameters, which include at least temperature, carbon monoxide concentration, hydrogen concentration, smoke concentration, total risk score, records of alarm signals triggered by the sound and light alarm module, and on-site image sequences;

[0011] The sound and light alarm module includes an LED indicator light and a high-decibel alarm.

[0012] Optionally, the host computer module performs data standardization on the monitoring parameters in the processed monitoring parameter time series, which is specifically implemented by formula (1), which is:

[0013]

[0014] Where X' is the monitoring parameter in the standardized monitoring parameter time series, X is the monitoring parameter in the processed monitoring parameter time series, μ is the average value of the monitoring parameter in the processed monitoring parameter time series, and σ is the standard deviation of the monitoring parameter in the processed monitoring parameter time series;

[0015] Based on formula (1), the temperature time series, carbon monoxide concentration time series, smoke concentration time series and hydrogen concentration time series in the monitoring parameter time series are standardized to obtain a standardized temperature time series, a standardized carbon monoxide concentration time series, a standardized smoke concentration time series and a standardized hydrogen concentration time series.

[0016] Optionally, the host computer module performs time series analysis on the standardized monitoring parameter time series to obtain the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, specifically including:

[0017] Calculate the first-order derivative of the standardized monitoring parameter time series, specifically by formula (2), which is:

[0018]

[0019] Among them, X′(t) represents the standardized monitoring parameter time series with the independent variable at time t, X′(t-Δt) represents the standardized monitoring parameter time series with the independent variable at time t-Δt, and Δt is the sampling time interval;

[0020] The second-order derivative of the standardized monitoring parameter time series is calculated by formula (3), which is:

[0021]

[0022] Based on formula (2) and formula (3), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the first-order derivative of the standardized temperature time series: And the second derivative of the standardized temperature time series The first derivative of the normalized carbon monoxide concentration time series And the second derivative of the normalized carbon monoxide concentration time series The first derivative of the normalized smoke concentration time series And the second derivative of the normalized smoke concentration time series The first derivative of the normalized hydrogen concentration time series And the second derivative of the normalized hydrogen concentration time series

[0023] Optionally, the host computer module performs feature extraction on the standardized monitoring parameter time series to obtain a moving standard deviation, a moving average deviation and an exponentially weighted moving average of the standardized monitoring parameter time series, specifically including:

[0024] The moving standard deviation of the standardized monitoring parameter time series is calculated by formula (4), which is:

[0025]

[0026] Among them, σ x′ (t) is the moving standard deviation of the standardized monitoring parameter time series at time t, Q is the preset time interval, t′ is the time Q seconds before time t, X′(t′) is the monitoring parameter of the standardized monitoring parameter time series at time t′, It represents the average value of the monitoring parameter in the standardized monitoring parameter time series within Q seconds corresponding to time t;

[0027] Based on formula (4), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the moving standard deviation σ of the standardized temperature time series T′ (t), the moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving standard deviation σ of the standardized smoke concentration time series S′ (t) and the moving standard deviation σ of the standardized hydrogen concentration time series H′ (t);

[0028] The moving average difference of the standardized monitoring parameter time series is calculated by formula (5), which is:

[0029]

[0030] Among them, Δ x′ (t) is the moving average difference of the normalized monitoring parameter time series at time t;

[0031] Based on formula (5), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the moving average difference Δ T′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the moving average difference Δ of the standardized hydrogen concentration time series H′ (t);

[0032] The exponentially weighted moving average of the standardized monitoring parameter time series is calculated by formula (6), which is:

[0033] μ x′ (t) = βX′(t) + (1-β)μ x′ (t-1) (6)

[0034] Among them, μ X′ (t) is the exponentially weighted moving average of the standardized monitoring parameter time series at time t, β is the smoothing factor, X'(t) is the monitoring parameter of the standardized monitoring parameter time series at time t, μ X′ (t-1) is the exponentially weighted moving average of the normalized monitoring parameter time series at time t-1;

[0035] Based on formula (6), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the exponentially weighted moving average μ of the standardized temperature time series: T′ (t), exponentially weighted moving average μ of the standardized carbon monoxide concentration time series C′ (t), exponentially weighted moving average μ of the standardized smoke concentration time series S′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time seriesH′ (t).

[0036] Optionally, the standardized monitoring parameter time series includes a standardized temperature time series, a standardized carbon monoxide concentration time series, a standardized smoke concentration time series, and a standardized hydrogen concentration time series; the first-order derivative and the second-order derivative of the standardized monitoring parameter time series include the first-order derivative of the temperature time series. And the second derivative of the standardized temperature time series The first derivative of the normalized carbon monoxide concentration time series And the second derivative of the normalized carbon monoxide concentration time series The first derivative of the normalized smoke concentration time series And the second derivative of the normalized smoke concentration time series The first derivative of the normalized hydrogen concentration time series And the second derivative of the normalized hydrogen concentration time series

[0037] The moving standard deviation of the standardized monitoring parameter time series includes the moving standard deviation of the standardized temperature time series σ T′ (t), the moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving standard deviation σ of the standardized smoke concentration time series S′ (t) and the moving standard deviation σ of the standardized hydrogen concentration time series H′ (t);

[0038] The moving average difference of the standardized monitoring parameter time series includes the moving average difference of the standardized temperature time series Δ T′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the moving average difference Δ of the standardized hydrogen concentration time series H′ (t); The exponentially weighted moving average of the standardized monitoring parameter time series includes the exponentially weighted moving average of the standardized temperature time series μ T′ (t), exponentially weighted moving average μ of the standardized carbon monoxide concentration time series C′ (t), exponentially weighted moving average μ of the standardized smoke concentration time series S′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t);

[0039] Based on this, the host computer module calculates the total risk score of the industrial dust removal facility according to the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, as well as the moving standard deviation, moving average deviation and exponentially weighted moving average of the standardized monitoring parameter time series, including:

[0040] According to the first derivative of the standardized temperature time series Normalized second derivative of temperature time series The moving standard deviation σ of the standardized temperature time series T′ (t), the moving average difference of the standardized temperature time series Δ T′ (t) and the exponentially weighted moving average μ of the standardized temperature time series T′ (t), calculate the temperature risk score R T (t) is calculated by formula (7), which is:

[0041]

[0042] Among them, a T 、b T 、c T ,d T 、e T All are weight coefficients;

[0043] According to the first derivative of the normalized carbon monoxide concentration time series The second derivative of the normalized carbon monoxide concentration time series The moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t) and the exponentially weighted moving average μ of the normalized carbon monoxide concentration time series C′ (t), calculate the carbon monoxide concentration risk score R C (t), is calculated by formula (8), which is:

[0044]

[0045] Among them, a C 、b C 、c C ,d C 、e C All are weight coefficients;

[0046] According to the first derivative of the standardized smoke concentration time series Normalized second-order derivative of smoke concentration time series The moving standard deviation σ of the standardized smoke concentration time series S′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the exponentially weighted moving average μ of the normalized smoke concentration time series S′ (t), calculate the smoke concentration risk score R S (t), is calculated by formula (9), which is:

[0047]

[0048] Among them, a S 、b S 、c S ,d S 、e S All are weight coefficients;

[0049] According to the first derivative of the normalized hydrogen concentration time series Normalized second derivative of hydrogen concentration time series The moving standard deviation σ of the standardized hydrogen concentration time series H′ (t), moving average difference of the normalized hydrogen concentration time series Δ H′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t), calculate the hydrogen concentration risk score R H (t), is calculated by formula (10), which is:

[0050]

[0051] Among them, a H 、b H 、c H ,d H 、e H All are weight coefficients;

[0052] Then according to the temperature risk score R T (t), carbon monoxide concentration risk score R C (t), smoke concentration risk score R S (t) and hydrogen concentration risk score R H (t), calculate the total risk score R(t) of the industrial dust removal facility, which is specifically calculated by formula (11). Formula (11) is specifically:

[0053] R(t)=w T R T (t)+w C R C (t)+w HR H (t)+w S R S (t) (11)

[0054] Among them, w T 、w C 、w H 、w S are all weight coefficients.

[0055] Optionally, the deep learning module performs image processing on the scene image sequence, including:

[0056] The size of each frame of the scene image in the scene image sequence is adjusted to a preset size, and the pixel values ​​in the resized image are scaled to the interval [0,1], and data enhancement is performed on the scaled image to obtain an enhanced image sequence.

[0057] Optionally, the deep learning module performs optical flow calculation and motion feature extraction on the image sequence after image processing to obtain multiple target images including:

[0058] The optical flow calculation and motion feature extraction are performed on the image sequence after image processing. Specifically, the horizontal motion intensity V of each pixel position is extracted between every two consecutive frames in the enhanced image sequence. p (p,q) and vertical motion intensity V q (p,q), and then according to the horizontal motion intensity V p (p,q) and vertical motion intensity V q (p, q), calculate the motion amplitude M(p, q) and motion direction E(p, q) of each pixel, which are calculated by formula (12) and formula (13). Formula (12) and formula (13) are specifically:

[0059]

[0060]

[0061] Among them, (p, q) represents the pixel coordinates in the enhanced image, p represents the horizontal position, q represents the vertical position, V p (p, q) represents the horizontal motion intensity of the pixel (p, q) in the enhanced image, V q (p, q) represents the vertical motion intensity of the pixel (p, q) in the enhanced image, E(p, q) represents the motion direction at the pixel (p, q) in the enhanced image, and M(p, q) represents the motion amplitude at the pixel (p, q) in the enhanced image;

[0062] The motion amplitude M(p, q) and the motion direction E(p, q) are encoded into an RGB image to obtain a target image, and based on this, a plurality of target images are obtained for the image sequence after image processing.

[0063] Optionally, the deep learning module performs model training on a deep neural network model based on spatiotemporal feature fusion according to a training sample set, specifically including:

[0064] For each training sample in the training sample set, the input sample is input into the deep neural network model based on spatiotemporal feature fusion. Specifically, the input sample is subjected to the first convolution layer, batch normalization processing, the second convolution layer, batch normalization processing again, maximum pooling processing, the third convolution layer, global average pooling processing, the fully connected layer and the output layer in the deep neural network model based on spatiotemporal feature fusion to obtain the predicted value of the risk level of the input sample; based on the predicted value of the risk level of the input sample, the output sample and the loss function, the parameters in the deep neural network model based on spatiotemporal feature fusion are updated, and the model parameters in the loss function are also updated. When the loss function is less than a preset threshold, the model training ends;

[0065] The loss function is expressed by formula (14):

[0066]

[0067] in, represents the predicted value of the risk level of the input sample by the deep neural network model based on spatiotemporal feature fusion, x represents the input sample, y represents the output sample, and D train is a training sample set containing multiple training samples, N represents the total number of training samples in the training set, represents y and The loss function between , θ is the model parameter;

[0068] Among them, the model parameter θ in the loss function is updated, which is specifically implemented by formula (15). Formula (15) is specifically:

[0069]

[0070] Among them, θ t represents the model parameters at time t, θ t+1 represents the updated model parameters at time t+1, α represents the learning rate, Denotes the loss function J(θ t ) for the current parameter θ t gradient.

[0071] Among them, the combustible metal dust spontaneous combustion early warning system used in industrial dust removal facilities will accumulate new training samples during operation, and then based on the previous training sample set D train And the new training sample set D new Update the parameters of the deep neural network model based on spatiotemporal feature fusion, and also update the model parameters in the loss function. When the loss function is less than the preset threshold, the model training ends.

[0072] Among them, the model parameters in the loss function are updated by formula (16), and formula (16) is specifically:

[0073] θ new =argmin θ J(θ;D train ∪D new )(16)

[0074] Among them, θ new The deep neural network model based on spatiotemporal feature fusion is trained on the new training sample set D new And the previous training sample set D train The new model parameters are obtained after retraining.

[0075] Optionally, the sensor module is used to collect monitoring parameters in real time, the monitoring parameters including temperature, carbon monoxide concentration, smoke concentration and hydrogen concentration, and then obtain a monitoring parameter time series, the monitoring parameter time series including a temperature time series, a carbon monoxide concentration time series, a smoke concentration time series and a hydrogen concentration time series, and send the monitoring parameter time series to the industrial control unit;

[0076] The image acquisition module is used to collect a sequence of on-site images of the industrial dust removal facility in real time, and send the sequence of on-site images to the deep learning module;

[0077] The sound and light alarm module is used to automatically trigger alarm signals of different levels according to the risk level.

[0078] Optionally, the combustible metal dust spontaneous combustion warning system used in industrial dust removal facilities, after the deep neural network model based on spatiotemporal feature fusion in the deep learning module is trained, uses at least the image acquisition module, the deep learning module, the data storage module and the sound and light alarm module. Specifically, the image acquisition module acquires image sequences in real time, and the acquired image sequences are input into the deep learning module to obtain risk levels corresponding to the acquired image sequences. The sound and light alarm module triggers alarm signals of different levels according to the risk levels corresponding to the acquired image sequences. The data storage module stores the acquired image sequences, the risk levels corresponding to the acquired image sequences and the alarm signals of the sound and light alarm module.

[0079] The beneficial effects of adopting the above technical solution are:

[0080] 1. The present invention provides a combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities. By comprehensively monitoring multiple parameters such as temperature, carbon monoxide concentration, hydrogen concentration, and smoke concentration, and using deep learning technology for intelligent analysis, it realizes comprehensive monitoring and accurate early warning of the risk of spontaneous combustion of combustible metal dust. The system can timely capture small changes in the early stage of spontaneous combustion, accurately predict the risk of spontaneous combustion through multi-parameter correlation analysis, and automatically trigger different levels of alarm signals according to the risk level, which significantly improves the accuracy and timeliness of early warning and effectively reduces the incidence of spontaneous combustion accidents of combustible metal dust.

[0081] 2. The present invention provides a combustible metal dust self-ignition warning system for industrial dust removal facilities, which uses an industrial control unit for data collection and preliminary processing, transmits data in real time to a host computer module and a deep learning module via Ethernet for analysis and processing, and combines an acoustic and visual alarm module to achieve fully automatic and intelligent monitoring and early warning. The system can run 24 hours a day without manual supervision, greatly reducing labor costs, and is particularly suitable for safety monitoring of large-scale production or storage sites, improving the overall safety of industrial production.

[0082] 3. The present invention provides a combustible metal dust spontaneous combustion warning system for industrial dust removal facilities. In the deep learning module, a deep neural network model based on spatiotemporal feature fusion is adopted. By introducing the optical flow method to extract dynamic motion information, intelligent analysis and risk prediction of on-site images are realized. The system also has adaptive optimization capabilities, and continuously optimizes the warning strategy through long-term accumulated historical data, improves the sensitivity and reliability of the system, and provides data support and decision-making basis for industrial safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1This is a schematic structural diagram of a combustible metal dust spontaneous combustion early warning system used in an industrial dust removal facility in an embodiment of the present invention;

[0084] Figure 2 Schematic diagram of the deep learning module processing flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0086] In view of the problems existing in the prior art, the present invention provides a combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities. By comprehensively monitoring environmental parameters such as temperature, carbon monoxide, hydrogen and smoke concentration, combined with a deep learning model based on spatiotemporal feature fusion, the system can analyze data change trends in real time. Through the integration of multi-parameter monitoring and intelligent algorithms, the system can accurately capture the early signs of spontaneous combustion risks, provide risk level classification, and automatically take different measures according to the risk level to ensure the accuracy and timeliness of the early warning.

[0087] Combination Figure 1 , the present invention provides a combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities, including a sensor module, an industrial control unit, a host computer module, an image acquisition module, a deep learning module, a data storage module and an audio-visual alarm module;

[0088] The sensor module is connected to the industrial control unit, that is, the industrial control unit is connected to the sensor module by wire, and the sensor module includes a temperature sensor, a carbon monoxide concentration sensor, a hydrogen concentration sensor and a smoke concentration sensor;

[0089] The sensor module consists of two smoke sensors, two hydrogen sensors, two carbon monoxide sensors and two temperature sensors. The hydrogen sensor is installed in the upper area of ​​the dust collector and near the ash hopper. One thermocouple of the temperature sensor is placed in the dust collector where dust is easily accumulated, and the other is installed in the middle of the filter. Two carbon monoxide sensors are installed in the middle area of ​​the dust collector, and the smoke sensor is installed in the top area of ​​the dust collector and near the air inlet. All sensors are dust-proof and explosion-proof, and are connected to the industrial control unit outside the dust collector through wired connections.

[0090] Wherein, the sensor module is used to collect monitoring parameters in real time, and the monitoring parameters include temperature, carbon monoxide concentration, smoke concentration and hydrogen concentration, so as to obtain a monitoring parameter time series. Specifically, the temperature sensor is used to collect the temperature in the industrial dust removal facility in real time, and the real-time collected temperature forms a temperature time series in chronological order; the carbon monoxide concentration sensor is used to collect the carbon monoxide concentration in the industrial dust removal facility in real time, and the real-time collected carbon monoxide concentration forms a carbon monoxide concentration time series in chronological order; the hydrogen concentration sensor is used to collect the hydrogen concentration in the industrial dust removal facility in real time, and the real-time collected hydrogen concentration forms a hydrogen concentration time series in chronological order; the smoke concentration sensor is used to collect the smoke concentration in the industrial dust removal facility in real time, and the real-time collected smoke concentration forms a smoke concentration time series in chronological order; the temperature time series, the carbon monoxide concentration time series, the hydrogen concentration time series and the smoke concentration time series form a monitoring parameter time series;

[0091] The monitoring parameter time series is sent to the industrial control unit through the connection line between the industrial control unit and the sensor module;

[0092] The industrial control unit includes a signal processing unit and a data transmission unit, wherein the data transmission unit is used to receive the monitoring parameter time series transmitted by the sensor module; the signal processing unit is used to perform analog-to-digital conversion and filtering processing on the monitoring parameter time series to obtain the processed monitoring parameter time series, and the data transmission unit is also used to transmit the processed monitoring parameter time series to the host computer module; in the present invention, the processed monitoring parameter time series can be transmitted to the host computer module through the Ethernet port;

[0093] Among them, the industrial control unit supports multi-tasking and can simultaneously process time series from different sensors in the sensor module.

[0094] The host computer module receives the processed monitoring parameter time series and performs data standardization on the monitoring parameters in the processed monitoring parameter time series, which is specifically implemented by formula (1). Formula (1) is:

[0095]

[0096] Where X' is the monitoring parameter in the standardized monitoring parameter time series, X is the monitoring parameter in the processed monitoring parameter time series, μ is the average value of the monitoring parameter in the processed monitoring parameter time series, and σ is the standard deviation of the monitoring parameter in the processed monitoring parameter time series;

[0097] Based on formula (1), the temperature time series, carbon monoxide concentration time series, smoke concentration time series and hydrogen concentration time series in the monitoring parameter time series are standardized to obtain a standardized temperature time series, a standardized carbon monoxide concentration time series, a standardized smoke concentration time series and a standardized hydrogen concentration time series.

[0098] Then, the standardized monitoring parameter time series is analyzed to obtain the first-order derivative and second-order derivative of the standardized monitoring parameter time series. Time series analysis is to quantify the rate of change and acceleration of data through derivative calculation (including first-order derivative and second-order derivative), so as to better understand the dynamic changes of data, including:

[0099] Calculate the first-order derivative of the standardized monitoring parameter time series, specifically by formula (2), which is:

[0100]

[0101] Among them, X′(t) represents the standardized monitoring parameter time series with the independent variable at time t, X′(t-Δt) represents the standardized monitoring parameter time series with the independent variable at time t-Δt, and Δt is the sampling time interval;

[0102] The second-order derivative of the standardized monitoring parameter time series is calculated by formula (3), which is:

[0103]

[0104] Based on formula (2) and formula (3), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the first-order derivative of the standardized temperature time series: And the second derivative of the standardized temperature time series The first derivative of the normalized carbon monoxide concentration time series And the second derivative of the normalized carbon monoxide concentration time series The first derivative of the normalized smoke concentration time series And the second derivative of the normalized smoke concentration time series The first derivative of the normalized hydrogen concentration time series And the second derivative of the normalized hydrogen concentration time series

[0105] At the same time, feature extraction is performed on the standardized monitoring parameter time series to obtain the moving standard deviation, moving average difference and exponentially weighted moving average of the standardized monitoring parameter time series. Feature extraction is to extract the key features of the standardized monitoring parameter time series, namely, the moving standard deviation, moving average difference and exponentially weighted moving average, which are used to capture the volatility, trend and long-term changes of the data, including:

[0106] The moving standard deviation of the standardized monitoring parameter time series is calculated by formula (4), which is:

[0107]

[0108] Among them, σ x′ (t) is the moving standard deviation of the standardized monitoring parameter time series at time t, Q is the preset time interval, t′ is the time Q seconds before time t, X′(t′) is the monitoring parameter of the standardized monitoring parameter time series at time t′, It represents the average value of the monitoring parameter in the standardized monitoring parameter time series within Q seconds corresponding to time t;

[0109] Based on formula (4), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the moving standard deviation σ of the standardized temperature time series T′ (t), the moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving standard deviation σ of the standardized smoke concentration time series S′ (t) and the moving standard deviation σ of the standardized hydrogen concentration time series H′ (t);

[0110] The moving average difference of the standardized monitoring parameter time series is calculated by formula (5), which is:

[0111]

[0112] Among them, Δ x′ (t) is the moving average difference of the normalized monitoring parameter time series at time t;

[0113] Based on formula (5), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the moving average difference Δ T′(t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the moving average difference Δ of the standardized hydrogen concentration time series H′ (t);

[0114] The exponentially weighted moving average of the standardized monitoring parameter time series is calculated by formula (6), which is:

[0115] μ x′ (t) = βX′(t) + (1-β)μ x′ (t-1) (6)

[0116] Among them, μ X′ (t) is the exponentially weighted moving average of the standardized monitoring parameter time series at time t, β is the smoothing factor, X'(t) is the monitoring parameter of the standardized monitoring parameter time series at time t, μ X′ (t-1) is the exponentially weighted moving average of the normalized monitoring parameter time series at time t-1;

[0117] Based on formula (6), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the exponentially weighted moving average μ of the standardized temperature time series: T′ (t), exponentially weighted moving average μ of the standardized carbon monoxide concentration time series C′ (t), exponentially weighted moving average μ of the standardized smoke concentration time series S′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t).

[0118] Based on this, the total risk score of industrial dust removal facilities is calculated based on the first-order derivative and second-order derivative of the standardized monitoring parameter time series, as well as the moving standard deviation, moving average deviation and exponentially weighted moving average of the standardized monitoring parameter time series, including:

[0119] According to the first derivative of the standardized temperature time series Normalized second derivative of temperature time series The moving standard deviation σ of the standardized temperature time series T′ (t), the moving average difference of the standardized temperature time series Δ T′ (t) and the exponentially weighted moving average μ of the standardized temperature time seriesT′ (t), calculate the temperature risk score R T (t) is calculated by formula (7), which is:

[0120]

[0121] Among them, a T 、b T 、c T ,d T 、e T All are weight coefficients;

[0122] According to the first derivative of the normalized carbon monoxide concentration time series The second derivative of the normalized carbon monoxide concentration time series The moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t) and the exponentially weighted moving average μ of the normalized carbon monoxide concentration time series C′ (t), calculate the carbon monoxide concentration risk score R C (t), is calculated by formula (8), which is:

[0123]

[0124] Among them, a C 、b C 、c C ,d C 、e C All are weight coefficients;

[0125] According to the first derivative of the standardized smoke concentration time series Normalized second-order derivative of smoke concentration time series The moving standard deviation σ of the standardized smoke concentration time series S′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the exponentially weighted moving average μ of the normalized smoke concentration time series S′ (t), calculate the smoke concentration risk score R S (t), is calculated by formula (9), which is:

[0126]

[0127] Among them, a S 、b S 、c S dS 、e S All are weight coefficients;

[0128] According to the first derivative of the normalized hydrogen concentration time series Normalized second derivative of hydrogen concentration time series The moving standard deviation σ of the standardized hydrogen concentration time series H′ (t), moving average difference of the normalized hydrogen concentration time series Δ H′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t), calculate the hydrogen concentration risk score R H (t), is calculated by formula (10), which is:

[0129]

[0130] Among them, a H 、b H 、c H ,d H 、e H All are weight coefficients;

[0131] Then according to the temperature risk score R T (t), carbon monoxide concentration risk score R C (t), smoke concentration risk score R S (t) and hydrogen concentration risk score R H (t), calculate the total risk score R(t) of the industrial dust removal facility, which is specifically calculated by formula (11). Formula (11) is specifically:

[0132] R(t)=w T R T (t)+w C R C (t)+w H R H (t)+w S R S (t) (11)

[0133] Among them, w T 、w C 、w H 、w S are all weight coefficients.

[0134] Then, according to the total risk score, a risk level is determined, and the risk level is low risk, medium risk or high risk;

[0135] Low risk (R(t) is low): indicates that the system as a whole is in a stable state, the sensor data is normal, and the system continues to monitor.

[0136] Medium risk (R(t) is at a medium level): indicates that the temperature, carbon monoxide concentration, hydrogen concentration or smoke concentration monitored by the system has exceeded the normal threshold, and the system is facing potential risks. With the support of the deep learning module, the host computer module determines that the system is in the possible precursor stage of spontaneous combustion, and the system automatically starts the workshop ventilation equipment to enhance air circulation, avoid dust accumulation, and reduce the possibility of local overheating.

[0137] High risk (R(t) is high): High risk indicates that multiple monitoring parameters of the system have been significantly abnormal, such as a significant temperature rise, excessive carbon monoxide and hydrogen concentrations, or smoke detection. The possibility of spontaneous combustion is already high, and the system is in a dangerous state. The host computer module will further weight the sensor input, combined with the analysis of the deep learning module, the system will issue a high-intensity sound and light alarm, automatically shut down the equipment and start the fire extinguishing system, and send an emergency notice to the relevant person in charge.

[0138] The host computer module is also used to display real-time monitoring data, intelligent risk assessment results and alarm information, and receive operator instructions such as system parameter adjustment and manual alarm triggering, and pass these instructions to the corresponding system modules for execution.

[0139] At the same time, the host computer module supports dynamic adjustment of risk level thresholds, and continuously optimizes threshold settings based on the analysis results of historical data to improve the accuracy of early warnings.

[0140] The image acquisition module is used to collect the on-site image sequence of the industrial dust removal facility in real time, and send the on-site image sequence to the deep learning module. The image acquisition module is an industrial camera, which is set at a key position in the combustible metal dust storage and production area, and communicates with the deep learning module through a local area network. The industrial camera obtains the real-time on-site image sequence I at a fixed frame rate f t , where I t Represents the image frame at time t.

[0141] It should be noted that, combined with Figure 1 In the specific implementation process, the image acquisition module can be connected to the host computer module, that is, the image acquisition module communicates with the host computer module through the local area network, and sends the scene image sequence to the host computer module, and the host computer module transmits the scene image sequence to the deep learning module, thereby determining the risk level corresponding to the scene image sequence, but the scene image sequence is not processed in the host computer module, and then the scene image sequence and its corresponding risk level can be transmitted to the deep learning module together.

[0142] Combination Figure 2The deep learning module is used to receive the scene image sequence sent by the image acquisition module, and perform image processing on the scene image sequence. Specifically, the size of each frame of the scene image in the scene image sequence is adjusted to a preset size WxY, and the pixel value in the resized image is scaled to the [0,1] interval, and data enhancement is performed on the scaled image. Specifically, random cropping, rotation, flipping and other data enhancement operations are performed on the scaled image to improve the generalization ability of the model to obtain an enhanced image sequence.

[0143] Among them, the scene image sequence is Figure 2 The risk area image sequence in the scene is processed by image processing. Figure 2 Image preprocessing in .

[0144] Among them, the deep learning module is a high-performance computer equipped with a GPU;

[0145] During the spontaneous combustion process, combustible metal dust often causes phenomena such as rapid increase of local smoke and dust disturbance. Optical flow images can capture these motion characteristics and use them as input data for model training, helping neural networks to identify early signs of dust spontaneous combustion and make timely warnings.

[0146] Then, the present invention performs optical flow calculation and motion feature extraction on the image sequence after image processing to obtain multiple target images, wherein the motion feature extraction is Figure 2 The optical flow feature extraction in , specifically includes:

[0147] For every two consecutive frames of image F in the enhanced image sequence t and F t+1 The horizontal motion intensity V of each pixel position is extracted by p (p,q) and vertical motion intensity V q (p,q).

[0148] Among them, the horizontal motion intensity V at each pixel position p (p,q) and vertical motion intensity V q (p,q) characterizes the flow state of dust and smoke in the system. Before spontaneous combustion of metal dust, local temperature rise and airflow disturbance are usually accompanied, resulting in abnormal changes in the speed and pattern of dust flow and smoke diffusion. The optical flow method can accurately capture these tiny movement changes, especially the abnormal flow in the key parts of the dust collector (filter, ash hopper, air inlet), which are possible early signals of spontaneous combustion.

[0149] In order to further quantify the motion state, according to the horizontal motion intensity V p (p,q) and vertical motion intensity V q(p, q), calculate the motion amplitude M(p, q) and motion direction E(p, q) of each pixel, which are calculated by formula (12) and formula (13). Formula (12) and formula (13) are specifically:

[0150]

[0151]

[0152] Among them, (p, q) represents the pixel coordinates in the enhanced image, p represents the horizontal position, q represents the vertical position, V p (p, q) represents the horizontal motion intensity of the pixel (p, q) in the enhanced image, V q (p, q) represents the vertical motion intensity of the pixel (p, q) in the enhanced image, E(p, q) represents the motion direction at the pixel (p, q) in the enhanced image, and M(p, q) represents the motion amplitude at the pixel (p, q) in the enhanced image;

[0153] The motion amplitude M(p, q) and the motion direction E(p, q) are encoded into an RGB image to obtain a target image, and based on this, a plurality of target images are obtained for the image sequence after image processing.

[0154] For each target image, the target image is used as an input sample of the deep neural network model based on spatiotemporal feature fusion in the deep learning module, the risk level corresponding to the target image is determined according to the upper computer module, and the risk level corresponding to the target image is used as an output sample of the deep neural network model based on spatiotemporal feature fusion in the deep learning module, the input sample and the output sample constitute a training sample, and a plurality of training samples constitute a training sample set, and the deep neural network model based on spatiotemporal feature fusion is trained according to the training sample set to obtain a trained deep neural network model based on spatiotemporal feature fusion, and the image to be detected is identified by the trained deep neural network model based on spatiotemporal feature fusion to obtain the risk level of the image to be detected, and the risk level of the image to be detected is transmitted to the sound and light alarm module;

[0155] The training sample set is divided into training set, validation set and test set in a ratio of 7:2:1, which is used for model training and evaluation. The training set is used to train the deep neural network model based on spatiotemporal feature fusion, the validation set is used to evaluate the model performance, and the model is optimized (adjusting the network structure and modifying the hyperparameters) based on the verification results. The test set is used to conduct a final evaluation of the optimized model to ensure that the model has good generalization ability.

[0156] Among them, the model training of the deep neural network model based on spatiotemporal feature fusion includes:

[0157] Combination Figure 2 For each training sample in the training sample set, the input sample is input into the deep neural network model based on spatiotemporal feature fusion. Specifically, the input sample is passed through the first convolution layer, batch normalization processing, the second convolution layer, batch normalization processing again, maximum pooling processing, the third convolution layer, global average pooling processing, the fully connected layer and the output layer in the deep neural network model based on spatiotemporal feature fusion to obtain the predicted value of the risk level of the input sample, that is, Figure 2 The prediction results in ;

[0158] Specifically, the input sample is processed by the first convolution layer, and multiple convolution kernels of different sizes and directions are used to scan and extract features of the target image. The basic features such as edges, textures, and color gradients in the target image are identified through convolution operations, and the pixel data of the target image is converted into a feature map. The output of the first convolution layer is processed by batch normalization. This step stabilizes the learning process and accelerates training by normalizing the distribution of activation values ​​for each batch, standardizing the feature map, and reducing internal covariate shifts. The batch-normalized feature map passes through the second convolution layer again. This layer uses convolution kernels with larger receptive fields to further extract more complex and abstract features, such as specific shape patterns or local structures, enhancing the model's ability to recognize high-level features. The output of the second convolution layer is batch normalized again to maintain the stability of the deep network. The normalized feature map is downsampled through the maximum pooling operation. This step selects the maximum value of each area in the feature map, which not only reduces the spatial dimension of the feature map, but also reduces the computational cost. The computational complexity is reduced, and the robustness of the model to small image displacements is improved; the downsampled feature map passes through the third convolution layer again to further extract higher-level abstract features, which enables the model to capture more global and semantic image information; the feature map after multiple convolutions and pooling is subjected to the global average pooling operation, which compresses the spatial dimension of each feature map to 1x1, effectively converting the spatial information into a channel descriptor, significantly reducing the number of parameters while retaining the global response of each feature map; the output of the global average pooling is input to the fully connected layer, in which all the extracted features are comprehensively analyzed to establish a complex nonlinear relationship between the features, and realize the mapping from feature representation to task-specific semantics. Finally, the result of the fully connected layer is processed by the output layer, and the predicted value of the risk level of the input sample is generated through the softmax activation function.

[0159] Through this series of processing steps, key features can be gradually extracted and analyzed from the target image, ultimately achieving an accurate assessment of the risk of spontaneous combustion of combustible metal dust. This multi-level, progressive image processing method enables the system to effectively learn and identify complex image patterns related to the risk of spontaneous combustion of combustible metal dust, providing reliable technical support for timely warning and risk management.

[0160] Based on the predicted value of the risk level of the input sample, the output sample and the loss function, the parameters in the deep neural network model based on spatiotemporal feature fusion are updated, and the model parameters in the loss function are also updated. When the loss function is less than the preset threshold, the model training ends;

[0161] The loss function is expressed by formula (14):

[0162]

[0163] in, represents the predicted value of the risk level of the input sample by the deep neural network model based on spatiotemporal feature fusion, x represents the input sample, y represents the output sample, and D train is a training sample set containing multiple training samples, N represents the total number of training samples in the training set, represents y and The loss function between θ and θ is the model parameter;

[0164] Among them, the model parameter θ in the loss function is updated, which is specifically implemented by formula (15). Formula (15) is specifically:

[0165]

[0166] Among them, θ t represents the model parameters at time t, θ t+1 represents the updated model parameters at time t+1, α represents the learning rate, ▽J(θ t ) represents the loss function J(θ t ) for the current parameter θ t gradient.

[0167] Among them, the combustible metal dust spontaneous combustion early warning system used in industrial dust removal facilities will accumulate new training samples during operation, and then based on the previous training sample set D train And the new training sample set D new Update the parameters of the deep neural network model based on spatiotemporal feature fusion, and also update the model parameters in the loss function. When the loss function is less than the preset threshold, the model training ends.

[0168] Among them, the model parameters in the loss function are updated by formula (16), and formula (16) is specifically:

[0169] θ new =argmin θ J(θ;D train ∪D new )(16)

[0170] Among them, θ new The deep neural network model based on spatiotemporal feature fusion is trained on the new training sample set D new And the previous training sample set D train Up, argmin θ : It means finding the parameter θ that can minimize the function J and obtaining the new model parameters after retraining.

[0171] The parameters of the deep neural network model based on spatiotemporal feature fusion are updated regularly so that the model can adapt to dynamic changes in the environment.

[0172] The data storage module is a large-capacity hard disk array, which is used to store historical data of monitoring parameters, and the historical data of monitoring parameters at least includes temperature, carbon monoxide concentration, hydrogen concentration, smoke concentration, total risk score, records of alarm signals triggered by the sound and light alarm module, and on-site image sequences;

[0173] The sound and light alarm module includes an LED indicator light and a high-decibel alarm, which is used to automatically trigger alarm signals of different levels according to the risk level.

[0174] The combustible metal dust spontaneous combustion warning system used in industrial dust removal facilities, after the deep neural network model based on spatiotemporal feature fusion in the deep learning module is trained, at least uses the image acquisition module, the deep learning module, the data storage module and the sound and light alarm module. Specifically, the image acquisition module acquires image sequences in real time, and the acquired image sequences are input into the deep learning module to obtain the risk level corresponding to the acquired image sequences. The sound and light alarm module triggers alarm signals of different levels according to the risk level corresponding to the acquired image sequences. According to the alarm signal, the staff can take measures such as starting ventilation equipment and shutting down equipment. The data storage module stores the acquired image sequences, the risk levels corresponding to the acquired image sequences and the alarm signals of the sound and light alarm module.

[0175] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. A combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities, characterized in that: It includes sensor module, industrial control unit, host computer module, image acquisition module, deep learning module, data storage module and sound and light alarm module; The sensor module is connected to the industrial control unit, and the sensor module includes a temperature sensor, a carbon monoxide concentration sensor, a hydrogen concentration sensor and a smoke concentration sensor; The industrial control unit includes a signal processing unit and a data transmission unit, the data transmission unit is used to receive the monitoring parameter time series transmitted by the sensor module, the monitoring parameter time series includes a temperature time series, a carbon monoxide concentration time series, a smoke concentration time series and a hydrogen concentration time series, the monitoring parameter time series includes monitoring parameters, and the monitoring parameters include temperature, carbon monoxide concentration, smoke concentration and hydrogen concentration; the signal processing unit is used to perform analog-to-digital conversion and filtering processing on the monitoring parameter time series to obtain a processed monitoring parameter time series, and the data transmission unit is also used to transmit the processed monitoring parameter time series to the host computer module; The host computer module receives the processed monitoring parameter time series, and performs data standardization on the monitoring parameters in the processed monitoring parameter time series, and then performs time series analysis on the standardized monitoring parameter time series to obtain the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, and at the same time performs feature extraction on the standardized monitoring parameter time series to obtain the moving standard deviation, the moving average difference and the exponentially weighted moving average of the standardized monitoring parameter time series, and calculates the total risk score of the industrial dust removal facility according to the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, and the moving standard deviation, the moving average difference and the exponentially weighted moving average of the standardized monitoring parameter time series, and then determines the risk level according to the total risk score, and the risk level is low risk, medium risk or high risk; The deep learning module is used to receive the on-site image sequence sent by the image acquisition module, and perform image processing on the on-site image sequence, perform optical flow calculation and motion feature extraction on the image sequence after image processing, and obtain multiple target images. For each target image, the target image is used as an input sample of the deep neural network model based on spatiotemporal feature fusion in the deep learning module, and the risk level corresponding to the target image is determined according to the upper computer module, and the risk level corresponding to the target image is used as the output sample of the deep neural network model based on spatiotemporal feature fusion in the deep learning module. The input sample and the output sample constitute a training sample, and multiple training samples constitute a training sample set. The deep neural network model based on spatiotemporal feature fusion is trained according to the training sample set to obtain a trained deep neural network model based on spatiotemporal feature fusion. The image to be detected is identified by the trained deep neural network model based on spatiotemporal feature fusion to obtain the risk level of the image to be detected, and the risk level of the image to be detected is transmitted to the sound and light alarm module; The data storage module is used to store historical data of monitoring parameters, which include at least temperature, carbon monoxide concentration, hydrogen concentration, smoke concentration, total risk score, records of alarm signals triggered by the sound and light alarm module, and on-site image sequences; The sound and light alarm module includes an LED indicator light and a high-decibel alarm.

2. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 1 is characterized in that: The host computer module performs data standardization on the monitoring parameters in the processed monitoring parameter time series, which is specifically implemented through formula (1), which is: Where X' is the monitoring parameter in the standardized monitoring parameter time series, X is the monitoring parameter in the processed monitoring parameter time series, μ is the average value of the monitoring parameter in the processed monitoring parameter time series, and σ is the standard deviation of the monitoring parameter in the processed monitoring parameter time series; Based on formula (1), the temperature time series, carbon monoxide concentration time series, smoke concentration time series and hydrogen concentration time series in the monitoring parameter time series are standardized to obtain a standardized temperature time series, a standardized carbon monoxide concentration time series, a standardized smoke concentration time series and a standardized hydrogen concentration time series.

3. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 2 is characterized in that: The host computer module performs time series analysis on the standardized monitoring parameter time series to obtain the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, specifically including: Calculate the first-order derivative of the standardized monitoring parameter time series, specifically by formula (2), which is: Among them, X′(t) represents the standardized monitoring parameter time series with the independent variable at time t, X′(t-Δt) represents the standardized monitoring parameter time series with the independent variable at time t-Δt, and Δt is the sampling time interval; The second-order derivative of the standardized monitoring parameter time series is calculated by formula (3), which is: Based on formula (2) and formula (3), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the first-order derivative of the standardized temperature time series: And the second derivative of the standardized temperature time series The first derivative of the normalized carbon monoxide concentration time series And the second derivative of the normalized carbon monoxide concentration time series The first derivative of the normalized smoke concentration time series And the second derivative of the normalized smoke concentration time series The first derivative of the normalized hydrogen concentration time series And the second derivative of the normalized hydrogen concentration time series 4. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 2 is characterized in that: The host computer module performs feature extraction on the standardized monitoring parameter time series to obtain the moving standard deviation, moving average deviation and exponentially weighted moving average of the standardized monitoring parameter time series, specifically including: The moving standard deviation of the standardized monitoring parameter time series is calculated by formula (4), which is: Among them, σ x′ (t) is the moving standard deviation of the standardized monitoring parameter time series at time t, Q is the preset time interval, t′ is the time Q seconds before time t, X′(t′) is the monitoring parameter of the standardized monitoring parameter time series at time t′, It represents the average value of the monitoring parameter in the standardized monitoring parameter time series within Q seconds corresponding to time t; Based on formula (4), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the moving standard deviation σ of the standardized temperature time series T′ (t), the moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving standard deviation σ of the standardized smoke concentration time series S′ (t) and the moving standard deviation σ of the standardized hydrogen concentration time series H′ (t); The moving average difference of the standardized monitoring parameter time series is calculated by formula (5), which is: Among them, Δ x′ (t) is the moving average difference of the normalized monitoring parameter time series at time t; Based on formula (5), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the moving average difference Δ T′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the moving average difference Δ of the standardized hydrogen concentration time series H′ (t); The exponentially weighted moving average of the standardized monitoring parameter time series is calculated by formula (6), which is: m x′ (t)=βX′(t)+(1-β)μ x′ (t-1) (6) Among them, μ X′ (t) is the exponentially weighted moving average of the standardized monitoring parameter time series at time t, β is the smoothing factor, X'(t) is the monitoring parameter of the standardized monitoring parameter time series at time t, μ X′ (t-1) is the exponentially weighted moving average of the normalized monitoring parameter time series at time t-1; Based on formula (6), the standardized temperature time series, the standardized carbon monoxide concentration time series, the standardized smoke concentration time series and the standardized hydrogen concentration time series in the standardized monitoring parameter time series are calculated respectively to obtain the exponentially weighted moving average μ of the standardized temperature time series: T′ (t), exponentially weighted moving average μ of the standardized carbon monoxide concentration time series C′ (t), exponentially weighted moving average μ of the standardized smoke concentration time series S′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t).

5. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 1 is characterized in that: The standardized monitoring parameter time series includes a standardized temperature time series, a standardized carbon monoxide concentration time series, a standardized smoke concentration time series and a standardized hydrogen concentration time series; the first-order derivative and the second-order derivative of the standardized monitoring parameter time series include the first-order derivative of the temperature time series. And the second derivative of the standardized temperature time series The first derivative of the normalized carbon monoxide concentration time series And the second derivative of the normalized carbon monoxide concentration time series The first derivative of the normalized smoke concentration time series And the second derivative of the normalized smoke concentration time series The first derivative of the normalized hydrogen concentration time series And the second derivative of the normalized hydrogen concentration time series The moving standard deviation of the standardized monitoring parameter time series includes the moving standard deviation of the standardized temperature time series σ T′ (t), the moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving standard deviation σ of the standardized smoke concentration time series S′ (t) and the moving standard deviation σ of the standardized hydrogen concentration time series H′ (t); The moving average difference of the standardized monitoring parameter time series includes the moving average difference of the standardized temperature time series Δ T′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the moving average difference Δ of the standardized hydrogen concentration time series H′ (t); The exponentially weighted moving average of the standardized monitoring parameter time series includes the exponentially weighted moving average of the standardized temperature time series μ T′ (t), exponentially weighted moving average μ of the standardized carbon monoxide concentration time series C′ (t), exponentially weighted moving average μ of the standardized smoke concentration time series S′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t); Based on this, the host computer module calculates the total risk score of the industrial dust removal facility according to the first-order derivative and the second-order derivative of the standardized monitoring parameter time series, as well as the moving standard deviation, moving average deviation and exponentially weighted moving average of the standardized monitoring parameter time series, including: According to the first derivative of the standardized temperature time series Normalized second derivative of temperature time series The moving standard deviation σ of the standardized temperature time series T′ (t), the moving average difference of the standardized temperature time series Δ T′ (t) and the exponentially weighted moving average μ of the standardized temperature time series T′ (t), calculate the temperature risk score R T (t) is calculated by formula (7), which is: Among them, a T 、b T 、c T ,d T 、e T All are weight coefficients; According to the first derivative of the normalized carbon monoxide concentration time series The second derivative of the normalized carbon monoxide concentration time series The moving standard deviation σ of the standardized carbon monoxide concentration time series C′ (t), the moving average difference of the standardized carbon monoxide concentration time series Δ C′ (t) and the exponentially weighted moving average μ of the normalized carbon monoxide concentration time series C′ (t), calculate the carbon monoxide concentration risk score R C (t), is calculated by formula (8), which is: Among them, a C 、b C 、c C ,d C 、e C All are weight coefficients; According to the first derivative of the standardized smoke concentration time series Normalized second-order derivative of smoke concentration time series The moving standard deviation σ of the standardized smoke concentration time series S′ (t), the moving average difference of the standardized smoke concentration time series Δ S′ (t) and the exponentially weighted moving average μ of the normalized smoke concentration time series S′ (t), calculate the smoke concentration risk score R S (t) is calculated by formula (9), which is: Among them, a S 、b S 、c S ,d S 、e S All are weight coefficients; According to the first derivative of the normalized hydrogen concentration time series Normalized second derivative of hydrogen concentration time series The moving standard deviation σ of the standardized hydrogen concentration time series H′ (t), moving average difference of the normalized hydrogen concentration time series Δ H′ (t) and the exponentially weighted moving average μ of the normalized hydrogen concentration time series H′ (t), calculate the hydrogen concentration risk score R H (t), is calculated by formula (10), which is: Among them, a H 、b H 、c H ,d H 、e H All are weight coefficients; Then according to the temperature risk score R T (t), carbon monoxide concentration risk score R C (t), smoke concentration risk score R S (t) and hydrogen concentration risk score R H (t), calculate the total risk score R(t) of the industrial dust removal facility, which is specifically calculated by formula (11). Formula (11) is specifically: R(t)=w T R T (t)+w C R C (t)+w H R H (t)+w S R S (t) (11) Among them, w T 、w C 、w H 、w S are all weight coefficients.

6. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 1 is characterized in that: The image processing of the scene image sequence in the deep learning module includes: The size of each frame of the scene image in the scene image sequence is adjusted to a preset size, and the pixel values ​​in the resized image are scaled to the interval [0,1], and data enhancement is performed on the scaled image to obtain an enhanced image sequence.

7. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 1 is characterized in that: The deep learning module performs optical flow calculation and motion feature extraction on the image sequence after image processing to obtain multiple target images including: The optical flow calculation and motion feature extraction are performed on the image sequence after image processing. Specifically, the horizontal motion intensity V of each pixel position is extracted between every two consecutive frames in the enhanced image sequence. p (p,q) and vertical motion intensity V q (p,q), and then according to the horizontal motion intensity V p (p,q) and vertical motion intensity V q (p, q), calculate the motion amplitude M(p, q) and motion direction E(p, q) of each pixel, which are calculated by formula (12) and formula (13). Formula (12) and formula (13) are specifically: Among them, (p, q) represents the pixel coordinates in the enhanced image, p represents the horizontal position, q represents the vertical position, V p (p, q) represents the horizontal motion intensity of the pixel (p, q) in the enhanced image, V q (p, q) represents the vertical motion intensity of the pixel (p, q) in the enhanced image, E(p, q) represents the motion direction at the pixel (p, q) in the enhanced image, and M(p, q) represents the motion amplitude at the pixel (p, q) in the enhanced image; The motion amplitude M(p, q) and the motion direction E(p, q) are encoded into an RGB image to obtain a target image, and based on this, a plurality of target images are obtained for the image sequence after image processing.

8. The combustible metal dust spontaneous combustion early warning system for industrial dust removal facilities according to claim 1 is characterized in that: The deep learning module performs model training on a deep neural network model based on spatiotemporal feature fusion according to a training sample set, specifically including: For each training sample in the training sample set, the input sample is input into the deep neural network model based on spatiotemporal feature fusion. Specifically, the input sample is subjected to the first convolution layer, batch normalization processing, the second convolution layer, batch normalization processing again, maximum pooling processing, the third convolution layer, global average pooling processing, the fully connected layer and the output layer in the deep neural network model based on spatiotemporal feature fusion to obtain the predicted value of the risk level of the input sample; based on the predicted value of the risk level of the input sample, the output sample and the loss function, the parameters in the deep neural network model based on spatiotemporal feature fusion are updated, and the model parameters in the loss function are also updated. When the loss function is less than a preset threshold, the model training ends; The loss function is expressed by formula (14): in, represents the predicted value of the risk level of the input sample by the deep neural network model based on spatiotemporal feature fusion, x represents the input sample, y represents the output sample, and D train is a training sample set containing multiple training samples, N represents the total number of training samples in the training set, represents y and The loss function between θ and θ is the model parameter; Among them, the model parameter θ in the loss function is updated, which is specifically implemented by formula (15). Formula (15) is specifically: Among them, θ t represents the model parameters at time t, θ t+1 represents the updated model parameters at time t+1, α represents the learning rate, Denotes the loss function J(θ t ) for the current parameter θ t The gradient of Among them, the combustible metal dust spontaneous combustion early warning system used in industrial dust removal facilities will accumulate new training samples during operation, and then based on the previous training sample set D train And the new training sample set D new Update the parameters of the deep neural network model based on spatiotemporal feature fusion, and also update the model parameters in the loss function. When the loss function is less than the preset threshold, the model training ends. Among them, the model parameters in the loss function are updated by formula (16), and formula (16) is specifically: i new =argmin θ J(θ;D train ∪D new (16) Among them, θ new The deep neural network model based on spatiotemporal feature fusion is trained on the new training sample set D new And the previous training sample set D train The new model parameters are obtained after retraining.

9. The combustible metal dust spontaneous combustion warning system for industrial dust removal facilities according to claim 1 is characterized in that: The sensor module is used to collect monitoring parameters in real time, the monitoring parameters include temperature, carbon monoxide concentration, smoke concentration and hydrogen concentration, and then obtain a monitoring parameter time series, the monitoring parameter time series includes a temperature time series, a carbon monoxide concentration time series, a smoke concentration time series and a hydrogen concentration time series, and send the monitoring parameter time series to the industrial control unit; The image acquisition module is used to collect on-site image sequences of industrial dust removal facilities in real time, and send the on-site image sequences to the deep learning module; The sound and light alarm module is used to automatically trigger alarm signals of different levels according to the risk level.

10. The combustible metal dust spontaneous combustion warning system for industrial dust removal facilities according to claim 1, characterized in that: The combustible metal dust spontaneous combustion warning system used in industrial dust removal facilities, after the deep neural network model based on spatiotemporal feature fusion in the deep learning module is trained, at least uses the image acquisition module, the deep learning module, the data storage module and the sound and light alarm module. Specifically, the image acquisition module acquires an image sequence in real time, and the acquired image sequence is input into the deep learning module to obtain the risk level corresponding to the acquired image sequence. The sound and light alarm module triggers alarm signals of different levels according to the risk level corresponding to the acquired image sequence. The data storage module stores the acquired image sequence, the risk level corresponding to the acquired image sequence and the alarm signal of the sound and light alarm module.

Images