A hazardous waste package integrity detection and early warning method

Abstract

The present application relates to the technical field of image data processing, and more particularly to a hazardous waste packaging integrity detection and early warning method, which comprises the following steps: collecting multispectral image data of the surface of a hazardous waste packaging container, and obtaining reflectance spectral data by preprocessing; screening the reflectance spectral data to locate suspicious areas; using non-negative matrix factorization to spectrally unmix the reflectance multispectral data of the suspicious areas to obtain an abundance map containing a leakage endmember and a container background endmember; extracting the abundance map of the leakage endmember, calculating a risk quantification index thereof, and forming time series data reflecting the development of the leakage; inputting the time series data into a long short-term memory network model for prediction to obtain the future trend of the risk quantification index; and generating an early warning signal according to the future trend. The present application integrates leakage physical prior knowledge into the algorithm model, and solves the problems of inaccurate early-stage weak leakage identification and high false alarm rate of traditional methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, and in particular to a dangerous waste packaging integrity detection and early warning method. BACKGROUND

[0002] In the field of industrial safety, the sealing of dangerous waste packaging containers is of great importance. Small leaks caused by aging, micro-cracks or corrosion may lead to serious environmental pollution and safety accidents if not discovered in time.

[0003] Traditional leak detection mainly relies on manual visual inspection and simple instrument detection. These two methods have obvious limitations in modern industrial mass production. Manual visual inspection is highly subjective and cannot effectively detect micro-cracks or early signs of leakage that are not visible to the naked eye. Simple instrument detection is sensitive to larger-scale leaks, but its detection sensitivity is low for early-stage micro-amount substance precipitation or small oil film formation, making it difficult to play an effective early warning role.

[0004] In related technologies, spectral analysis technology is used for defect detection. The reflectivity data of the packaging surface is captured through multi-spectral imaging, and then a spectral unmixing algorithm is used to analyze the composition of the surface adhering substances. The non-negative matrix factorization (NMF) algorithm is a commonly used technique in spectral unmixing, which can decompose mixed spectra into endmember spectra of multiple pure substances and their corresponding content ratios, i.e., abundance matrix. However, the traditional non-negative matrix factorization algorithm has inherent defects when applied to early-stage leak detection in this specific scenario. First, the algorithm lacks consideration of the physical characteristics of the leak, which can easily misjudge background noise or spectral fluctuations of the material itself as a leak, resulting in a high false alarm rate. Second, in the early stages of leakage, the spectral signal of the leakage precursor is extremely weak and highly mixed with the background of the packaging material. The traditional NMF algorithm cannot effectively separate the two, resulting in insufficient recognition ability for weak leakage signals. SUMMARY

[0005] To solve the technical problem of the traditional non-negative matrix factorization algorithm lacking consideration of the physical characteristics of the leak in spectral unmixing, easily misjudging background noise, and difficulty in separating the mixed spectrum of weak leakage precursors and packaging materials, affecting the accuracy of detection, the present application provides a dangerous waste packaging integrity detection and early warning method, which comprises the following steps:

[0006] The multispectral image data of the surface of a hazardous waste packaging container is collected and preprocessed to obtain reflectance spectral data; the reflectance spectral data is screened to locate suspicious areas; non-negative matrix decomposition is used to spectrally unmix the reflectance multispectral data of the suspicious areas to obtain an abundance map containing a leakage endmember and a container background endmember; the objective function of the non-negative matrix decomposition contains a fidelity term for ensuring spectral reconstruction accuracy and a regularization term for constraining the spatial distribution of the abundance matrix; the regularization term includes a spatial smoothing constraint term based on the total variation sum of the abundance map and a spatial competitive repulsion constraint term based on the product of the abundance values of the leakage endmember and the container background endmember at the same pixel point; the abundance map of the leakage endmember is extracted, and a risk quantification index thereof is calculated to form time series data reflecting the development of leakage; the time series data is input into a pre-trained long short-term memory network model for prediction to obtain the future trend of the risk quantification index; and an early warning signal is generated according to the future trend.

[0007] The present application introduces spatial smoothing and spatial competitive repulsion dual constraints into the objective function, integrates the physical prior knowledge that the leakage should be continuously distributed in space and cover the original background into the algorithm model, which enables the present application to effectively suppress the interference caused by background noise and material itself fluctuation, improves the recognition accuracy of early weak leakage signals, and reduces the false alarm rate. In addition, the present application further uses the long short-term memory network LSTM to predict the trend of the quantified risk index time series data, realizes the leap from post-discovery to pre-warning, can timely issue an alarm before the leakage problem worsens, and provides a decision-making advance for safety management.

[0008] Preferably, the objective function of the non-negative matrix decomposition satisfies the relationship:

[0009] ;

[0010] Wherein, is the objective function value of the suspicious area; is the spectral matrix of the suspicious area; is the endmember matrix of the suspicious area; is the abundance matrix of the suspicious area; is the square of the Frobenius norm; is the regularization constraint term of the abundance matrix of the suspicious area.

[0011] The present application clearly adds a regularization constraint term, mathematically establishes a basic framework for integrating physical prior knowledge into the optimization objective, so that the optimization direction of the algorithm is no longer only to pursue the fitting accuracy of the data, but also to consider the physical reality of the understanding, thereby improving the reliability of the spectral unmixing result.

[0012] Preferably, the regularization constraint term satisfies the following relation:

[0013] ;

[0014] in, It is the abundance matrix of the suspicious region. Regularization constraints; This is the preset total number of endpoints; It is a total variational operator; It is an abundance matrix The OK; The leakage end-member is in the first Abundance value per pixel, It is an index of the leaked material end-member; The container background terminator is in the first... Abundance values ​​per pixel; It is the index of the container background terminator; , These are the weight coefficients of the first and second terms in the relation, respectively. It represents the total number of pixels within the suspicious area.

[0015] This invention further defines the regularization constraint terms. By using the total variation term and the abundance value product term, the spatial smoothness constraint and the competitive mutual exclusion constraint are transformed into mathematical functions. This combined constraint term, which is designed for the physical characteristics of leakage, can guide the algorithm to separate the leakage area that conforms to the physical laws, ensuring the continuity of the leakage area and clearly defining the boundary between the leakage area and the container background.

[0016] Preferably, the abundance update rule of the leakage endmembers satisfies the following relation:

[0017] ;

[0018] in, The updated index is The leakage end element in the first Abundance values ​​per pixel; It is the endmember matrix of the suspicious region. transpose; It is the spectral matrix of the suspicious region; yes and product The Middle line, number Column elements; It is the abundance matrix of the suspicious regions; yes , , product In the first In the first The element in the first Is a spatial continuity constraint term Partial derivative of the abundance value The value of the first Pixel; Is a preset first infinitesimal value.

[0019] The present application integrates two constraint terms derived from spatial smoothing and competitive exclusion constraint terms in the denominator, which makes the pixel point of the leakage abundance not only its own spectrum like the leakage, but also must not destroy the smoothness of the region, and cannot forcibly increase in the position where the background abundance is already high. This mechanism makes each step of the algorithm iteration towards a more physically reasonable direction, avoiding the problem of isolated noise points being misjudged as leakage.

[0020] Preferably, the reflectance spectral data is screened to locate the suspicious area, comprising: calculating the spectral angle between the reflectance spectral data of each pixel point in the image and the preset reference spectrum, and determining the pixel point with the spectral angle greater than the preset angle threshold as the suspicious pixel; calculating the first order derivative of the spectrum of the suspicious pixel, and comparing the similarity with the pre-stored first order derivative spectrum template of the known leakage substance, and recording the area formed by the suspicious pixel with the similarity higher than the preset similarity threshold as the suspicious area.

[0021] The present application firstly performs a wide range of preliminary screening through the fast spectrum angle matching calculation, and excludes most of the normal background; then only a small amount of suspicious pixels are subjected to more detailed spectral derivative matching for secondary confirmation, and this coarse screening plus fine screening strategy reduces the amount of data that needs to be calculated subsequently.

[0022] Preferably, the similarity is the Pearson correlation coefficient method.

[0023] Preferably, the method for obtaining the first order derivative spectrum template is: performing spectral measurement on the known leakage substance to obtain its standard spectrum, and calculating the first order derivative of the standard spectrum.

[0024] Preferably, the method for obtaining the preset reference spectrum comprises: selecting a non-polluted background area in the currently collected multi-spectral image data, and taking the average value of all pixel spectra in the background area as the preset reference spectrum.

[0025] Preferably, the risk quantification index comprises: the total abundance value of the leakage endmember abundance map, or the maximum connected region area of the leakage endmember abundance map.

[0026] Preferably, generating the early warning signal according to the future trend comprises: generating a first-level early warning when the future trend shows that the risk quantification index will continue to grow; and generating a second-level early warning when the future trend shows that the risk quantification index will exceed a preset safety threshold.

[0027] The present application has the following advantages: firstly, the present application obtains suspicious areas by combining spectral angle screening and spectral derivative screening, thereby reducing the amount of data that needs to be calculated by NMF subsequently; secondly, the present application performs non-negative matrix factorization on the reflectance spectral data of the suspicious areas, and then introduces spatial smoothing and spatial competitive repulsion double constraints into the objective function of spectral unmixing, thereby incorporating the physical prior knowledge that the leakage should be continuously distributed in space and cover the original background into the algorithm model, which makes the present application capable of effectively suppressing the interference caused by background noise and fluctuations of the material itself, improving the recognition accuracy of early weak leakage signals, and reducing the false alarm rate; and thirdly, the present application uses a long short-term memory network (LSTM) to perform trend prediction on the quantified risk index time series data, thereby realizing the leap from post-discovery to pre-warning, and being capable of issuing an alarm in a timely manner before the leakage problem worsens, and providing a decision-making lead time for safety management. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A flowchart of a hazardous waste packaging integrity detection and early warning method provided by an embodiment of the present application is shown in FIG. 1. DETAILED DESCRIPTION

[0029] An embodiment of the present application provides a hazardous waste packaging integrity detection and early warning method, as shown in FIG. 1, which comprises steps S100-S400. Figure 1

[0030] Step S100: collect multi-spectral image data of the surface of a hazardous waste packaging container, and pre-process the multi-spectral image data to obtain reflectance spectral data.

[0031] It should be noted that this step is the data acquisition basis of the entire detection and early warning process, and the spectral reflectance information of the surface of the packaging container is obtained by collecting images of the packaging container at multiple discrete wavebands, thereby providing prerequisite data support for subsequent identification of the optical property changes caused by the leakage precursors.

[0032] In a specific operation, a multi-spectral camera covering the ultraviolet, visible light and near-infrared wavebands is used to periodically scan and image the surface of the hydraulic oil packaging container; the periodic scanning is used to continuously monitor the dynamic changes of the spectral data of the container surface, so as to capture the early signs of the adsorption or deposition of the leakage precursors in a timely manner.

[0033] ​In addition, in order to eliminate the influence of ambient light changes and camera sensor dark current noise on data accuracy, the collected raw light intensity data needs to be corrected to reflectivity, so that the original multi-dimensional light intensity data is converted into a standardized reflectivity spectral data cube. The data cube takes pixels as the basic unit, and the spectral information of each pixel point corresponds to a vector containing several different band reflectivity values. This vector is the spectral fingerprint required for subsequent analysis. When the spectral fingerprint of a certain pixel point changes, it indicates that the optical properties of the packaging container surface have changed, and it is highly likely that the leakage of the precursor substance has caused adsorption or deposition on the surface. This change will serve as a key basis for subsequent feature extraction and risk judgment. How to correct reflectivity is known in the prior art, and will not be described here.

[0034] At this point, the reflectivity spectral data is obtained.

[0035] Step S200, screening the reflectivity spectral data to locate suspicious areas.

[0036] It should be noted that, considering the huge amount of multi-spectral image data, directly performing complex spectral unmixing operation is inefficient. Moreover, the leakage precursor usually only appears in a local area, therefore, it is necessary to screen out the area with obvious spectral difference from the normal background from the whole image, which is referred to as a suspicious area, so as to reduce the range of subsequent fine analysis and improve the detection efficiency. In this embodiment, a screening strategy combining spectral angle mapping screening and spectral derivative screening is used.

[0037] For spectral angle mapping screening, under normal and clean conditions, the surface of the same packaging container or the same batch of packaging containers has consistency in material and state, and therefore the shape of its spectral reflectance curve should also be highly similar. When a small amount of leakage precursor substance such as oil film formed by condensation of volatile organic compounds or salt generated by acid mist corrosion adheres to the surface, it will change the spectral response characteristics of the local area, specifically the change of the shape of the spectral curve. The spectral angle is an effective indicator for measuring the shape similarity of two spectral vectors, and is not sensitive to the brightness change of the whole spectrum, and is suitable for identifying such spectral shape variation caused by changes in material composition.

[0038] According to the above logic, taking any pixel point as a pixel to be detected, the spectral angle between the spectrum of the pixel to be detected and the preset normal background reference spectrum is calculated to judge the abnormality degree. The spectral angle satisfies the relationship:

[0039] ;

[0040] Among them, is the spectral angle between the spectral vector of the pixel to be detected and the preset normal background reference spectral vector, and its value range is ; is the spectral vector of the pixel to be detected, ,in For this pixel in the th Reflectivity of each band This represents the total number of spectral bands. It is a preset normal background reference spectral vector; as a preferred implementation, the reference spectrum can be obtained by selecting a confirmed uncontaminated and representative background area in the current image and calculating the average value of the spectrum of all pixels in the area, so as to adapt to changes in current lighting and environment. , They are vectors and The L2 norm is used to normalize the spectral vector to eliminate overall brightness differences caused by uneven illumination or surface angle. It is an inverse cosine function.

[0041] In this relationship, the closer the spectral shape of the pixel to be detected is to the reference spectrum, the more consistent the directions of the two vectors are, and the closer their inner product, after norm normalization, is to 1, thus making the calculated spectral angle more consistent. The closer it is to 0. Conversely, when a pixel is contaminated by the precursor substance, its spectral shape changes, leading to a change in the vector. and The direction deviates. The value increases. Therefore, The value quantifies the difference in spectral shape between the region where the pixel to be detected is located and the normal background.

[0042] By performing the above operation on each pixel in the image, a spectral angle plot can be generated. Analyzing the pixel value histogram of this plot usually reveals two main peaks: one corresponding to a large area of ​​normal background, and the other... The values ​​are concentrated in the main peak of the low-value region because the normal background accounts for a much larger proportion of the image than the local abnormal region, resulting in a large number of pixels with low θ values; the other corresponds to a small number of abnormal regions. The values ​​are distributed in the secondary peak of the high-value region because the abnormal region accounts for a small proportion and the number of pixels with high θ values ​​is small.

[0043] As a preferred implementation, the segmentation threshold used to distinguish between normal and abnormal regions can be set to a value slightly larger than the distribution range of the main peak. For example, the segmentation threshold can be set to the mean of the spectral angle distribution of the normal background region plus three times its standard deviation. All pixels with spectral angles greater than the segmentation threshold will be marked as preliminary suspicious regions and enter the next screening process.

[0044] For spectral derivative screening, after initially identifying suspected regions through spectral angle mapping, this step further refines them. Many common leakage precursors, such as specific solvents, additives, and acidic substances, possess specific chemical bonds in their molecular structures. These chemical bonds generate characteristic absorption within specific wavelength ranges, forming weak absorption valleys in the reflectance spectrum. These absorption valleys may be difficult to detect in the original spectrum due to weak signals or being masked by noise, but in the first derivative curve of the spectrum, they transform into zero-crossing or peak-valley features, thus being effectively amplified and identified. Based on the above logic, this step performs the following operations on each pixel within the initially suspected region:

[0045] First, extract the complete spectral vector of the pixel to be detected. As input, the first derivative of this spectral vector is then calculated to generate a new first-derivative spectrum that amplifies the absorption characteristics. This process calculates reflectance. Along the spectral dimension band number The rate of change.

[0046] Preferably, the central difference method is used for calculation:

[0047] ;

[0048] in, It is the first The pixel in the first The first-order spectral derivative values ​​for each band; It is the first The pixel in the first Reflectivity of each band; It is the spectral band number; It is the band increment, usually taken as 1.

[0049] This invention utilizes all wavebands By performing this calculation, we can obtain the value belonging to the th A complete first-order derivative spectral vector of pixels. First derivative spectral vector Each value in the curve reflects the original spectral curve. The intensity of the upward or downward trend in the corresponding wave band.

[0050] The first derivative spectrum of the pixel to be detected is calculated. Then, it is matched with the first derivative spectral templates of known leaking substances, such as specific VOCs, acid mist, oil film, etc., which are pre-stored in the spectral fingerprint database. This spectral fingerprint database is constructed by experimentally measuring and storing the first derivative spectra of known leaking substances.

[0051] As a feasible implementation, the similarity matching here is realized by calculating the Pearson correlation coefficient The calculated correlation coefficient is compared with a preset similarity threshold As a preferred implementation, the similarity threshold can be set to When is lower than , it is considered that the spectrum of the pixel to be detected is significantly different from the template spectrum, which may be caused by other types of noise or non-target substances, and is determined as a non-target anomaly; when is higher than the threshold, it is considered that the spectral derivative feature of the pixel to be detected is highly matched with the fingerprint feature of the known leakage substance, for example, the threshold is set to , which can balance the sensitivity and specificity of detection, effectively exclude the interference of background noise, and at the same time ensure the reliable identification of the target leakage precursor.

[0052] Finally, all the pixels to be detected with a similarity greater than the threshold are confirmed as highly suspicious pixels, and the region composed of these highly suspicious pixels is the final screened suspicious region. Through the above two-step screening, the data amount that needs to be operated complexly can be reduced while ensuring the detection rate.

[0053] Step S300, using non-negative matrix factorization to perform spectral unmixing on the reflectance multispectral data of the suspicious region to obtain an abundance map containing leakage endmembers and container background endmembers; the objective function of the non-negative matrix factorization contains a fidelity term for ensuring spectral reconstruction accuracy and a regularization term for constraining the spatial distribution of the abundance matrix.

[0054] It should be noted that the spectral signal in the suspicious region screened is usually a linear mixture of spectral features of multiple substances, that is, endmembers, and the endmember refers to a single substance with pure spectral characteristics, such as a pure packaging container base material, a pure adsorbed leakage precursor substance, and pure air dust.

[0055] Non-negative matrix factorization (NMF) algorithm is a commonly used spectral decomposition technique, which can decompose a complex mixed spectral matrix into the product of two non-negative matrices, and the two matrices correspond to the pure endmember spectral matrix and the abundance matrix respectively, and the decomposition process conforms to the physical nature of the non-negative spectral signal, can avoid the appearance of negative coefficients without practical significance, and has natural adaptability when dealing with multi-component mixed spectrum. However, the traditional NMF algorithm has limitations when applied to leakage detection scene, it only takes the minimization of the error between the mixed spectrum and the reconstructed spectrum after decomposition as the target, without considering the physical characteristics of the leakage process, which is easy to lead to the endmember abundance distribution obtained by decomposition not conforming to the spatial distribution rule of leakage, for example, discrete and non-continuous high-value area of abundance, which cannot reflect the real leakage. Therefore, the target function and updating rule of the traditional non-negative matrix factorization are optimized.

[0056] The non-negative matrix factorization technology is adopted to decompose the mixed spectrum matrix of the screened suspicious area into an abundance matrix and an endmember matrix, realize spectral unmixing, and the decomposition formula satisfies the relationship:

[0057] ;

[0058] Among them, is the spectral matrix of the suspicious area; is the abundance matrix of the suspicious area; is the endmember matrix of the suspicious area. The specific decomposition process is the prior art, and will not be described here.

[0059] Specifically, the NMF target function f (W, H) is constructed by f (W, H) = ||X- WH||2 + λ1||WH||1 + λ2||WH||2, wherein the spatial double constraints correspond to the spatial sparsity and spatial continuity of the leakage respectively. The regularization term is added to the reconstruction error term of the traditional NMF, the regularization term is used to constrain the solution which does not conform to the physical characteristics of the leakage, and then guide the decomposition result to tend to the physical real situation, so that the decomposition result is more suitable for the actual spatial distribution and material composition of the leakage.

[0060] According to the above logic, the improved target function satisfies the relationship:

[0061] ;

[0062] Among them, is the target function value of the suspicious area, representing the overall optimization target of spectral unmixing, and the target of the algorithm is to find the endmember matrix and the abundance matrix which make minimum; is the spectral matrix of the suspicious area, and the dimension is , is the number of spectral bands, is the total number of pixels in the suspicious area; is the endmember matrix of the suspicious area, and the dimension is , This is the preset total number of endpoints; This is the abundance matrix of the suspicious regions, with dimensions of . ; It is the square of the Frobenius norm; It is the abundance matrix of the suspicious region. The regularization constraint term is used to embed the physical characteristics of leakage, and the specific calculation method is detailed below.

[0063] In this relation, the first term This is the data fidelity item, used to ensure the reconstruction accuracy of spectral data while maintaining the basic decomposition capability of the NMF algorithm. (Second item) This is the regularization term introduced in this invention. By applying dual constraints, the decomposition results are more reasonable at the physical level, which can effectively avoid the large-area fuzzy solutions or isolated false points that may be generated by the traditional NMF method, which do not conform to the physical process of leakage spreading from a local point source.

[0064] It should be noted that leakage, as a fluid diffusion process, should result in a contaminated area that is continuous and smooth in space, rather than isolated, randomly distributed noisy spots. At any physical point, when leakage occurs, it will inevitably cover or replace the original packaging container. This means that the abundance of the leakage and the container background are mutually exclusive and cannot coexist in large quantities at the same time.

[0065] Based on this, the constraint terms of the abundance matrix of the suspicious region satisfy the following relation:

[0066] ;

[0067] in, It is the abundance matrix of the suspicious region. Regularization constraints; This is the preset total number of endpoints; It is a total variation operator that acts on the abundance map of a single endmember to calculate the L1 norm of its gradient, which is used to measure the smoothness of the abundance map. The smoother the abundance map, the lower its total variation value. It is an abundance matrix The Line, it constitutes the first Abundance map of endmembers throughout the suspected region; The leakage end-member is in the first Abundance value per pixel, It is an index of the leaked material end-member; The container background terminator is in the first... Abundance values ​​per pixel; It is the index of the container background terminator; , are the weight coefficients of the first and second terms of the relationship respectively; is the total number of pixels in the suspicious region.

[0068] In the relationship, the first term is the spatial continuity constraint term, by minimizing the total sum of total variation of all endmember abundance maps, it promotes the generated abundance distribution of each material to be more smooth and continuous in space, effectively inhibiting the generation of isolated noise points. The second term is the competitive repulsion constraint term introduced by the present application, by minimizing the competitive repulsion term, the product of each pair tends to zero, which means that at any pixel point , a clear two-choice decision must be made between the leakage and the container background, and ambiguity cannot be tolerated, i.e., both of their abundances are high at the same time, thus converting the physical phenomenon of leakage covering the background into a mathematical constraint, so that the boundary of the leakage region in the unmixing result has a more clear physical meaning.

[0069] It should be noted that the weight coefficients and the values of these two parameters need to be balanced between data fidelity, spatial smoothness and material repulsion.

[0070] Specifically, the weight coefficient controls the strength of the spatial continuity constraint, if it is set too large, it may cause the abundance map to be overly smooth, blurring the fine boundary of the leakage region; if it is set too small, the constraint will not be effective enough to suppress isolated noise points. The weight coefficient determines the strength of the competitive repulsion constraint, if it is set too large, it may force the distinction between the leakage and the background at the early stage of decomposition, affecting the stable convergence of the algorithm; if it is set too small, it cannot effectively realize the repulsion between the two endmember abundances, resulting in an unclear physical meaning of the decomposition result.

[0071] As a preferred embodiment, and the specific values can be determined by cross-validation.

[0072] For example, on a sample data set containing known leakage conditions, by methods such as grid search, systematically test multiple groups of combinations, and select the one that optimizes the unmixing result as the final setting value. In the absence of a large number of prior samples for training, it can also be set according to experience. For example, in order to effectively apply the constraint while ensuring the stability of the algorithm, the preferred range of to and The preferred range can be set as to By such setting, a good balance between spectral reconstruction accuracy and physical constraint rationality can be obtained. The grid search method and the like are prior art, and will not be described herein.

[0073] Up to now, the optimized objective function is obtained.

[0074] It should be noted that, in order to solve the objective function containing double constraints, the present application adopts a multiplication update rule to iteratively optimize the abundance matrix This method is derived from the gradient descent idea, by splitting the gradient of the objective function with respect to the element in the abundance matrix, and constructing a multiplication iterative form with non-negative characteristics, so as to drive the objective function to converge to the minimum value while ensuring the non-negativity of the abundance value.

[0075] Specifically, the update rule of the present application contains two parts, which are the update rules of the leak endmember and the container background endmember abundance, which will be described in the following.

[0076] For the update rule of the leak abundance, the purpose is to integrate the mathematical accuracy of spectral unmixing and the physical law of leak diffusion, on the one hand, based on the similarity between the leak endmember and the pixel spectrum, to ensure that the abundance growth always relies on the real spectral signal, on the other hand, through the triple key constraints to form an iterative framework, both to retain the standard spectral reconstruction correlation term to ensure the mathematical accuracy of unmixing, and to add the smoothness constraint to force the leak abundance to present a continuous distribution, which is consistent with the spatial characteristics of the leak as fluid diffusion, and the mutual exclusion constraint is introduced to realize the mutual elimination of the leak and the container background abundance, which is consistent with the physical logic of the leak covering the background. In this process, the multiplication iterative form is adopted to ensure the non-negativity and iterative convergence of the abundance, and finally the problems such as discrete false points, abundance double high ambiguity solution and other problems that violate the physical common sense in the traditional NMF in leak detection are solved, so that the unmixing result not only meets the spectral accuracy requirement, but also matches the actual leak scene.

[0077] According to the above logic, the update rule of the leak abundance satisfies the relationship:

[0078] ;

[0079] Wherein, is the abundance value of the updated leak endmember with index at the th pixel; is the abundance value of the leak endmember with index at the th pixel in the current iteration step; is the endmember matrix of the suspicious region ; is the spectral matrix of the suspicious region ; is the product of ; , the element in the i-th row and j-th column of , measures the similarity between the endmember spectrum and the spectrum of the j-th pixel ; is the abundance matrix of the suspicious region ; is the product of , , ; , the element in the i-th row and j-th column of , represents the reconstruction of the original spectrum based on the current solution, is the denominator term in the traditional NMF update rule is the spatial continuity constraint term ; , the value of the partial derivative of the abundance value with respect to the i-th pixel, quantifies the impact of the change of on the smoothness of the entire spill abundance map is the abundance value of the i-th pixel for the container background endmember with index ; , , are the weight coefficients of the spatial continuity constraint term and the competitive repulsion constraint term, respectively is a preset first infinitesimal value, used to prevent the denominator from being zero, which can be set to 0.001 or adjusted according to requirements

[0080] In this relationship, three thresholds are set for the growth of spill abundance: first, the numerator requires that the spectral characteristics of the pixel itself match the spectral fingerprint of the spill, which is the basis for growth; second, the term in the denominator constitutes a smoothness constraint, if the growth of will destroy the local smoothness of the abundance map it belongs to, the derivative value will become larger, thus increasing the denominator and inhibiting its growth; the term in the denominator constitutes a competitive constraint, on the same pixel , if the abundance of the container background is already high, this term will become a large constraint value, increasing the denominator and thus inhibiting the growth of the spill abundance .

[0081] Similarly, the update rule for the container background abundance also follows a symmetric logic, and its update is competitively inhibited by the spill abundance.​

[0082] According to the above logic, the update rule of the container background abundance satisfies the following relationship:

[0083] ;

[0084] wherein, is the abundance value of the container background endmember with index at the th pixel after update; is the abundance value of the container background endmember with index at the th pixel in the current iteration step; is the transpose of the endmember matrix of the suspicious region; is the spectral matrix of the suspicious region; is the product of and , which is the element in the th row and the th column of the matrix , and this term is used to measure the similarity between the spectrum of the container background endmember and the spectrum of the th pixel; is the abundance matrix of the suspicious region; is the product of , , , which is the element in the th row and the th column of the matrix , and this term represents the reconstruction of the original spectrum based on the current solution, which is the denominator term in the traditional NMF update rule; is the partial derivative of the spatial continuity constraint term with respect to the abundance value at the th pixel, which quantifies the influence of the change of on the smoothness of the entire container background abundance map; is the abundance value of the leak endmember with index at the th pixel; , are the weight coefficients of the spatial continuity constraint term and the competitive exclusion constraint term, respectively; is a preset second infinitesimal value used to prevent the denominator from being zero, which can be set to 0.001 or adjusted according to requirements.

[0085] In this relationship, the term in the denominator constitutes a competitive penalty for the growth of the container background abundance, and when the leak abundance is high, the growth of the abundance of the package will be inhibited.

[0086] In summary, by the above two symmetric multiplication update rules containing double constraints, the application constructs a dynamic iterative system, in which the abundances of each pixel point and the container background are in mutual competition and mutual restraint in each iteration. This update mechanism ensures that the abundance map obtained after the algorithm converges is not only spatially continuous and smooth, but also clearly mutually exclusive in material composition, thereby restoring the physical reality of the seepage covered background.

[0087] By iteratively calculating the above update rule until the abundance matrix converges, the final abundance map containing seepage information can be obtained.

[0088] Step S400, extract the abundance map of the seepage endmember and calculate its risk quantification index to form time series data reflecting the development of seepage; input the time series data into the pre-trained long short-term memory network model for prediction to obtain the future trend of the risk quantification index; generate an early warning signal according to the future trend.

[0089] It should be noted that the abundance value detected at a single time can only reflect the static pollution condition at the current time, while seepage is a dynamic development process, and the real risk lies in the continuous growth of the pollution area or concentration. Therefore, this step performs time series analysis on the seepage abundance data obtained at consecutive time points to determine its development trend, thereby realizing the upgrade from static detection to dynamic early warning.

[0090] In specific operation, after completing the detection process of step S100-step S300 each time, the abundance map representing the seepage is extracted from the final obtained abundance matrix. Subsequently, the key indicators in the abundance map are calculated as the risk quantification value at the current time.

[0091] As a preferred embodiment, the key indicator can be selected as the total abundance value or the maximum connected region area. The total abundance value is the sum of all pixel abundance values in the abundance map, reflecting the total volume or total area of pollution; the maximum connected region area reflects the size of the largest single pollution region.

[0092] ​The risk quantification value calculated for each period is stored in a time series database in chronological order to form a time series reflecting the development process of the leakage. Then the time series is input into a pre-trained long short-term memory (LSTM) model. LSTM is a recurrent neural network that is good at learning and predicting long-term dependencies of time series. Through training on a large amount of real or simulated leakage development data, the model can accurately capture various growth patterns of the leakage from nothing to something and from slow to fast. The LSTM model receives a sequence of risk quantification values in the past, for example, in the past 24 hours, as input and outputs a prediction of the risk quantification value in the future, for example, in the next 1-2 hours. The LSTM model is prior art and will not be described in detail here.

[0093] Finally, based on the prediction results of the LSTM, a hierarchical warning strategy is executed:

[0094] Primary warning: When the model predicts that the future risk quantification value will show a continuous and stable growth trend, even if the current value is still below the warning line, the system will mark the container as a potential risk point and highlight it on the monitoring interface to prompt the management personnel to pay attention.

[0095] Secondary warning: When the model predicts that the future risk quantification value will exceed the pre-set safety threshold, the system will immediately trigger a high-level alarm and also push it to the relevant safety management personnel so that they can take timely intervention measures. The setting of the safety threshold can be determined comprehensively according to historical safety data, the physicochemical properties of the specific hazardous waste such as the volatilization rate, the corrosion level or relevant safety production management regulations. For example, for high-volatile organic solvents, the safety threshold can be set to a low value to ensure that the early stage of leakage can trigger a warning.

[0096] The above are preferred embodiments of the present application, but do not limit the protection scope of the present application, therefore: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A hazardous waste package integrity detection and warning method, characterized in that, The method comprises the steps of: collecting multi-spectral image data of the surface of a hazardous waste packaging container and pre-processing the same to obtain reflectance spectral data; screening the reflectance spectral data to locate a suspicious area; performing spectral unmixing on the reflectance multi-spectral data of the suspicious area by using non-negative matrix factorization to obtain an abundance map containing a leakage endmember and a container background endmember; a target function of the non-negative matrix factorization contains a fidelity term for ensuring spectral reconstruction accuracy and a regularization term for constraining the spatial distribution of the abundance matrix; the regularization term includes a spatial smoothing constraint term based on the total variation sum of the abundance map and a spatial competitive repulsion constraint term based on the product of the abundance values of the leakage endmember and the container background endmember at the same pixel point; The objective function of the non-negative matrix factorization satisfies the relationship: ; wherein, is the objective function value of the suspicious region; is the spectral matrix of the suspicious region; is the endmember matrix of the suspicious region; is the abundance matrix of the suspicious region; is the square of the Frobenius norm; is the regularization constraint term of the abundance matrix of the suspicious region; The regularization constraint term satisfies a relationship: ; wherein, is a preset total number of end members; is a total variation operator; is an abundance matrix of the first row; is an abundance value of a leaky end member at a first pixel; is an index of the leaky end member; is an abundance value of a container background end member at a first pixel; is an index of the container background end member; , are weight coefficients of the first term and the second term in the relationship, respectively; is a total number of pixels in the suspicious region; the abundance update rule of the leakage endmember satisfies the relationship: ； wherein, is the abundance value of the leaky endmember at the th pixel; is the transpose of the endmember matrix of the suspicious region; is the product of and is the element in the th row, the th column of the matrix is the product of , , is the element in the th row, the th column of the matrix is the partial derivative of the spatial continuity constraint term with respect to the abundance value at the th pixel; is a preset first infinitesimal value; an abundance map of the leaky endmember is extracted, and a risk quantification index thereof is calculated to form time series data reflecting leak development; the time series data is input into a pre-trained long short-term memory network model for prediction to obtain a future trend of the risk quantification index; and a warning signal is generated according to the future trend. ​​​ 2. The method of claim 1, wherein the method further comprises: The screening of the reflectance spectral data to locate a suspicious area comprises: calculating the spectral angle between the reflectance spectral data of each pixel in the image and a preset reference spectrum, and determining the pixel with a spectral angle greater than a preset angle threshold as a suspicious pixel; calculating the first-order derivative of the spectrum of the suspicious pixel, and comparing it with a pre-stored spectral first-order derivative template of a known leakage substance; the region formed by the suspicious pixels with a similarity higher than a preset similarity threshold is recorded as a suspicious area.

3. The method of claim 2, wherein the method further comprises: The similarity is a Pearson correlation coefficient method.

4. The method of claim 3, wherein the method further comprises: The method for obtaining the spectral first-order derivative template comprises: performing spectral measurement on the known leakage substance to obtain a standard spectrum, and calculating the first-order derivative of the standard spectrum.

5. The method of claim 3, wherein the method further comprises: The method for obtaining the preset reference spectrum comprises: selecting a non-polluted background area in the currently collected multi-spectral image data, and taking the average value of the spectrum of all pixels in the background area as the preset reference spectrum.

6. The hazardous waste package integrity detection and warning method of claim 1, wherein, The risk quantification index comprises: the total abundance value of the leakage endmember abundance map, or the area of the largest connected region of the leakage endmember abundance map.

7. The method of claim 1, wherein the method further comprises: Generating an early warning signal according to the future trend comprises: generating a first-level early warning when the future trend shows that the risk quantification index will continue to grow; generating a second-level early warning when the future trend shows that the risk quantification index will exceed a preset safety threshold.

Images