Artificial intelligence-based chemical substance rapid identification method and system

Abstract

The application provides a kind of based on artificial intelligence chemical substance fast identification method and system, belong to intelligent detection analysis technical field;The method includes the preprocessing and structural conversion processing of physical signal and social context information, extracts the physical evidence characteristics and social evidence characteristics of the sample to be measured;According to physical evidence characteristics and social evidence characteristics, through double drive hypothesis generation, the bayesian inference network containing candidate substance hypothesis node, physical evidence node and social evidence node is constructed;Solving the posterior probability value corresponding to each candidate substance hypothesis, according to the posterior probability value and evidence chain analysis result executes decision, generates and outputs chemical substance identification report.The application generates double drive hypothesis by fusing physical signal and social context information, and uses bayesian network to carry out probability inference, which significantly improves the accuracy and reliability of chemical substance identification in complex scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and analysis technology, specifically to a method and system for rapid identification of chemical substances based on artificial intelligence. Background Technology

[0002] In recent years, with the development of artificial intelligence technology, especially the application of machine learning algorithms in spectral analysis, some intelligent substance identification methods have emerged. These methods automatically learn features in spectra or mass spectra by training deep learning models, thus improving the automation and speed of identification to some extent. For example, some systems can perform pattern recognition on input infrared spectra and directly output the most likely substance name. However, these methods view the detection signal itself in isolation, leading to poor interpretability of the decision-making process. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for rapid identification of chemical substances based on artificial intelligence. This method integrates the physical signals and social context information of the sample under test, and performs collaborative analysis and probability calculation in an interpretable dynamic Bayesian inference network to achieve rapid and reliable automated identification of chemical substances.

[0004] To achieve the above objectives, this invention provides a rapid chemical substance identification method based on artificial intelligence, comprising: collecting physical signals and socio-contextual information of a sample to be tested; preprocessing and structuring the physical signals and socio-contextual information to extract physical evidence features and socio-evidence features of the sample to be tested; constructing a Bayesian inference network containing candidate substance hypothesis nodes, physical evidence nodes, and socio-evidence nodes based on the physical evidence features and socio-evidence features through dual-driven hypothesis generation; performing abductive calculation and probabilistic inference based on the Bayesian inference network to solve for the posterior probability value corresponding to each candidate substance hypothesis and generate evidence chain analysis results; executing a decision based on the posterior probability value and the evidence chain analysis results, generating and outputting a chemical substance identification report, and completing the identification of the chemical substance in the sample to be tested.

[0005] Optionally, the step of acquiring the physical signal and socio-contextual information of the sample to be tested, and preprocessing and structuring the physical signal and socio-contextual information, includes: acquiring at least one original physical signal of the sample to be tested; performing quality assessment on the original physical signal to obtain a signal quality quantification index; if the signal quality quantification index is greater than or equal to a preset signal quality threshold, then converting the original physical signal into a standard format spectrum; if the signal quality quantification index is less than the preset signal quality threshold, then optimizing the signal acquisition parameters and re-acquiring the signal.

[0006] Optionally, the step of collecting the physical signals and socio-contextual information of the sample to be tested, and preprocessing and structuring the physical signals and socio-contextual information, further includes: acquiring at least one original socio-contextual information of the sample to be tested; evaluating the integrity of the original socio-contextual information to obtain a quantitative index of information integrity; if the quantitative index of information integrity is greater than or equal to a preset integrity threshold, converting the original socio-contextual information into a structured social evidence feature vector; if the quantitative index of information integrity is less than the preset integrity threshold, executing guidance for contextual information collection or generating instructions for supplementary contextual information.

[0007] Optionally, the extraction of physical evidence features and social evidence features of the sample to be tested includes: generating a first numerical feature vector as physical evidence features based on the standard format spectrum, wherein the elements in the first numerical feature vector are used to characterize the spectrum features; and generating a second numerical feature vector as social evidence features based on the social context information, wherein the elements in the second numerical feature vector are used to characterize contextual attributes.

[0008] Optionally, constructing a Bayesian inference network comprising candidate substance hypothesis nodes, physical evidence nodes, and social evidence nodes includes: based on the physical evidence features, searching in a preset standard substance spectral library using a spectral matching algorithm, calculating the matching value between the physical evidence features and the standard spectra of candidate substances in the standard substance spectral library, sorting the candidate substances based on the matching value, and selecting a first candidate substance list with a score higher than a first preset threshold based on the sorting result; based on the social evidence features, calculating the correlation degree between the social evidence features and entity nodes in the chemical knowledge graph by traversing a preset chemical knowledge graph; sorting the candidate substances based on the correlation degree, and selecting a second candidate substance list with a score higher than a second preset threshold based on the sorting result; merging and deduplicating the first candidate substance list and the second candidate substance list to obtain a candidate substance set.

[0009] Optionally, the construction of the Bayesian inference network including candidate substance hypothesis nodes, physical evidence nodes, and social evidence nodes further includes: instantiating each candidate substance in the candidate substance set as an independent candidate substance hypothesis node in the Bayesian inference network; instantiating each specific spectral feature in the physical evidence features as an independent physical evidence node in the Bayesian inference network; instantiating each specific contextual attribute in the social evidence features as an independent social evidence node in the Bayesian inference network; and establishing directed edges from the candidate substance hypothesis nodes to each physical evidence node and each social evidence node, respectively, to characterize the probabilistic dependencies between the candidate substance hypothesis nodes and the physical evidence nodes, and between the candidate substance hypothesis nodes and the social evidence nodes.

[0010] Optionally, the step of solving for the posterior probability value corresponding to each candidate substance hypothesis includes: using an approximate inference algorithm, with the observation values ​​of all the physical evidence nodes and all the social evidence nodes as conditions, calculating the posterior probability value of each candidate substance hypothesis node; sorting all candidate substance hypotheses according to the posterior probability values ​​to obtain a candidate substance probability ranking list arranged in descending order of the posterior probability values.

[0011] Optionally, the generation of evidence chain analysis results includes: selecting one or more candidate substance hypotheses with the highest posterior probability values, quantitatively calculating the contribution of each physical evidence node and each social evidence node to the posterior probability value through attribution analysis; identifying nodes with positive contribution values ​​as supporting evidence and nodes with negative contribution values ​​as conflicting evidence; associating and encapsulating the supporting evidence, the conflicting evidence, the quantitative contribution weight corresponding to the supporting evidence, the quantitative adjustment magnitude corresponding to the conflicting evidence, and the selected candidate substance hypotheses to obtain structured evidence chain analysis results.

[0012] Optionally, the step of executing a decision based on the posterior probability value and the evidence chain analysis result, and generating and outputting a chemical substance identification report, includes: comparing the highest posterior probability value among the candidate substance hypotheses with a preset first confidence threshold and a second confidence threshold; if the highest posterior probability value is greater than or equal to the first confidence threshold, then executing a high-confidence identification decision, generating a chemical substance identification report including the identification result of the most likely substance and a complete evidence chain; if the highest posterior probability value is less than the first confidence threshold but greater than or equal to the second confidence threshold, then executing a medium-confidence assessment decision, generating a report including a ranked list of candidate substances and further testing suggestions; if the highest posterior probability value is less than the second confidence threshold, then executing a low-confidence request decision, generating a report including recommendations for subsequent testing schemes.

[0013] On the other hand, the present invention provides an artificial intelligence-based rapid chemical substance identification system for implementing an artificial intelligence-based rapid chemical substance identification method. The system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the artificial intelligence-based rapid chemical substance identification method.

[0014] The above technical solution significantly improves the accuracy and reliability of chemical substance identification in complex scenarios by integrating physical signals and social context information for dual-driven hypothesis generation and using Bayesian networks for probabilistic reasoning. It not only overcomes the shortcomings of traditional methods that rely on a single signal, but also makes the decision-making process transparent and credible by generating interpretable evidence chain analysis results.

[0015] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a rapid chemical substance identification method based on artificial intelligence.

[0017] Figure 2 This is a flowchart of Bayesian inference network modeling based on dual-source information fusion.

[0018] Figure 3 It is a three-level decision-making flowchart based on posterior probability and confidence threshold.

[0019] Figure 4 This is a schematic diagram of the association analysis of a dual-source evidence chain. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 -Appendix Figure 4 The specific implementation methods of the embodiments of the present invention will be described in detail below. It should be understood that the specific implementation methods described herein are only for illustrating and explaining the embodiments of the present invention, and are not intended to limit the embodiments of the present invention.

[0021] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0022] In the process of realizing this invention, the inventors of this application discovered that the prior art relies solely on isolated spectral matching based on a single instrument detection signal, resulting in poor reliability in identifying complex real samples and isomers, and the decision-making process lacks interpretability.

[0023] Example 1 Reference Figures 1-4 This is the first embodiment of the present invention, which provides a method for rapid identification of chemical substances based on artificial intelligence, including: S100: Collects physical signals and social context information of the sample to be tested, performs preprocessing and structured transformation on the physical signals and social context information, and extracts physical evidence features and social evidence features of the sample to be tested.

[0024] In the embodiments of this application, at least one original physical signal of the sample to be tested is acquired; the original physical signal is evaluated to obtain a signal quality quantification index; if the signal quality quantification index is greater than or equal to a preset signal quality threshold, the original physical signal is converted into a standard format spectrum; if the signal quality quantification index is less than the preset signal quality threshold, the signal acquisition parameters are optimized and reacquired.

[0025] In a preferred embodiment of this application, a corresponding physical signal acquisition device is first selected based on the physical state (solid, liquid, gas) of the sample to be tested. The physical signal can be one or more of Raman spectroscopy, infrared spectroscopy, mass spectrometry, chromatographic signals (high-performance liquid chromatography, gas chromatography), and ultraviolet-visible absorption spectroscopy. Then, acquisition parameters are preset, which can be one or more of spectral resolution, scan range, integration time, and column temperature. The initial parameter values ​​are determined based on the estimated substance type of the sample to be tested and the corresponding industry standards. At least one raw physical signal is acquired through the acquisition device, and environmental parameters such as temperature, humidity, and air pressure are recorded simultaneously.

[0026] Secondly, the quality of the acquired raw physical signals is assessed, three core quantification indicators are calculated, and the signal quality quantification index (out of 100 points) is obtained through weighted summation. The preferred weight allocation for each indicator is as follows: Signal-to-noise ratio (SNR), with a weight of 0.4, is calculated as the ratio of the mean effective peak value of the signal to the standard deviation of the background noise. It is used to characterize the degree of separation between signal and noise. Baseline stability, with a weight of 0.3, is characterized by the ratio of baseline drift to the signal peak range, where a baseline drift of ≤5% is considered acceptable. Signal integrity, with a weight of 0.3, is characterized by the percentage of valid signal points (no missing or saturated signals), where a percentage of valid signal points ≥ 95% is considered acceptable.

[0027] It should be noted that the weights of the above three core quantitative indicators were cross-validated using a dataset of 500 standard substances. The optimal signal quality assessment accuracy was achieved when the grid search yielded a signal-to-noise ratio weight of 0.4, a baseline stability weight of 0.3, and a signal integrity weight of 0.3.

[0028] A preset signal quality threshold is established (preferably 70 points, which is statistically calibrated based on a large amount of standard sample data and can be dynamically adjusted according to the signal type). The signal quality quantification index is compared with the signal quality threshold. If the signal quality quantification index is greater than or equal to the preset signal quality threshold, a standardization transformation is performed. If the signal quality quantification index is less than the preset signal quality threshold, signal acquisition parameter optimization is performed.

[0029] It should be noted that signal acquisition parameter optimization and re-detection are based on the quantitative indicators that fail to meet the quality assessment, and the acquisition parameters are adjusted accordingly. For example, if the signal-to-noise ratio is insufficient, the integration time is extended or the light source intensity is increased; if the baseline drift exceeds the limit, the equipment baseline is corrected and the acquisition environment is stabilized; if the signal integrity is insufficient, the scanning range is supplemented or the equipment sensitivity is adjusted. Acquisition is then re-executed using the optimized parameters. If the quality score is still below the threshold after three consecutive optimizations, equipment fault diagnosis instructions and sample preprocessing suggestions are generated.

[0030] Next, the qualified raw physical signals are standardized and converted into data formats that conform to industry standards, such as converting spectral signals into JDX format and chromatographic signals into CDF format, and unifying the coordinate units.

[0031] Preferably, based on previously recorded environmental parameters, a baseline correction algorithm (such as polynomial fitting) and a background subtraction algorithm are used to eliminate environmental interference and inherent equipment noise, resulting in a clean standard format spectrum.

[0032] In the embodiments of this application, at least one original social context information of the sample to be tested is obtained; the integrity of the original social context information is evaluated to obtain a quantitative index of information integrity; if the quantitative index of information integrity is greater than or equal to a preset integrity threshold, the original social context information is converted into a structured social evidence feature vector; if the quantitative index of information integrity is less than the preset integrity threshold, the guidance for context information collection or the instruction to generate supplementary context information is executed.

[0033] In a preferred embodiment of this application, the first step is to collect original social context information. Four core information collection dimensions are defined around the sample to be tested: source information (location, time, personnel, batch affiliation, etc.), application scenario information (estimated use, industry, usage environment, etc.), related document information (previous reports, process parameters, industry standards, etc.), and auxiliary descriptive information (appearance characteristics, accompanying substances, etc.). A combination of automatic data collection and manual supplementation is employed. For example, with authorized permissions, data such as the location, time, submitting personnel, and batch affiliation of the sample are automatically collected from the laboratory management system. Data is also collected through standardized electronic questionnaires on mobile terminals (such as tablets and mobile phones), guiding users to input or select estimated use, industry, and usage environment. Furthermore, the system supports uploading photos of the sample's appearance via the mobile device's camera, using built-in image recognition algorithms to automatically extract key appearance features such as color, physical state, and morphology to assist in generating text descriptions. All information collected through interactive methods is based on explicit user input or confirmation, thereby ensuring the legality and compliance of the data source.

[0034] Secondly, the completeness of the collected original social context information is assessed. The information is broken down into specific information points according to dimensions, and a quantitative index of information completeness (out of 100 points) is calculated based on the ratio of the number of valid information points to the total number of information points. Source information and application scenario information are defined as key information points, with 20 points deducted for each missing item; auxiliary descriptive information is considered secondary information points, with 5 points deducted for each missing item. A preset completeness threshold is set (preferably 80 points, calibrated based on knowledge graph association and matching requirements). If the quantitative index of information completeness is greater than or equal to this threshold, the structured transformation step is initiated; otherwise, an information supplementation guidance process is triggered.

[0035] In the process of guiding the supplementation of social contextual information, precise supplementation instructions are generated for missing information points, such as prompting the user to select their industry or guiding them to upload relevant documents. The semantic consistency of the supplemented information also needs to be verified. If key information points still cannot be supplemented, a missing information label is added, and the weight of the corresponding feature is reduced in subsequent feature inference.

[0036] Next, the unstructured raw social context information is transformed into a structured form, including text cleaning (deduplication, word segmentation, and stop word removal) and information encoding. The encoding methods are as follows: for categorical variables such as industry and purpose, one-hot encoding is used to convert them into binary vectors; for numerical variables such as time and batch, standardization is performed; for textual variables such as appearance descriptions, word embedding algorithms are used to convert them into low-dimensional numerical vectors.

[0037] In the embodiments of this application, a first numerical feature vector is generated as a physical evidence feature based on a standard format spectrogram, and the elements in the first numerical feature vector are used to characterize the spectrogram features; a second numerical feature vector is generated as a social evidence feature based on social context information, and the elements in the second numerical feature vector are used to characterize contextual attributes.

[0038] In a preferred embodiment of this application, two types of features are extracted from the standardized spectrum: first, feature point information, which includes the position, intensity, peak width, and area of ​​characteristic peaks for spectral signals, and retention time, peak height, area, and peak shape parameters for chromatographic signals; second, global features, such as the overall trend of the spectrum, the number and distribution density of characteristic peaks, and the intensity ratio of key peaks. Based on these features, a first numerical feature vector is generated. To eliminate the influence of dimensions, the elements of the first numerical feature vector are normalized (preferably using the Min-Max normalization algorithm) to obtain standardized physical evidence features. Core features strongly associated with chemical substance identification are selected from the structured data, and redundant information is removed. The various coded features are concatenated according to preset dimensions (source, application scenario, associated document, auxiliary description) to generate a second numerical feature vector with a fixed dimension (e.g., 256 dimensions). This vector is normalized to obtain the final social evidence features.

[0039] It should be noted that social evidence features play a dual role in this method. On the one hand, they are used to narrow down the hypothesis space of candidate substances, and on the other hand, they serve as observational evidence to participate in subsequent Bayesian inference. The two are independent of each other in terms of processing stage and functional purpose.

[0040] The above scheme collects the physical signals and social context information of the sample to be tested, and performs preprocessing and structure transformation processing respectively, so that the original heterogeneous data is uniformly mapped to a computable feature space. This effectively reduces the differences in dimensions, noise and expression forms of different data sources. It not only improves the availability and stability of physical detection signals and context information, but also avoids the noise amplification and uncertainty accumulation problems caused by directly performing joint modeling on the original data, thereby improving the robustness of the overall recognition process from the source.

[0041] S200: Based on the characteristics of physical evidence and social evidence, a Bayesian inference network containing candidate material hypothesis nodes, physical evidence nodes, and social evidence nodes is constructed through dual-driven hypothesis generation.

[0042] In the embodiments of this application, based on physical evidence features, a spectral matching algorithm is used to search a preset standard substance spectral library, and the matching value between the physical evidence features and the standard spectra of candidate substances in the standard substance spectral library is calculated. The matching value is used to characterize the similarity between the two features. Candidate substances are sorted based on the matching value. According to the sorting result, substances with scores higher than a first preset threshold are selected. [The first preset threshold is preferably 0.7. The determination of this value is based on statistical calibration of a large number of standard substance spectral matching experiments: on a training set containing 500 representative standard samples, the accuracy and recall of spectral matching are used as the core evaluation indicators to plot the receiver operating characteristic curve (ROC curve) and calculate the Youden index; the matching degree threshold that makes the Youden index reach its maximum value (i.e., the critical point that can most effectively distinguish between correct and incorrect matches) is determined as the optimal threshold, which is statistically calculated to be approximately 0.7.] This threshold can be dynamically fine-tuned for different spectral types (such as Raman spectroscopy and infrared spectroscopy) to adapt to their unique feature distribution and noise level. Based on social evidence features, the correlation between social evidence features and entity nodes in the chemical knowledge graph is calculated by traversing a preset chemical knowledge graph. The correlation degree is used to characterize the degree of fit between candidate substances and social context attributes. Candidate substances are ranked based on the correlation degree. According to the ranking results, those with scores higher than the second preset threshold are selected. [The second preset threshold is preferably 0.65. This value is determined based on the statistical calibration of entity correlation rules in the chemical knowledge graph: on a validation set covering 200 typical application scenarios, the impact of social context information on the accuracy and recall of candidate substance ranking under different correlation degree thresholds is analyzed. Through precision-recall curve (PR curve) analysis, a threshold point that can effectively recall key candidate substances while ensuring high accuracy is selected. Statistically, this value is approximately 0.65.] This threshold can be adaptively adjusted according to the importance differences of specific social context information types (such as industry information, usage information) to form a second candidate substance list; the first candidate substance list and the second candidate substance list are merged and deduplicated to obtain a candidate substance set.

[0043] In some implementations, the candidate chemical substance set also includes a class of unknown substances or other substance nodes to characterize potential chemical substances not covered by the existing chemical substance database, thereby ensuring the integrity and normalization of the posterior probability distribution during subsequent Bayesian inference. The unknown substance nodes are only used to absorb unmatched probability masses, and their corresponding posterior probabilities do not participate in the ranking and identification decisions of known chemical substances.

[0044] In a preferred embodiment of this application, the matching value is calculated using a cosine similarity algorithm, and the first numerical feature vector corresponding to the physical evidence feature is denoted as . The eigenvector corresponding to the standard spectrum of a candidate substance in the standard substance spectral library is: The matching value is calculated as follows:

[0045] Where M represents the matching value between the physical evidence features and the standard spectral features of the candidate substance, and the value ranges from [0,1]. The larger the value, the higher the similarity between the two features. The i-th element of the first numerical feature vector X represents the i-th core feature parameter (such as feature peak intensity, peak area, peak width, etc.) of the standard format spectrum of the sample to be tested. The i-th element of the standard spectrum feature vector Y represents the i-th core feature parameter of the standard spectrum of the candidate substance. and All values ​​are dimensionless standardized values ​​obtained through normalization; n represents the dimension of the feature vector, which is determined by the spectrum type and the number of core features (e.g., the vector dimension corresponding to Raman spectrum is set to 512 dimensions); i represents the dimension index.

[0046] In a preferred embodiment of this application, the relevance R is calculated using a weighted fusion algorithm, fusing semantic relevance and attribute relevance. Let the text vector processed by word embedding in the social evidence features be... In a chemical knowledge graph, the text description vector corresponding to a certain entity node is: The semantic relevance is calculated using the cosine similarity algorithm, as shown in the following formula:

[0047] in, This represents the semantic correlation between the text vector of social evidence features and the text description vector of entity nodes, with a value range of [0,1]. The larger the value, the higher the semantic fit between the two. Let i represent the i-th element of text vector A, and let i represent the i-th semantic feature of the social context text information of the sample to be tested. represents the i-th element of the text description vector B, which represents the i-th semantic feature of the substance corresponding to the entity node; m represents the dimension of the text vector, which is determined by the word embedding algorithm and the complexity of the text information; i represents the dimension index.

[0048] Let C be the set of classification attributes in social evidence features, and D be the set of attributes corresponding to a certain entity node in the chemical knowledge graph. The Jaccard coefficient algorithm is used to calculate the attribute association degree, and the calculation formula is as follows:

[0049] in, The value ranges from [0,1] to represent the correlation between the social evidence feature classification attribute and the entity node attribute. The larger the value, the higher the matching degree between the two attributes. C represents the set of classification attributes in the social evidence feature of the sample to be tested, which consists of structured classification information such as the industry to which the sample belongs, the collection location, and the usage environment. D represents the set of attributes corresponding to a certain entity node in the chemical knowledge graph, which consists of classification information such as the industry affiliation and applicable scenarios of the substance corresponding to the node.

[0050] The formula for calculating the total correlation degree R is as follows:

[0051] Where R represents the total correlation between social evidence features and a certain entity node in the chemical knowledge graph, and the value ranges from [0,1]. The larger the value, the higher the degree of fit between the candidate substance and the social context attribute. Indicates semantic relevance The weighting coefficient, with a value of 0.6, is used to experimentally calibrate the degree of influence of social context text semantics on substance recognition. Indicates the degree of attribute association The weight coefficient is set to 0.4, and the influence of social context classification attributes on material recognition is experimentally calibrated.

[0052] Furthermore, the first and second candidate substance lists are merged, and a set deduplication algorithm is used to remove duplicate candidate substances, resulting in a candidate substance set. This candidate substance set covers substances that meet both physical feature matching and socio-contextual relevance requirements, balancing feature matching accuracy and scene adaptability.

[0053] In the embodiments of this application, each candidate substance in the candidate substance set is instantiated as an independent candidate substance hypothesis node in a Bayesian inference network; each specific spectral feature in the physical evidence features is instantiated as an independent physical evidence node in a Bayesian inference network; each specific contextual attribute in the social evidence features is instantiated as an independent social evidence node in a Bayesian inference network; directed edges are established from the candidate substance hypothesis nodes to each physical evidence node and each social evidence node to represent the probabilistic dependencies between the candidate substance hypothesis nodes and the physical evidence nodes, and between the candidate substance hypothesis nodes and the social evidence nodes.

[0054] In a preferred embodiment of this application, each candidate substance in the candidate substance set is instantiated as an independent candidate substance hypothesis node in a Bayesian inference network. Each node uniquely corresponds to a candidate substance and has a unidirectional mapping, specifically used to accurately characterize the core hypothesis that the substance is the sample to be tested. Each specific spectral feature in the physical evidence features is instantiated as an independent physical evidence node in the Bayesian inference network. Each node corresponds to a specific spectral feature parameter (such as characteristic peak position, peak intensity ratio, retention time, etc.) used to quantitatively characterize the inherent physical properties of the sample to be tested. Each specific contextual attribute in the social evidence features is instantiated as an independent social evidence node in the Bayesian inference network. Each node corresponds to a specific contextual attribute (such as sample purpose, industry, environmental conditions, etc.) used to quantitatively characterize the scene-related attributes of the sample to be tested.

[0055] Furthermore, based on causal logic, probabilistic dependencies between network nodes are constructed. Directed edges are established from each candidate substance hypothesis node to all physical evidence nodes and all social evidence nodes, clarifying the unidirectional dependency between candidate substance hypotheses and these two types of evidence nodes. The inherent physical properties of candidate substances determine their corresponding spectral characteristics, while their application scenario attributes determine their corresponding social context characteristics. The direction of the directed edges strictly conforms to objective technical laws, avoiding logical inversion. Simultaneously, each directed edge is assigned an initial conditional probability, specifically the conditional probability of the corresponding evidence node appearing when the candidate substance hypothesis is true. This initial conditional probability is based on a large amount of standard material data and implemented using a statistical learning algorithm. Taking physical evidence nodes as an example, the conditional probability P(B|A) between "candidate substance hypothesis node A" and "physical evidence node B" is calibrated by selecting the NIST 2020 standard mass spectrometry library as the training set and using an algorithm based on frequency statistics and Laplace smoothing. The formula for calculating the conditional probability P(B|A) is as follows:

[0056] Where A represents the candidate substance category variable, a is the specific value of the variable, B represents the numerical representation of the observational evidence, and b represents the specific observed value of the evidence variable. This represents the number of data records in the training set that simultaneously satisfy both conditions. |B| represents the total number of samples of substance A; |B| represents the number of all possible values ​​of evidence node B (e.g., if B is "whether a characteristic peak exists", then |B|=2). This is a smoothing factor, set to 1 in this formula, to prevent the probability of a non-occurring situation from being zero.

[0057] Furthermore, the calculated probability value P(B|A) is assigned to the directed edge from node A to node B, completing the initialization of the conditional probability table. The conditional probability labeling of social evidence nodes adopts a similar principle, but its data source is a pre-defined chemical knowledge graph. The association probability is obtained by calculating the correlation between entities and contextual attributes and performing normalization processing.

[0058] In some implementations, in response to the potential correlation between physical evidence nodes or social evidence nodes, evidence features with strong correlations can be combined to construct composite evidence nodes, or intermediate feature nodes can be introduced to characterize the combined effect of multiple evidence features, so as to reduce the impact of evidence redundancy on the accuracy of Bayesian inference results.

[0059] Further preferred, to avoid dimensionality explosion of the conditional probability table (CPT), it is preferable to use a parametric model (e.g., log-linear model, factorized approximation, or probabilistic regression model based on training data) for the conditional probability, or to use approximate inference and sampling methods to estimate the high-dimensional conditional probability, thereby balancing accuracy and computability.

[0060] The above scheme constructs a candidate substance set containing known chemical substance nodes and unknown chemical substance nodes, and introduces physical feature evidence and social context evidence into the Bayesian inference framework, so that the identification process can uniformly model multiple potential chemical substance hypotheses while ensuring the integrity and normalization of probability distribution.

[0061] S300: Based on Bayesian inference networks, it performs abductive calculations and probabilistic inferences, solves the posterior probability values ​​corresponding to each candidate substance hypothesis, and generates evidence chain analysis results.

[0062] In the embodiments of this application, an approximate reasoning algorithm is used to calculate the posterior probability value of each candidate substance hypothesis node using the observation values ​​of all physical evidence nodes and all social evidence nodes as conditions; all candidate substance hypotheses are sorted according to the posterior probability values ​​to obtain a candidate substance probability ranking list arranged in descending order of posterior probability values.

[0063] In a preferred embodiment of this application, the constructed Bayesian inference network is first initialized by loading a pre-set initial conditional probability table (this table is based on the correlation probability calibration between standard substance characteristic statistical data and chemical knowledge graph, and is stored in the directed edge attributes of each candidate substance hypothesis node pointing to physical and social evidence nodes), clarifying the probability dependencies and initial likelihood probabilities between each node. Subsequently, all physical evidence node observations (such as measured values ​​of characteristic peak positions, quantified values ​​of peak intensity ratios, etc.) and social evidence node observations (such as the industry label of the sample, usage environment parameters, etc.) are input as hard evidence into the Bayesian inference network to trigger network abductive reasoning calculation.

[0064] An approximate inference algorithm adapted to high-dimensional node networks is employed (Markov chain Monte Carlo algorithm is preferred, suitable for complex networks with a large number of nodes; if the network node dimension is low, variational inference algorithm can be used to improve computational efficiency), and the posterior probability values ​​of each candidate substance hypothesis are calculated based on Bayes' theorem. Solve for the posterior probability value, which represents the confidence level that the corresponding candidate substance is the sample to be tested under the combined support of the current physical and social evidence combination event B. The value range is [0,1]. The larger the value, the higher the confidence level, and the sum of the posterior probabilities corresponding to all candidate substance hypotheses is 1.

[0065] It should be noted that although unknown substance nodes are included in the posterior probability calculation to ensure probability normalization, when used to generate the candidate substance ranking list and identification decision, priority is given to ranking and comparing only the candidate substances listed in the known database. When the posterior probability of an unknown substance node exceeds a preset judgment threshold, it is used to indicate that the confidence of the identification result is insufficient and trigger a supplementary detection process, or to label the corresponding identification result as an unknown substance.

[0066] Bayes' theorem is as follows:

[0067] The specific calculation method is as follows: Call the prior probability of event A corresponding to the candidate substance hypothesis in the Bayesian inference network. This prior probability is the initial probability that the corresponding candidate substance is the sample to be tested when there is no supporting evidence. It is preset in the hypothesis node attributes, based on the frequency statistics of similar substances in the standard database, and takes a value range of [0,1]. All candidate substances correspond to... The sum is 1.

[0068] Call the likelihood probability from the conditional probability table The likelihood probability is the probability that event B, corresponding to the combination of physical and social evidence, occurs when event A, corresponding to the hypothesis of the candidate substance, is true. It is used to achieve precise matching and calling through a conditional probability table associated with directed edges. The value range is [0,1], and its value is pre-calibrated based on the characteristic statistical data of the standard substance and the association probability of the chemical knowledge graph.

[0069] Calculate the marginal probability of event B corresponding to the combination of evidence. This marginal probability represents the event corresponding to all candidate substance hypotheses. The total probability of event B occurring corresponding to the current combination of input evidence is given below, with a value range of [0,1]. Let i be the hypothesis event corresponding to the i-th candidate substance, where the subscript i is the unique identifier of the candidate substance, used to distinguish the hypotheses of different candidate substances. The calculation formula is as follows:

[0070] Where k represents the total number of candidate substance hypothesis nodes, that is, the total number of candidate substances participating in the reasoning, and is a positive integer; Let represent the likelihood probability of event B occurring when the hypothesis about the i-th candidate substance is true. Let represent the prior probability of the hypothesis for the i-th candidate substance.

[0071] The marginal probability of evidence combination event B is obtained by summing the likelihood probability × prior probability of all candidate substances.

[0072] Furthermore, the posterior probability values ​​corresponding to each candidate substance hypothesis are collected. Sort the candidate substances in descending order of their numerical values ​​to generate a probability ranking list. This list must be associated with and store core information such as the unique identifier of each candidate substance, its corresponding posterior probability value, probability ranking, and prior probability. This includes labeling evidence groups for suitability. After ranking, the confidence levels of the top three candidate substances are verified by calculating the posterior probability difference between adjacent ranked candidate substances. If the difference is less than or equal to a preset confidence threshold (preferably 0.15, based on statistical calibration from a large number of substance identification experiments), it is marked as a "high-confidence competing hypothesis" and needs to be compared in subsequent evidence contribution analysis. If the difference is greater than the threshold, the hypothesis ranked first is taken as the core candidate substance, ensuring that the ranking results provide a clear priority basis for subsequent analysis.

[0073] In the embodiments of this application, one or more candidate substance hypotheses with the highest posterior probability values ​​are selected, and the contribution of each physical evidence node and each social evidence node to the posterior probability value is quantitatively calculated through attribution analysis; nodes with positive contribution values ​​are identified as supporting evidence, and nodes with negative contribution values ​​are identified as conflicting evidence; supporting evidence, conflicting evidence, the quantitative contribution weight corresponding to supporting evidence, the quantitative adjustment magnitude corresponding to conflicting evidence, and the selected candidate substance hypotheses are associated and encapsulated to obtain structured evidence chain analysis results.

[0074] In a preferred embodiment of this application, if a "high-confidence competition hypothesis" exists, the top three candidate substances are selected as the joint analysis objects; if there is no competition, the posterior probability is selected. The highest-ranking candidate substance was selected as the core analysis object. Based on the node tracing function of the Bayesian network, the contribution of each physical evidence node and social evidence node to the posterior probability of the core candidate substance was quantitatively calculated.

[0075] In this embodiment, the contribution of evidence can be characterized by comparing the change in the posterior probability of the candidate substance hypothesis before and after the introduction of a certain evidence node, so as to reflect the degree of influence of the evidence node on the identification result. In other embodiments, the relative contribution of each physical evidence node and social evidence node can also be estimated based on the method of multi-evidence combination analysis, so as to adapt to the analysis needs under different application scenarios.

[0076] It should be noted that when the difference between the posterior probability value and the benchmark value is positive, it indicates that the evidence node has a positive supporting effect on the validity of the event A corresponding to the hypothesis; when it is negative, it indicates that the evidence node has a negative restrictive effect on the validity of the hypothesis. The larger the absolute value of the contribution, the stronger the influence of the corresponding evidence.

[0077] Furthermore, supporting evidence is identified as a screening criterion. For evidence nodes with positive values, ranked by contribution. The absolute values ​​are sorted from high to low, and the quantified contribution weights corresponding to each piece of evidence are recorded simultaneously. Based on these quantified contribution weights, the evidence is categorized into strength levels: ≥0.2 indicates strong supporting evidence, [0.1, 0.2) indicates moderately supporting evidence, and <0.1 indicates weakly supporting evidence. The quantified contribution weights characterize the relative influence of each evidence node on the formation of the posterior probability of the candidate substance hypothesis and do not participate in the normalization calculation of the posterior probability. The formula for calculating the quantified contribution weights is as follows:

[0078] in, represents the quantitative contribution weight of the i-th supporting evidence, used to measure the proportion of positive support for the core candidate substance hypothesis among all supporting evidence, with a value range of [0,1]; m represents the total number of supporting evidence, that is, the total number of positive supporting evidence participating in the weight calculation, which is a positive integer; s represents the sequence number variable for summation, used to traverse all supporting evidence. This represents the contribution of the s-th supporting piece of evidence.

[0079] Conflicting evidence identification as a screening method For evidence nodes with negative values, categorized by contribution. Sort the absolute values ​​from highest to lowest, and record the quantitative adjustment range corresponding to each piece of evidence. The evidence was classified into several levels of intensity, with an adjustment range of ≥0.1 indicating strong conflict, [0.05, 0.1) indicating medium conflict, and <0.05 indicating weak conflict.

[0080] Preferably, for each piece of classified evidence, the evidence is traced back to the corresponding candidate substance hypothesis node through the directed edge of the Bayesian inference network, and the evidence type, specific feature parameters and associated network node ID are labeled to ensure that the evidence is traceable.

[0081] Furthermore, based on the selected candidate substance hypotheses, key data corresponding to these hypotheses are extracted simultaneously, including name, unique identifier, and posterior probability of full evidence. Ranking, prior probability The confidence verification results serve as the foundational supporting evidence and quantitative data for the association encapsulation. Supporting evidence is retrieved, sorted by strength level from high to low, and each piece of evidence is bound to its corresponding quantitative contribution weight, contribution degree, network node ID, and likelihood probability, then linked one by one to the core candidate substance hypothesis. Similarly, conflicting evidence is retrieved, sorted by strength level from high to low, and each piece of evidence is bound to its corresponding quantitative adjustment range, network node ID, and likelihood probability, simultaneously linked to the core candidate substance hypothesis to clarify their constraints. The reasoning is supplemented by the node dependencies of the Bayesian inference network (i.e., the unidirectional causal probability association between the candidate substance hypothesis node and each evidence node), with each piece of evidence explained separately. How do the observed values ​​of supporting evidence match the attribute characteristics corresponding to the candidate substance hypothesis? Explain the supporting logic for the hypothesis using the likelihood probability. Taking a candidate substance hypothesis of a certain type of organic compound as an example, the observed values ​​of its corresponding infrared spectral detection peak show a characteristic absorption peak near wavenumber 1600, which matches the attribute characteristic of this type of organic compound containing a benzene ring. Combined with a likelihood probability of 0.92, this demonstrates the strong supporting logic for the hypothesis.

[0082] What are the differences between the observed values ​​of conflicting evidence and the hypothetical attributes? Explain the reasons for the constraints on the hypothesis based on the corresponding likelihood probability. For example, if the hypothetical melting point of a candidate substance is 80℃, but the observed melting point evidence is 65℃, there is a significant difference between the two. Based on the corresponding likelihood probability of 0.15, explain why this evidence constrains the hypothesis: the melting point deviation exceeds a reasonable range, reducing the probability of the hypothesis being true.

[0083] It should be noted that the above-mentioned evidence contribution analysis and evidence chain construction steps are not used to change the posterior probability calculation results of the candidate substance hypothesis, but rather to enhance the interpretability of the reasoning results and express the traceability of the evidence while keeping the Bayesian reasoning results unchanged.

[0084] Based on the above analysis of evidence contribution and evidence chain, the chemical substance identification process possesses the technical characteristics of being interpretable, traceable, and verifiable, making it suitable for application scenarios with high requirements for identification reliability.

[0085] The aforementioned scheme calculates the posterior probability of each candidate substance hypothesis and further constructs evidence contribution analysis and structured expression of the evidence chain based on this. This enables the chemical substance identification results to not only provide probabilistic judgments but also clarify the relative influence of different evidence nodes in the reasoning process. Without altering the Bayesian inference results, an interpretability enhancement mechanism is introduced, giving the identification process traceable and verifiable technical characteristics. This improves the acceptability and credibility of the identification results in regulatory, emergency response, and high-reliability application scenarios.

[0086] S400: Executes decisions based on posterior probability values ​​and evidence chain analysis results, generates and outputs a chemical substance identification report, and completes the identification of chemical substances in the sample to be tested.

[0087] In the embodiments of this application, the highest posterior probability value among the candidate substance hypotheses is compared with a preset first confidence threshold and a second confidence threshold. If the highest posterior probability value is greater than or equal to the first confidence threshold, a high-confidence identification decision is executed, generating a chemical substance identification report including the identification result of the most likely substance and a complete chain of evidence. If the highest posterior probability value is less than the first confidence threshold but greater than or equal to the second confidence threshold, a medium-confidence assessment decision is executed, generating a report including a ranked list of candidate substances and further testing suggestions. If the highest posterior probability value is less than the second confidence threshold, a low-confidence request decision is executed, generating a report including recommendations for subsequent testing schemes.

[0088] In a preferred embodiment of this application, the highest posterior probability value is first extracted from the candidate substance hypothesis set. At the same time, a preset dual confidence threshold is retrieved (e.g., the first confidence threshold is set to 0.8, which is set based on experimental data statistical optimization. Around this threshold, the system can achieve the highest automated processing efficiency while ensuring high accuracy; the second confidence threshold is set to 0.5, which is the theoretical benchmark point for random guessing in binary classification problems. When the highest posterior probability value among the candidate substance hypotheses is lower than this threshold, it indicates that the confidence level of all current hypotheses does not exceed the random probability, and their identification results are not statistically significant. Moreover, the first confidence threshold is greater than the second confidence threshold). The highest posterior probability value is then compared with the two thresholds in turn.

[0089] If the highest posterior probability value is greater than or equal to the first confidence threshold (e.g., ≥0.8), then a high-confidence identification decision is executed. The generated chemical substance identification report must clearly include the identification results such as the name of the most likely substance, purity range, and structural characteristics, as well as a complete encapsulated chain of evidence (supporting evidence and conflicting evidence, quantitative data and reasoning explanations) for subsequent review and archiving.

[0090] If the highest posterior probability value is less than the first confidence threshold and greater than or equal to the second confidence threshold (e.g., 0.5 ≤ highest posterior probability value < 0.8), then a medium confidence assessment decision is made. The report should list the candidate substances ranked from highest to lowest posterior probability (including at least the first 3, with corresponding probability values ​​marked), and specify further testing recommendations, such as supplementary infrared spectroscopy re-examination, accurate melting point determination, and other targeted solutions.

[0091] If the highest posterior probability value is less than the second confidence threshold (e.g., less than 0.5), then a low confidence request decision is executed. The generated report should prioritize recommending subsequent detection schemes, clearly suggesting new detection items (e.g., gas chromatography-mass spectrometry, elemental analysis), detection parameters, and priorities, to provide a basis for reconstructing candidate substance hypotheses and improving identification credibility.

[0092] The above scheme generates the final identification decision based on the posterior probability of candidate substances and a pre-set confidence threshold, effectively transforming the identification result from probabilistic reasoning into an engineering-executable decision. By distinguishing between unknown substance nodes and known substance ranking decisions, the system can output clear chemical substance identification results under high confidence conditions, and can promptly trigger supplementary detection or labeling of unknown results under low confidence conditions or when unknown substances dominate. This forms a complete, closed-loop identification decision mechanism, improving the system's security and reliability in practical applications.

[0093] Example 2 Reference Figures 1-4 This is the second embodiment of the present invention, which scientifically verifies the rapid identification method for chemical substances based on artificial intelligence in Embodiment 1.

[0094] Specifically, using an unknown transparent liquid (suspected organic solvent) found in a pharmaceutical factory's wastewater pond as the testing target, multi-source information of the sample was first collected and preprocessed. The Raman spectral data of the sample underwent quality assessment, with a signal quality quantification index of 85 points (signal-to-noise ratio 35, baseline drift 3%, signal integrity 98%). After passing the assessment, it was converted to standard JDX format spectra, and 1600... 750 The intensity and width of characteristic peaks form the physical evidence features. Simultaneously collected social context information includes: "Source: Wastewater pool of a pharmaceutical factory," "Industry: Pharmaceutical industry, Usage environment: Normal temperature and pressure," and "Appearance: Transparent liquid." This information completeness score is 90 points, and after encoding, a 256-dimensional social evidence feature vector is generated.

[0095] Subsequently, based on physical evidence characteristics, matching was performed in a standard spectral library to generate a first candidate substance list: toluene (match value 0.82), xylene (match value 0.78), and benzene (match value 0.73). Simultaneously, based on social evidence characteristics, the correlation was calculated in a chemical knowledge graph to generate a second candidate substance list: toluene (correlation 0.75), ethanol (correlation 0.68), and acetone (correlation 0.66). After merging and deduplicating the two lists, the candidate substance set {toluene, xylene, benzene, ethanol, acetone} was obtained.

[0096] The five substances in the candidate set (toluene, xylene, benzene, ethanol, and acetone) are instantiated as five candidate substance hypothesis nodes, and the key features are instantiated as evidence nodes, including three physical evidence nodes ("feature peak located at 1600"). "The characteristic peak is located at 750". "Characteristic peak intensity ratio 1600 / 750" The algorithm inputs observations for all candidate substances and two social evidence nodes (e.g., “Industry = Pharmaceutical Industry”, “Use = Reaction Solvent”). After inputting the observations for all evidence nodes, the posterior probability of each candidate substance hypothesis is calculated using an approximate inference algorithm. The results are: toluene (0.85), xylene (0.12), benzene (0.02), ethanol (0.006), and acetone (0.004).

[0097] Finally, a decision is made based on the posterior probability, and an interpretable report is generated. Since the posterior probability of toluene (0.85) is higher than the first confidence threshold (0.8), the system automatically outputs a high-confidence identification result for toluene. The report also includes a chain of evidence analysis, indicating that "the characteristic peak is located at 1600..." The evidence presented was "(contribution +0.25)" and "related to the pharmaceutical industry scenario" (contribution +0.20), which were key supporting evidence; while "physical state and potential solid-state risk" was weakly conflicting evidence (contribution -0.02). These pieces of evidence, fused using a Bayesian network, ultimately resulted in a high posterior probability for the toluene hypothesis, constituting a reliable identification conclusion. Verification using gas chromatography-mass spectrometry (GC-MS) confirmed that the results were completely consistent with the conclusions of this method, fully demonstrating its reliability.

[0098] The data above shows that although ethanol and acetone have a low physical compatibility, they were selected due to their strong relevance to the pharmaceutical industry, demonstrating the value of social context in broadening the candidate pool.

[0099] Furthermore, candidate substances are instantiated as hypothesis nodes, and physical characteristics (such as characteristic peak positions and intensity ratios) and social attributes (such as industry and application) are instantiated as evidence nodes to establish probabilistic dependencies. After inputting all evidence observations, the posterior probability of each substance is calculated using an approximate inference algorithm. The results are: toluene (0.85), xylene (0.12), benzene (0.02), ethanol (0.006), and acetone (0.004).

[0100] This embodiment demonstrates that the AI-based rapid chemical substance identification method can achieve rapid, accurate, and interpretable chemical substance identification in complex scenarios by integrating physical signals and social context information.

[0101] Example 3 Reference Figures 1-4 This is the third embodiment of the present invention, which compares and verifies the rapid chemical substance identification method based on artificial intelligence in Embodiment 1 with existing conventional chemical substance identification methods.

[0102] To quantify the contribution of social context information to the method of this invention, this embodiment compares the comprehensive performance differences in complex scenarios between traditional recognition schemes that use only physical signals and the scheme of this invention that integrates physical signals and social context information.

[0103] The experiment selected 100 mixed-scenario samples as the test set. This set specifically included 50 pairs of isomers with highly similar physical characteristics and 30 challenging samples with low signal-to-noise ratio and low concentration to simulate the difficulties in real-world testing. All samples were randomly divided into two groups, A and B, and tested under identical hardware and software conditions.

[0104] Group A (control group): Only the physical signal of the sample (such as Raman spectrum) is input, and the traditional spectral matching and threshold judgment method is used for identification.

[0105] Group B (Experimental Group): The method of this invention is used, and the physical signal of the sample is input simultaneously with the social context information preset according to the sample background (such as "chemical production intermediate", "pharmaceutical wastewater", "electronic cleaning agent" etc.).

[0106] Both experimental groups used the same standard mass spectrum library. For group B, structured social context information matching the scenario was matched to each sample. Table 1 summarizes the experimental results.

[0107] Table 1 Summary of Comparative Experiment Results

[0108] As shown in Table 1, the significant improvement in the recognition accuracy of Group B, especially the nearly 30% improvement in the recognition rate of isomers, directly proves that social context information provides a key basis for distinguishing substances with similar physical characteristics.

[0109] The reduced decision-making time demonstrates that socio-contextual information effectively focuses the candidate pool and reduces the computational overhead of the system on irrelevant assumptions. Meanwhile, the significantly increased proportion of high-confidence results indicates that the proposed solution outputs more decisive results, reduces uncertainty, and alleviates the burden of subsequent manual review.

[0110] In summary, this comparative experiment fully verifies that by integrating social context information, the present invention has achieved substantial breakthroughs in recognition accuracy, decision-making efficiency, and result certainty, significantly improving the overall performance of chemical substance recognition in complex scenarios.

[0111] The present invention also provides an artificial intelligence-based rapid chemical substance identification system for implementing an artificial intelligence-based rapid chemical substance identification method. The system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the artificial intelligence-based rapid chemical substance identification method.

[0112] This invention provides a storage medium storing a program that, when executed by a processor, implements the artificial intelligence-based rapid chemical substance identification method.

[0113] This invention provides a processor for running a program, wherein the program executes the artificial intelligence-based rapid chemical substance identification method during runtime.

[0114] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements an artificial intelligence-based method for rapid identification of chemical substances. The device described herein can be a server, PC, tablet, mobile phone, etc.

[0115] This application also provides a computer program product that, when executed on a data processing device, is suitable for performing an artificial intelligence-based method for rapid identification of chemical substances.

[0116] Those skilled in the art will understand that embodiments of this application can provide methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0117] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0121] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0122] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0123] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0124] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for rapid identification of chemical substances based on artificial intelligence, characterized in that, include: The physical signals and social context information of the sample to be tested are collected, and the physical signals and social context information are preprocessed and structured transformation is performed to extract the physical evidence features and social evidence features of the sample to be tested. Based on the physical evidence features and the social evidence features, a Bayesian inference network containing candidate material hypothesis nodes, physical evidence nodes, and social evidence nodes is constructed through dual-driven hypothesis generation. Based on the Bayesian inference network, abductive calculation and probabilistic inference are performed to solve the posterior probability value corresponding to each candidate substance hypothesis and generate evidence chain analysis results. Decisions are made based on the posterior probability value and the chain of evidence analysis results, a chemical substance identification report is generated and output, and the identification of the chemical substance in the sample to be tested is completed.

2. The method for rapid identification of chemical substances based on artificial intelligence according to claim 1, characterized in that, The process of collecting physical signals and socio-contextual information from the sample to be tested, and preprocessing and structuring the physical signals and socio-contextual information, includes: Acquire at least one original physical signal of the sample to be tested; The original physical signal is subjected to quality assessment to obtain a signal quality quantification index; If the signal quality quantization index is greater than or equal to the preset signal quality threshold, the original physical signal is converted into a standard format spectrum. If the signal quality quantization index is less than the preset signal quality threshold, then the signal acquisition parameters are optimized and reacquired.

3. The method for rapid identification of chemical substances based on artificial intelligence according to claim 2, characterized in that, The process of collecting physical signals and socio-contextual information from the sample to be tested, and preprocessing and structuring the physical signals and socio-contextual information, further includes: Obtain the original social context information of the sample to be tested, wherein the social context information includes at least one of the following: sample source information, application scenario information, associated document information, and appearance description information; The integrity of the original social context information is assessed to obtain a quantitative index of information integrity. If the information integrity quantification index is greater than or equal to the preset integrity threshold, the original social context information is converted into a structured social evidence feature vector. If the information integrity quantification index is less than the preset integrity threshold, then the guidance for collecting contextual information or the instruction to generate supplementary contextual information will be executed.

4. The method for rapid identification of chemical substances based on artificial intelligence according to claim 3, characterized in that, The extraction of physical and social evidence characteristics from the sample to be tested includes: Based on the standard format spectrum, a first numerical feature vector is generated as a physical evidence feature, and the elements in the first numerical feature vector are used to characterize the spectrum features. Based on the social context information, a second numerical feature vector is generated as the social evidence feature, and the elements in the second numerical feature vector are used to characterize contextual attributes.

5. The method for rapid identification of chemical substances based on artificial intelligence according to claim 1, characterized in that, The construction of a Bayesian inference network comprising candidate substance hypothesis nodes, physical evidence nodes, and social evidence nodes includes: Based on the physical evidence features, a spectral matching algorithm is used to search in a preset standard substance spectral library, calculate the matching value between the physical evidence features and the standard spectra of candidate substances in the standard substance spectral library, sort the candidate substances based on the matching value, and filter out a first candidate substance list with a score higher than a first preset threshold based on the sorting result. Based on the social evidence features, the correlation between the social evidence features and entity nodes in the chemical knowledge graph is calculated by traversing a preset chemical knowledge graph; the candidate substances are sorted based on the correlation, and a second candidate substance list with a score higher than a second preset threshold is obtained based on the sorting result. The first candidate substance list and the second candidate substance list are merged and deduplicated to obtain a candidate substance set.

6. The method for rapid identification of chemical substances based on artificial intelligence according to claim 5, characterized in that, The construction of the Bayesian inference network, which includes candidate substance hypothesis nodes, physical evidence nodes, and social evidence nodes, also includes: Each candidate substance in the candidate substance set is instantiated as an independent candidate substance hypothesis node in the Bayesian inference network; Each specific spectral feature in the physical evidence features is instantiated as an independent physical evidence node in the Bayesian inference network; Each specific contextual attribute in the social evidence features is instantiated as an independent social evidence node in the Bayesian inference network. Directed edges are established from the candidate substance hypothesis nodes to each physical evidence node and each social evidence node, respectively, to represent the probabilistic dependencies between the candidate substance hypothesis nodes and the physical evidence nodes, and between the candidate substance hypothesis nodes and the social evidence nodes.

7. The method for rapid identification of chemical substances based on artificial intelligence according to claim 6, characterized in that, The posterior probability values ​​corresponding to each candidate substance hypothesis are as follows: Using an approximate inference algorithm, the posterior probability value of each candidate material hypothesis node is calculated, with the observation values ​​of all the physical evidence nodes and all the social evidence nodes as conditions. All candidate substance hypotheses are sorted according to the posterior probability values ​​to obtain a candidate substance probability ranking list arranged in descending order of the posterior probability values.

8. The method for rapid identification of chemical substances based on artificial intelligence according to claim 1, characterized in that, The results of the generated chain of evidence analysis include: Select one or more candidate substance hypotheses with the highest posterior probability values, and quantitatively calculate the contribution of each physical evidence node and each social evidence node to the posterior probability values ​​through attribution analysis. Nodes with positive contribution values ​​are identified as supporting evidence, and nodes with negative contribution values ​​are identified as conflicting evidence. The supporting evidence, the conflicting evidence, the quantitative contribution weights corresponding to the supporting evidence, the quantitative adjustment ranges corresponding to the conflicting evidence, and the selected candidate substance hypotheses are correlated and encapsulated to obtain structured evidence chain analysis results.

9. The method for rapid identification of chemical substances based on artificial intelligence according to claim 1, characterized in that, The step of making a decision based on the posterior probability value and the chain of evidence analysis results, and generating and outputting a chemical substance identification report, includes: The highest posterior probability value in the candidate substance hypothesis is compared with the preset first confidence threshold and second confidence threshold; If the highest posterior probability value is greater than or equal to the first confidence threshold, a high-confidence identification decision is executed to generate a chemical substance identification report that includes the identification result of the most likely substance and a complete chain of evidence. If the highest posterior probability value is less than the first confidence threshold and greater than or equal to the second confidence threshold, then a medium confidence assessment decision is performed, generating a report that includes a ranked list of candidate substances and further testing recommendations; If the highest posterior probability value is less than the second confidence threshold, a low confidence request decision is executed, generating a report that includes recommendations for subsequent detection schemes.

10. A rapid chemical substance identification system based on artificial intelligence, characterized in that, The system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the artificial intelligence-based rapid chemical substance identification method according to any one of claims 1-9.

Images