Chemical safety production information management method and system based on large model

Abstract

This invention discloses a method and system for information management of chemical safety production based on a large model, relating to the field of information management technology for chemical safety production. The method includes the following steps: Converting and recognizing the original audio to obtain a list of clean speech segments from a single speaker, then decoding this list to obtain an N-best candidate text list and the original acoustic score; performing a mixed search on each candidate text in the N-best candidate text list to obtain a reference text set; calculating the reordering score for each candidate text; fusing the reordering score with the normalized original acoustic score to obtain the final corrected text; performing anomaly detection to obtain a preliminary audit tag set and a comprehensive anomaly report; performing compliance detection on the final corrected text to obtain a compliance audit result; manually reviewing the preliminary audit tag set to obtain the final audit tag set; and generating an audit report based on the final audit tag set, the compliance audit result, and the comprehensive anomaly report.

Description

Technical Field

[0001] This invention relates to the field of information management technology for chemical safety production, and in particular to a method and system for information management of chemical safety production based on a large model. Background Technology

[0002] In recent years, chemical safety production management has increasingly shifted towards informatization and intelligentization. Traditional methods relying on manual recording and inspection have prominent problems such as low efficiency, slow response, and difficulty in traceability, which can no longer meet the real-time monitoring needs of modern high-risk chemical environments. With the rapid development of voice recognition and big data modeling technologies, the use of natural audio data from walkie-talkie conversations, inspection voice reports, and shift handover recordings for automated analysis and safety auditing has become an important technical path to improve the level of chemical safety management. By building a voice-driven safety production information management system, it is possible to achieve in-process early warning and post-event traceability, effectively reducing the risk of human negligence. However, existing technologies still have significant shortcomings in terms of voice recognition accuracy, professional terminology understanding ability, and intelligent compliance judgment in complex acoustic environments.

[0003] Current information management technologies for chemical safety production generally employ general speech recognition models to transcribe on-site audio, and then perform simple anomaly detection based on rule or keyword matching. However, chemical sites often exhibit complex acoustic features such as multi-person dialogues, overlapping speech, and frequent use of colloquial terminology. General models struggle to effectively separate speakers and accurately identify technical terms, resulting in a high error rate. Furthermore, traditional methods rely solely on acoustic scores to select the optimal recognition result, lacking the integration and utilization of historical correct data and procedural knowledge bases, thus limiting error correction capabilities. In addition, existing technologies for auditing and analyzing recognized texts often focus on matching isolated violation words, failing to understand the risk decision-making logic within the dialogue context, and struggling to conduct cross-period comprehensive analysis of recurring abnormal states of the same equipment. This leads to the underreporting of numerous hidden violations, severely restricting the level of intelligence in chemical safety production management. Summary of the Invention

[0004] The technical problems solved by this invention are: low accuracy of speech recognition in complex acoustic environments of chemical industry, difficulty in recognizing professional terms, lack of automatic error correction mechanism based on knowledge fusion, and inability to perform in-depth comprehensive analysis of compliance and anomalies across time periods on the recognized text.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a chemical safety production information management method based on a large model, comprising the following steps: Step S1: After converting the original audio, perform recognition to obtain a list of clean speech segments of a single speaker; Step S2: Decode the clean speech segment of a single speaker to obtain the N-best candidate text list and the original acoustic score; Step S3: Perform a mixed retrieval on each candidate text in the N-best candidate text list to obtain a reference text set. Calculate the re-ranking score for each candidate text based on the reference text set. Fuse the re-ranking score with the normalized original acoustic score to obtain the final corrected text. Step S4: Based on the final error correction text, perform anomaly detection to obtain a preliminary audit mark set and a comprehensive anomaly report; Step S5: Perform compliance assessment on the final corrected text to obtain compliance audit results, and manually review the preliminary audit mark set to obtain the final audit mark set. Based on the final audit mark set, compliance audit results, and comprehensive anomaly report, generate an audit report.

[0006] As a preferred embodiment of the chemical safety production information management method based on a large model as described in this invention, step S1 specifically includes: Step S101: Collect raw audio from the chemical production process; The original audio is converted using audio resampling technology to obtain audio data; Audio data is converted using channel conversion technology to obtain standardized audio data; Step S102: Perform speech recognition on the standardized audio data to obtain a list of clean speech segments from a single speaker.

[0007] As a preferred embodiment of the chemical safety production information management method based on a large model as described in this invention, step S2 specifically includes: Build a chemical engineering terminology database; For a single-speaker clean speech segment, the Whisper model is used for decoding. During the decoding process, a beam search algorithm based on hot word enhancement is called to construct a hot word prefix tree from the chemical terminology database to obtain candidate text sequences. When a candidate text sequence matches any hot word path in the hot word prefix tree, a preset positive score is added to the candidate text sequence, and the N-best candidate text list and the original acoustic score corresponding to each candidate text sequence are output.

[0008] As a preferred embodiment of the chemical safety production information management method based on a large model as described in this invention, step S3 specifically includes: Step S301: Construct a knowledge base based on the pairing of historical speech recognition text with the corrected text and the explanation of technical terms; Step S302 involves performing a mixed retrieval on each candidate text in the N-best candidate text list, specifically including: For each candidate text in the N-best candidate text list, vector semantic retrieval and BM25 keyword retrieval are used respectively to obtain historical paired vector retrieval results and historical paired BM25 retrieval results; Take the union of historical pairing vector retrieval results and historical pairing BM25 retrieval results, sort them in descending order according to the final similarity, and obtain the sorted merged retrieval result set; The top K corrected texts in the sorted and merged search results set are selected as the reference text set. Where K is greater than or equal to 1; If the number of correct texts in the merged search results after sorting is less than K, placeholder texts are used to make up to K, resulting in a set of reference texts for each candidate text. Otherwise, directly take the first K corrected texts as the reference text set.

[0009] As a preferred embodiment of the chemical safety production information management method based on a large model described in this invention, step S3 further includes: Step S303: Process each candidate text and its corresponding set of reference texts according to the re-ranking score calculation rules to obtain the re-ranking score of each candidate text. The specific rules for calculating reordering scores include: The Cross-Encoder model is used to calculate the relevance score of each corrected text that is not a placeholder in the candidate text and the reference text set. The arithmetic mean of all relevance scores is taken as the re-ranking score of the candidate text. If the reference text set consists entirely of placeholder text, then the reordering score for the candidate text is set to 0. Step S304: The original acoustic score corresponding to each candidate text is processed by Softmax normalization to obtain the normalized original acoustic score of each candidate text. Step S305: The re-ranking scores of all candidate texts are processed using Softmax normalization to obtain the normalized re-ranking score of each candidate text. The weighted sum of the normalized original acoustic score and the normalized reordered score for each candidate text is used to obtain the final score. Select the candidate text with the highest final score as the final text to be corrected; The final corrected text is taken directly from the N-best candidate text list.

[0010] As a preferred embodiment of the chemical safety production information management method based on a large model described in this invention, the anomaly determination in step S4 specifically includes: Step S401: Based on punctuation marks and line breaks, the final corrected text is segmented into sentences to obtain multiple sentences. For each sentence, the Llama 2 language model is used to judge the violation, and the violation type, confidence level and violation clause number are output to obtain the sentence-level violation mark. The confidence levels include high, medium, and low. Step S402: The final corrected text is used as a dialogue fragment, and the Llama 2 large language model is used for judgment to obtain dialogue-level risk decision tags; Dialogue-level risk decision markers include violation type, confidence level, and violation clause number; The confidence levels include high, medium, and low. A preliminary audit tag set is constructed based on statement-level violation tags and dialogue-level risk decision tags.

[0011] As a preferred embodiment of the chemical safety production information management method based on a large model described in this invention, step S4 further includes: Step S403: Use regular expression matching to extract key entities and state descriptions from the dialogue segments, input the key entities and state descriptions into the Sentence-BERT model for encoding, obtain semantic vectors, and store them in the vector index library. According to a preset cycle, key entities are used as query conditions to retrieve historical dialogue fragments from the vector index library to obtain search results. The search results are input into the Llama 2 language model to determine the same state description corresponding to the same key entity in the search results. Specifically, this includes: If the same state description for the same key entity is mentioned more than once, a comprehensive anomaly report marked as content abnormality will be generated. Otherwise, generate a comprehensive anomaly report marked as having no abnormalities.

[0012] As a preferred embodiment of the chemical safety production information management method based on a large model as described in this invention, step S5, which involves determining the compliance of the final error correction text, specifically includes: Step S501: Perform semantic role labeling on the final corrected text to obtain candidate operation behavior phrases; Perform keyword matching on candidate action phrases and output action descriptions; A safety procedure knowledge base is constructed, and behavioral descriptions are used as query conditions. Vector semantic retrieval is then used to search the safety procedure knowledge base to obtain safety procedure vector retrieval results. The BM25 keyword was used to search the safety regulations knowledge base to obtain the safety regulations BM25 search results. Take the union of the safety procedure vector retrieval results and the safety procedure BM25 retrieval results, sort them in descending order according to the final comprehensive similarity, and obtain the output retrieval result set; Select the first Q corrected texts from the output search results set as the reference text set; Step S502: Input the behavior description and procedure clause text into the Llama 2 language model to make a compliance judgment, and output the compliance judgment result, severity level and judgment basis text to obtain the compliance audit result; Step S503: Mark the initial audit flags with low confidence levels as pending manual review to obtain the audit list; Entries with a confidence level of medium or high are not marked.

[0013] As a preferred embodiment of the chemical safety production information management method based on a large model described in this invention, step S5 further includes: Step S504: Receive manual review instructions input by security management personnel through the interactive interface for the markers in the pending review list; Manual review instructions include confirmation, correction, or rejection; The confirmation command means to keep the original mark unchanged, the correction command means to modify the original mark to the correction value entered by the user, and the rejection command means to delete the original mark. The tags are manipulated according to the manual review instructions to generate feedback records, resulting in the final audit tag set and feedback records. Step S505: Generate an audit report from the final audit tag set, compliance audit results, and comprehensive anomaly report. The audit report includes a statistical summary, violation cases, and rectification recommendations. Generate a rectification task list based on the audit report, and output the audit report and rectification task list.

[0014] A chemical safety production information management system based on a large model, including an identification module, a decoding module, a sorting module, a judgment module, and a generation module; The recognition module converts the original audio and then recognizes it to obtain a list of clean speech segments of a single speaker. The decoding module decodes the clean speech segments of a single speaker to obtain the N-best candidate text list and the original acoustic score; The sorting module performs a mixed retrieval on each candidate text in the N-best candidate text list to obtain a reference text set. Based on the reference text set, it calculates the re-sorting score for each candidate text and merges the re-sorting score with the normalized original acoustic score to obtain the final corrected text. The judgment module performs anomaly judgment based on the final error correction text, and obtains a preliminary audit mark set and a comprehensive anomaly report; The generation module performs compliance assessment on the final corrected text to obtain compliance audit results, and manually reviews the preliminary audit tag set to obtain the final audit tag set. Based on the final audit tag set, compliance audit results, and comprehensive anomaly report, an audit report is generated.

[0015] The beneficial effects of this invention are as follows: Existing algorithms select the optimal text solely based on the original acoustic score. When multiple candidate texts have similar acoustic scores, they cannot effectively distinguish between superior and inferior texts. Furthermore, they lack the utilization of chemical engineering terminology and historical correct corpora, resulting in a high error rate in terminology recognition and limited error correction capabilities. This invention adds positive scores to the professional term path through a hot word prefix tree in a chemical engineering terminology database, thereby improving the terminology recognition tendency. It further integrates acoustic scores with scores based on vector, BM25 hybrid retrieval, and Cross-Encoder reordering, achieving a fundamental leap from pure acoustic optimization to semantic and historical empirical optimization. This effectively overcomes the error propagation problem caused by acoustic ambiguity. This improvement significantly reduces the error rate in recognizing chemical engineering colloquial terms, enhances the robustness of recognition in complex acoustic environments, and provides highly reliable text input for subsequent anomaly detection and compliance auditing, realizing intelligent and precise management of chemical safety production. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a chemical safety production information management method based on a large model, provided as an embodiment of the present invention.

[0017] Figure 2 This is a basic flowchart of a chemical safety production information management system based on a large model, provided as an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1, referring to Figure 1 As an embodiment of the present invention, a chemical safety production information management method based on a large model is provided, comprising the following steps: Step S1: After converting the original audio, perform recognition to obtain a list of clean speech segments of a single speaker; Step S2: Decode the clean speech segment of a single speaker to obtain the N-best candidate text list and the original acoustic score; Step S3: Perform a mixed retrieval on each candidate text in the N-best candidate text list to obtain a reference text set. Calculate the re-ranking score for each candidate text based on the reference text set. Fuse the re-ranking score with the normalized original acoustic score to obtain the final corrected text. Step S4: Based on the final error correction text, perform anomaly detection to obtain a preliminary audit mark set and a comprehensive anomaly report; Step S5: Perform compliance assessment on the final corrected text to obtain compliance audit results, and manually review the preliminary audit mark set to obtain the final audit mark set. Based on the final audit mark set, compliance audit results, and comprehensive anomaly report, generate an audit report.

[0020] In one embodiment, the present invention addresses the problems of low speech recognition accuracy, difficulty in understanding technical terms, and lack of in-depth semantic analysis in post-audit processes under complex chemical acoustic environments. Step S1 involves converting and recognizing the original audio to obtain a list of clean speech segments for a single speaker, effectively eliminating noise and overlapping speech interference, providing high-quality audio input for subsequent decoding. Step S2 decodes the clean speech segments for a single speaker to obtain an N-best candidate text list and the original acoustic score. Step S3 performs a mixed search on multiple candidate texts to obtain a reference text set. Based on the reference text set, a re-ranking score is calculated for each candidate text. The normalized original acoustic score is then weighted and fused with the re-ranking score, and the candidate text with the highest final score is selected as the final error-correcting text. This overcomes the limitations of relying solely on... The acoustic score has the disadvantage of being difficult to distinguish between good and bad when multiple candidate confidence levels are close, which significantly reduces the term recognition error rate; Step S4 automatically discovers violations and cross-time anomalies based on the final corrected text, and performs anomaly judgment based on the final corrected text to obtain a preliminary audit tag set and a comprehensive anomaly report, realizing comprehensive capture of isolated violations and hidden repetitive anomalies; Step S5 performs compliance judgment on the final corrected text to obtain compliance audit results, and performs manual review of the preliminary audit tag set to obtain the final audit tag set. Based on the final audit tag set, compliance audit results and comprehensive anomaly report, an audit report is generated, forming a closed-loop intelligent audit process, realizing full-process information management driven by voice data in chemical production, and improving the recognition robustness and the accuracy and comprehensiveness of safety audit in complex acoustic environments.

[0021] Step S1 specifically includes: Step S101: Collect raw audio from the chemical production process; The original audio is converted using audio resampling technology to obtain audio data; Audio data is converted using channel conversion technology to obtain standardized audio data; Step S102: Perform speech recognition on the standardized audio data to obtain a list of clean speech segments from a single speaker.

[0022] In one embodiment, step S1 separates a single-speaker clean speech segment from the raw audio of a complex chemical site, eliminating noise, non-human voices, overlapping speech, and speaker aliasing interference, providing high-quality audio input for subsequent decoding and auditing. Step S1 is the data entry point for the entire information management method. Only by obtaining a clean speech segment free of noise, overlapping speech, and distinguishable speakers can the accuracy of subsequent terminology recognition, error correction, and auditing be guaranteed. Step S1 specifically includes the following sub-steps: Step S101, collecting raw audio from the chemical production process, including walkie-talkie conversations, patrols, etc. The voice reports from inspections and shift handover recordings constitute the carriers of voice data at the chemical plant site. To ensure the uniformity and stability of subsequent processing, the original audio is converted using audio resampling technology, with the sampling rate uniformly set to 16kHz. This sampling rate was determined experimentally, specifically by collecting 200 typical voice samples from the chemical plant site, each approximately 10 seconds long, including walkie-talkie conversations, inspection voice reports, and shift handover recordings. These samples were then resampled to 8kHz, 16kHz, 22.05kHz, and 44.1kHz, respectively, and then the same Whisper model (OpenAI) was used. Transcription was performed using Whisper-base, and the processing time for each audio segment was recorded. The results showed that at an 8kHz sampling rate, the high-frequency band was truncated, and high-frequency information of consonants and fricatives was lost, resulting in a term recognition error rate that was about 15% higher than that at 16kHz. Compared with 16kHz, the recognition accuracy improved by less than 2% at 22.05kHz and 44.1kHz, but the computational load increased by about 40% and 120% respectively. Considering both recognition performance and computational efficiency, 16kHz can retain the frequency band within 8kHz where the main energy of chemical engineering speech is located, meeting the accuracy requirements of speech activity detection and voiceprint feature extraction, while avoiding the computational redundancy caused by further increasing the sampling rate (such as 22.05kHz or 44.1kHz). Therefore, 16kHz was set. Then, the resampled audio data was converted using channel conversion technology to merge or extract multi-channel audio into mono, resulting in standardized audio data. The standardized audio data is in mono PCM format, which eliminates phase difference interference between channels. Step S102 involves performing speech recognition on the standardized audio data to obtain a list of clean speech segments for a single speaker. Specifically, this includes: using speech activity detection to identify human voice segments from non-human voice segments, retaining human voice segments to obtain a sequence of human voice segments. This step eliminates pure noise and equipment silence segments, reducing unnecessary computation. For the obtained human voice segment sequence, the Mel-frequency cepstral coefficient extraction algorithm is used to extract MFCC features, resulting in an MFCC feature vector sequence. MFCC features can characterize the spectral envelope of human voice and have a certain robustness to mechanical noise in chemical plant environments. A change point detection algorithm based on the Bayesian information criterion is used on the MFCC feature vector sequence to identify the switching points of different speakers, obtaining a set of speaker switching point locations. The Bayesian information criterion, by comparing the likelihood gain and complexity penalty before and after segmentation, can automatically determine switching points without pre-specifying the number of speakers. When the incremental threshold of the Bayesian information criterion is set to 0.95, the sum of the false negative and false positive rates for switch point detection is minimized. This threshold is set based on traversing from 0.5 to 1.0 with a step size of 0.01 on a 1000-segment labeled multi-person dialogue test set in the chemical industry, and taking the 0.95 corresponding to the highest F1 score. Simultaneously, an overlapping speech detection algorithm based on harmonic energy ratio and a preset threshold is used for human voice speech segments. The ratio of harmonic energy to total energy in each frame is calculated. If the ratio is less than the energy threshold, it is determined to be a suspected overlapping speech region, and a speech overlapping region set is obtained; otherwise, it is determined to be a non-overlapping speech region and is not added to the speech overlapping region set. The energy threshold is set to 0.32, which specifically includes: collecting 200 segments of on-site dialogue in the chemical industry with clear overlapping and non-overlapping annotations, calculating the harmonic energy ratio of each frame, and determining the threshold of 0.32 corresponding to the maximum Youden index point through ROC curve analysis. This value can make the accuracy of overlapping speech detection reach more than 92%. The human voice speech segment sequence is segmented by combining the speaker switching point location set and the speech overlap region set. The segmentation specifically includes: removing all time intervals belonging to the speech overlap region set from the human voice speech segment sequence to obtain a non-overlapping speech segment sequence; then, further segmenting the non-overlapping speech segment sequence according to the speaker switching point location set, ensuring that each segment does not contain any switching points, resulting in independent speech segments arranged in chronological order. This segmentation order guarantees that each independent speech segment does not contain overlapping speech or speaker switching points, ensuring the purity of the segment. If the duration of an independent speech segment is less than 0.1 seconds, the segment is discarded because segments that are too short (less than 0.1 seconds) cannot contain identifiable speech units (such as complete syllables or phonemes) and have no practical significance for subsequent auditing. Assigning speaker IDs to these independent speech segments involves: Calculating the cosine similarity between the speaker ID vectors of each segment using a hierarchical clustering algorithm, based on the speaker ID feature vector of each segment. The speaker ID feature vector extraction method is as follows: For each segmented independent speech segment, a pre-trained ECAPA-TDNN speaker recognition model is used to extract the speaker embedding vector. This model is based on a time-delay neural network architecture, pre-trained on the VoxCeleb large open-source speaker dataset, and outputs a fixed-dimensional d-vector with a dimension of 512. If the duration of an independent speech segment is less than 0.5 seconds, the segment is discarded, as segments shorter than 0.5 seconds are insufficient to extract reliable speaker-discriminative speaker ID features. The extracted speaker ID feature vectors of each segment are then hierarchically clustered, and the cosine similarity between segments is calculated. Similarity is used to classify segments with similarity greater than a preset clustering threshold as belonging to the same speaker and assigning them the same speaker ID; segments with similarity lower than the preset clustering threshold are assigned different speaker IDs. The clustering threshold is set to 0.75, based on the following criteria: using 200 sets of known-labeled speech segments from the chemical plant site, traversing from 0.60 to 0.90 with a step size of 0.01, the optimal threshold of 0.75 is selected when the clustering purity and silhouette coefficient are combined. This threshold can effectively distinguish different speakers while avoiding the incorrect segmentation of speech segments from the same speaker into multiple different speaker IDs. Finally, each independent speech segment is attached with a timestamp and the assigned speaker ID label to obtain a list of pure speech segments for a single speaker. Each element in this list clearly records the start and end times of the speech, the speaker identifier, and excludes overlapping speech content. Step S1 achieves high-quality preprocessing of complex audio from chemical plant sites: speech activity detection reduces subsequent computation by approximately 40%; a 16kHz sampling rate and mono PCM format ensure the stability of feature extraction; switch point detection based on Bayesian information criterion accurately captures speaker changes without requiring a preset number of speakers; a harmonic energy ratio threshold of 0.32 enables high-precision filtering of overlapping speech; a hierarchical clustering threshold of 0.75 keeps the error rate of merging same speakers and distinguishing different speakers below 5%; discarding segments shorter than 0.5 seconds avoids clustering noise caused by invalid feature extraction, and the final output list of clean single-speaker speech segments eliminates multi-speaker aliasing and noise interference from the source.

[0023] Step S2 specifically includes: Build a chemical engineering terminology database; For a single-speaker clean speech segment, the Whisper model is used for decoding. During the decoding process, a beam search algorithm based on hot word enhancement is called to construct a hot word prefix tree from the chemical terminology database to obtain candidate text sequences. When a candidate text sequence matches any hot word path in the hot word prefix tree, a preset positive score is added to the candidate text sequence, and the N-best candidate text list and the original acoustic score corresponding to each candidate text sequence are output.

[0024] In one embodiment, step S2 converts a single-speaker clean speech segment into a machine-readable candidate text sequence and enhances the decoding process using chemical engineering terminology, thereby improving the accuracy of terminology recognition. This step connects speech signals with semantic understanding. By constructing a chemical engineering terminology database and assigning positive scores to hot words during decoding, the Whisper model has a priori bias in recognizing specialized terms, providing a high-quality N-best candidate list for subsequent hybrid retrieval error correction. Step S2 specifically includes the following sub-steps: Step S201: Construct a chemical terminology database. The database construction process specifically includes: collecting raw audio from walkie-talkie conversations, inspection voice reports, and shift handover recordings at the chemical production site, as well as corresponding work record texts for the relevant time periods, such as electronic work tickets and operation logs; using the work record text as the correct text, transcribing the raw audio into recognition text using the Whisper model; employing the minimum edit distance algorithm to align the recognition text and correct text word by word, locating inconsistencies between the two texts; extracting words or phrases from each inconsistency in the recognition text as candidates for colloquial expressions; and extracting corresponding words or phrases from the correct text as candidates for standardized terms, thus forming a candidate mapping. For mapping pairs, the frequency of each candidate mapping pair in all parallel corpora is counted. Candidate mapping pairs with a frequency of 3 or higher are retained. The frequency threshold of 3 is based on the following: statistical analysis of 1000 sets of parallel corpora shows that when the frequency is below 3, about 70% of the mapping pairs are accidental recognition errors rather than stable spoken variations, and retaining them would introduce noise; when the frequency is 3 or higher, the accuracy of the mapping pairs can reach over 95%. Therefore, the frequency threshold is set to 3. Simultaneously, the pronunciation of standard terms and colloquial expressions are converted into phoneme sequences. Specifically, the Pinyin scheme is used, converting each Chinese character into a phoneme string composed of the corresponding initial consonant and final vowel. For example, "emergency parking system" is converted to "jin ji ting che xi". The phoneme list for "tong" is used to calculate the edit distance between the standardized terminology phoneme sequence and the colloquial expression phoneme sequence. The larger of the lengths of the standardized terminology phoneme sequence and the colloquial expression phoneme sequence is selected, and the quotient of the edit distance and the larger value is calculated to obtain the normalized distance. If the normalized distance is less than or equal to 0.6, the mapping pair is retained; otherwise, it is considered a recognition error and discarded. The normalized distance threshold of 0.6 is set based on: collecting 500 manually annotated colloquial-standard terminology pairs, calculating the normalized distance distribution, and ensuring that the pronunciation similarity of mapping pairs less than or equal to 0.6 is distinguishable to the human ear and not easily confused, while mapping pairs greater than 0.6 are mostly true recognition errors. Leaving it behind will lead to incorrect hot word enhancement; the filtered mapping pairs will be stored in the chemical terminology database in key-value pair format, where the key is the standard term and the value is the corresponding set of colloquial expressions; for example, the set of colloquial expressions corresponding to the key "emergency stop system" may include "emergency stop system" and "urgent stop"; the update strategy of the chemical terminology database is to perform the above construction process on the newly generated parallel corpus every 30 days, add the newly filtered mapping pairs to the chemical terminology database, and delete old mapping pairs that have not appeared for two consecutive periods and have a frequency of less than 3. The 30-day update cycle is based on the fact that chemical field operation tickets and operation logs are usually archived monthly. 30 days can ensure that the terminology database reflects the newly emerging colloquial variations in a timely manner, while avoiding the computational overhead caused by frequent updates; Step S202 involves decoding the clean speech segment from a single speaker using the Whisper model, specifically the open-source Whisper-base version from OpenAI. The Whisper-base version exhibits robust recognition capabilities for Chinese speech in industrial noise environments, effectively suppressing background noise commonly found in chemical plants, such as pump operation and valve movement. The input audio sampling rate is 16kHz, consistent with the sampling rate of the standardized audio data output in step S101. During decoding, a bundle search algorithm based on hotword enhancement is invoked to construct a hotword prefix tree from the chemical terminology database built in step S201. The hotword prefix tree is a Trie tree structure, where each node represents a Chinese character. A hot word path is a sequence of nodes traversed from the root node to a leaf node, corresponding to a complete formal term or its colloquial expression. For example, the formal term "emergency stopping system" and its colloquial expression "emergency stopping system" are both inserted into the hot word prefix tree. During decoding, the Whisper model generates candidate text sequences at each time step. When a sequence matches any hot word path in the hot word prefix tree, a preset positive score is added to the candidate text sequence. The positive score is set to 3.2, based on the fact that on a test set of 500 chemical engineering terms, the F1 score from 2.0 to 5.0 is traversed with a step size of 0.2, and 3.2 is selected to achieve the highest F1 score for term recognition. This score can improve the term hit rate without being too high. To mitigate the issue of missed detections due to the suppression of non-terminal paths, the final output is an N-best candidate text list, with N set to 5, along with the original acoustic scores for each candidate text sequence. The setting of N to 5 is based on experiments. The experimental process specifically includes: collecting 2000 audio samples from the chemical plant site, each segment ranging from 5 to 15 seconds in length, covering three sources: walkie-talkie conversations, inspection voice reports, and shift handover recordings. These samples were manually transcribed into correct text, and the Whisper-base model was used to decode each audio segment, generating an N-best candidate list. The N values ​​were set to 3, 5, 10, and 20, respectively. The proportion of correct text appearing in the top N candidates (recall rate) and the average decoding time were statistically analyzed. Experimental results show that when N=3, the recall rate is [missing value]. The recall rate was 94.2%; with N=5, the recall rate was 98.6%; with N=10, the recall rate was 99.3%, an improvement of 0.7% compared to N=5; with N=20, the recall rate was 99.5%, an improvement of 0.9% compared to N=5. In terms of decoding time, with N=5 as the baseline, the decoding time increased by about 28% with N=10 and by about 65% with N=20. Considering the recall rate gain and computational cost, the recall rate reached 98.6% with N=5. Increasing N further would yield less than 1% in error correction but would increase the computational cost. Therefore, N=5 was determined to be the optimal value. The original acoustic score is output by the decoder of the Whisper model, representing the log probability of the Whisper model generating the text sequence given the acoustic features. Through the above steps S2, the automatic construction and dynamic updating of the chemical terminology database, as well as the decoding process of hot word enhancement, were realized. The normalized distance threshold of 0.6 for the mapping pairs in the terminology database ensured the similarity of pronunciation between spoken expressions and standardized terms, avoiding incorrect mapping; the frequency threshold of 3 eliminated accidental recognition errors; the positive score of 3.2 for hot words improved the recognition recall rate of chemical professional terms by about 18% compared with the absence of hot words; the N-best value of 5 achieved a balance between computational efficiency and error correction potential. The final output of the N-best candidate text list and the original acoustic score improved the terminology accuracy of speech recognition in complex chemical scenarios.

[0025] Step S3 specifically includes: Step S301: Construct a knowledge base based on the pairing of historical speech recognition text with the corrected text and the explanation of technical terms; Step S302 involves performing a mixed retrieval on each candidate text in the N-best candidate text list, specifically including: For each candidate text in the N-best candidate text list, vector semantic retrieval and BM25 keyword retrieval are used respectively to obtain historical paired vector retrieval results and historical paired BM25 retrieval results; Take the union of historical pairing vector retrieval results and historical pairing BM25 retrieval results, sort them in descending order according to the final similarity, and obtain the sorted merged retrieval result set; The top K corrected texts in the sorted and merged search results set are selected as the reference text set. Where K is greater than or equal to 1; If the number of correct texts in the merged search results after sorting is less than K, placeholder texts are used to make up to K, resulting in a set of reference texts for each candidate text. Otherwise, directly take the first K corrected texts as the reference text set.

[0026] Step S3 also includes: Step S303: Process each candidate text and its corresponding set of reference texts according to the re-ranking score calculation rules to obtain the re-ranking score of each candidate text. The specific rules for calculating reordering scores include: The Cross-Encoder model is used to calculate the relevance score of each corrected text that is not a placeholder in the candidate text and the reference text set. The arithmetic mean of all relevance scores is taken as the re-ranking score of the candidate text. If the reference text set consists entirely of placeholder text, then the reordering score for the candidate text is set to 0. Step S304: The original acoustic score corresponding to each candidate text is processed by Softmax normalization to obtain the normalized original acoustic score of each candidate text. Step S305: The re-ranking scores of all candidate texts are processed using Softmax normalization to obtain the normalized re-ranking score of each candidate text. The weighted sum of the normalized original acoustic score and the normalized reordered score for each candidate text is used to obtain the final score. Select the candidate text with the highest final score as the final text to be corrected; The final corrected text is taken directly from the N-best candidate text list.

[0027] In one embodiment, step S3 performs knowledge fusion correction on the N-best candidate text list output in step S2. It re-ranks each candidate text using historical correct corpus and professional terminology explanations, and then weights and fuses them with the original acoustic scores to select the final corrected text with the most accurate semantics. This step achieves a leap from purely acoustic optimization to semantic plus historical experience optimization, overcoming the deficiency that acoustic scores alone cannot distinguish between good and bad candidates when multiple candidate confidence levels are close, and reducing the terminology recognition error rate. Step S3 specifically includes the following sub-steps: Step S301: Construct a knowledge base based on the pairing of historical speech recognition text and corrected correct text, as well as the explanation of professional terms. Historical speech recognition text refers to the original recognition result text that has not been corrected in the previous speech recognition process; corrected correct text is the corresponding standard answer text that has been manually checked and confirmed; each historical speech recognition text and its corrected correct text constitute a pair; the explanation of professional terms is compiled based on the "Chemical Safety Operation Procedures", accident investigation reports, and equipment manuals. Each term corresponds to a piece of explanatory text to help understand the semantic connotation of the term. For example, the explanation of emergency shutdown system is a protection system used to automatically or manually trigger the production process to safely stop under dangerous conditions. Store the above pairings and explanations of terms in the knowledge base, and construct vector indexes and inverted indexes for all historical speech recognition texts in the knowledge base, so as to carry out vector semantic retrieval and BM25 keyword retrieval in the future. Step S302 involves performing a hybrid retrieval on each candidate text in the N-best candidate text list. Specifically, this includes: using vector semantic retrieval and BM25 keyword retrieval for each candidate text; retrieving historical pairings from the knowledge base constructed in step S301 to obtain historical pairing vector retrieval results and historical pairing BM25 retrieval results; using the Sentence-BERT model to encode candidate texts into vectors, calculating cosine similarity with the vectors of historical speech recognition texts in the knowledge base, and returning the top K historical pairings with the highest similarity; BM25 retrieval, based on word frequency and inverse document frequency, performs keyword matching on historical speech recognition texts in the knowledge base, returning the top K historical pairings with the highest relevance. The value of K is set to 5 experimentally. Specifically, K is set to 3, 5, and 10 on 500 test candidate texts, and the accuracy of the final corrected text is statistically analyzed. When K equals 5, the accuracy reaches 96.2%, while when K equals 10, the accuracy only increases by 0.5% but the retrieval time increases by approximately 40%, therefore K equals 5; historical pairings are retrieved... The union of vector retrieval results and historical paired BM25 retrieval results is sorted in descending order by final similarity to obtain the sorted merged retrieval result set. The final similarity is calculated as follows: for each result in the historical paired vector retrieval results, Min-Max normalization is used to map the cosine similarity to the [0,1] interval to obtain the historical vector normalization score. All results with historical vector normalization scores constitute the historical vector normalization result set. Min-Max normalization is calculated based on the maximum and minimum values ​​in the current retrieval result set. If the maximum value of all cosine similarities is equal to the minimum value, all scores are normalized to 0.5. For each result in the historical paired BM25 retrieval results, Min-Max normalization is used to map the BM25 score to the [0,1] interval to obtain the historical BM25 normalization score. All results with historical BM25 normalization scores constitute the historical BM25 normalization result set. Similarly, if the maximum value of all BM25 scores is equal to the minimum value, all scores are normalized to 0.5. If a result appears only in either the historical vector normalization result set or the historical BM25 normalization result set, the normalization score corresponding to that set is directly used as the final similarity. If a result appears in both result sets, the larger of the vector normalization score and the BM25 normalization score is taken as the final similarity. For the same historical pairing, only one record is retained, and its final similarity is determined according to the above rules. That is, if the pairing appears only in either the historical vector normalization result set or the historical BM25 normalization result set, the normalization score corresponding to that set is directly used as the final similarity; if the pairing appears only in one of the historical vector normalization result sets or the historical BM25 normalization result set, the normalization score corresponding to that set is directly used as the final similarity. If a pair appears in both result sets, the larger of the vector normalized score and the BM25 normalized score is taken as the final similarity. The final similarities are then sorted in descending order to obtain the merged search result set. The top K corrected texts in the merged search result set are selected as the reference text set. If the number of corrected texts in the merged search result set is less than K, placeholder texts are used to make up to K, specifically placeholder texts without similar references, thus obtaining the reference text set for each candidate text. Otherwise, the top K corrected texts are directly taken as the reference text set. Step S303: Process each candidate text and its corresponding reference text set according to the re-ranking score calculation rules to obtain the re-ranking score of each candidate text. The re-ranking score calculation rules specifically include: using a Cross-Encoder model to calculate the relevance score between the candidate text and each corrected correct text that is not a placeholder in the reference text set. The Cross-Encoder model uses a pre-trained bert-base-chinese model and is fine-tuned on chemical engineering historical pairing data. Fine-tuning means that based on the pre-trained bert-base-chinese model, using chemical engineering historical pairing data, i.e., pairs consisting of candidate text and corrected correct text, it continues to train for 5 rounds with a learning rate of 2e-5 and a batch size of 16. The cross-entropy loss function is used to adjust the parameters of the bert-base-chinese model so that the model can adapt to semantic relevance judgment in the chemical engineering field. The input is the concatenation of the candidate text and the corrected correct text, and the output is a relevance score between 0 and 1. The arithmetic mean of all relevance scores is taken as the re-ranking score of the candidate text. If all the reference text sets are placeholder texts, the re-ranking score of the candidate text is set to 0. Step S304: The original acoustic score corresponding to each candidate text is processed by Softmax normalization to obtain the normalized original acoustic score of each candidate text. The Softmax normalization process is as follows: the original acoustic scores are converted into normalized original acoustic scores with values ​​in the range [0,1] and a sum of 1. Step S305: The re-ranking scores of all candidate texts are processed using Softmax normalization to obtain a normalized re-ranking score for each candidate text. Then, the normalized original acoustic score and the normalized re-ranking score of each candidate text are weighted and summed to obtain the final score. Specifically, this includes: calculating the product of the normalized original acoustic score and λ to obtain the first product; calculating the product of the normalized re-ranking score and (1-λ) to obtain the second product; and calculating the sum of the first and second products to obtain the final score. The weight coefficient of the normalized original acoustic score is λ, and the weight coefficient of the re-ranking score is... The value of λ is 1-λ, and the range of λ is [0,1]. λ is set to 0.6 according to the experiment. Specifically, on 1000 test samples, the values ​​of λ are traversed from 0.3 to 0.9 with a step size of 0.1. The value of λ that makes the final error correction accuracy the highest is selected. The experiment shows that the accuracy is 97.3% when λ is equal to 0.6, 96.8% when λ is equal to 0.5, and 96.9% when λ is equal to 0.7. Therefore, 0.6 is selected. That is, while retaining acoustic information, the re-sorting score weight is given, and the candidate text with the highest final score is selected as the final error correction text. The final error correction text is directly taken from the N-best candidate text list. Step S3 achieves hybrid retrieval and reordering error correction based on historical corpus and terminology knowledge. The hybrid retrieval of vector and BM25 combines the advantages of semantic similarity and keyword matching, improving the recall rate by about 12% compared to a single retrieval method. Min-Max normalization eliminates the difference in the scale of scores of different retrieval results, making the final similarity comparable. The cross-encoder reordering score improves the accuracy of relevance judgment by 15% compared to the similarity using vector retrieval alone. In the weighted fusion, λ equals 0.6, which improves the accuracy of the final corrected text from 89% to over 97% compared to using only acoustic scores.

[0028] The anomaly detection in step S4 specifically includes: Step S401: Based on punctuation marks and line breaks, the final corrected text is segmented into sentences to obtain multiple sentences. For each sentence, the Llama 2 language model is used to judge the violation, and the violation type, confidence level and violation clause number are output to obtain the sentence-level violation mark. The confidence levels include high, medium, and low. Step S402: The final corrected text is used as a dialogue fragment, and the Llama 2 large language model is used for judgment to obtain dialogue-level risk decision tags; Dialogue-level risk decision markers include violation type, confidence level, and violation clause number; The confidence levels include high, medium, and low. A preliminary audit tag set is constructed based on statement-level violation tags and dialogue-level risk decision tags.

[0029] Step S4 also includes: Step S403: Use regular expression matching to extract key entities and state descriptions from the dialogue segments, input the key entities and state descriptions into the Sentence-BERT model for encoding, obtain semantic vectors, and store them in the vector index library. According to a preset cycle, key entities are used as query conditions to retrieve historical dialogue fragments from the vector index library to obtain search results. The search results are input into the Llama 2 language model to determine the same state description corresponding to the same key entity in the search results. Specifically, this includes: If the same state description for the same key entity is mentioned more than once, a comprehensive anomaly report marked as content abnormality will be generated. Otherwise, generate a comprehensive anomaly report marked as having no abnormalities.

[0030] In one embodiment, step S4 automatically detects violations and cross-period anomalies based on the final corrected text. Through statement-level violation judgment, dialogue-level risk decision-making, and cross-period entity state tracking, it achieves comprehensive capture of isolated violations and hidden recurring anomalies. This step combines the reasoning capabilities of the Llama 2 language models with Sentence-BERT semantic encoding to transform the text into structured audit markers and anomaly reports, providing a foundation for subsequent compliance assessments and manual review. Step S4 specifically includes the following sub-steps: Step S401: Based on punctuation and line breaks, the final corrected text is segmented into multiple sentences. For each sentence, the Llama 2 language model is used for violation judgment. The Llama 2 language model has been fine-tuned on a dataset including chemical safety regulation clause numbers and corresponding violation types. Fine-tuning refers to training the pre-trained Llama 2 language model for three more rounds using the chemical safety regulation clause dataset (each sample includes a violation sentence and its corresponding violation type and clause number). The learning rate is set to 1e-5, the batch size is 8, and the cross-entropy loss function is used to adjust the parameters of the Llama 2 language model so that it can output the corresponding clause number based on the input sentence. During inference, prompts are designed to make the Llama 2 language model output results in a specified format. Specifically, the sentence to be judged is concatenated with the task description and output format requirements to form a prompt text. The prompt text is input into the Llama 2 language model, which generates structured output that meets the requirements. By parsing the text generated by the Llama 2 language model, the violation type, confidence level, and violation clause number are obtained. The confidence level includes high, medium, and low. This yields a statement-level violation tag. This step allows individual operation instructions or reporting statements to be quickly categorized into specific violation clauses, achieving refined violation localization. Step S402 involves using the final corrected text as a dialogue fragment and employing the Llama 2 language model for risk decision assessment. The Llama 2 language model has also been fine-tuned to identify risk decision types and associate them with violation clauses. Output is guided by designed prompts, specifically: concatenating the dialogue fragment with the task description and output format requirements into a prompt text; inputting this prompt text into the Llama 2 language model; generating structured output; parsing the output to obtain the violation type, confidence level, and violation clause number, thus obtaining a dialogue-level risk decision marker; the confidence level also includes high, medium, and low; and constructing a preliminary audit marker set based on the statement-level violation markers and the dialogue-level risk decision markers. This step identifies risk decision behavior from the dialogue context, mitigating potential chain-of-command vulnerabilities that might be overlooked in single-sentence judgments. Step S403 involves using regular expression matching to extract key entities and state descriptions from the dialogue fragments. Specifically, this includes: writing regular expression rules for matching equipment numbers and work section names, such as matching equipment numbers consisting of letters and numbers (e.g., P-101, R02) and work section names (e.g., pyrolysis section, polymerization section); and regular expression rules for matching state verb phrases, such as abnormal noise, leakage, fluctuation, and overpressure; using the matched text as key entities and state descriptions respectively; and concatenating each extracted key entity and state description into a phrase, such as "pump-101 abnormal noise," inputting Sentence-BER. The T-model is encoded into a semantic vector, and then this vector, along with the original text of the dialogue segment and timestamp information, is stored in a vector index. The vector index supports approximate nearest neighbor retrieval based on cosine similarity, and performs the following operations at a preset period: each key entity and state description extracted from the current dialogue segment is used as a query condition, that is, the key entity and state description are first concatenated into a phrase, and then encoded into a query vector using the same Sentence-BERT model. Then, historical dialogue segments with a cosine similarity greater than 0.85 with the query vector are retrieved from the vector index to obtain the retrieval results. The preset period is set to 24 hours, and chemical production is usually carried out by shift handover. Four hours can cover all dialogues of a complete production day, while avoiding excessively frequent searches that would lead to computational burden. The cosine similarity threshold of 0.85 is set based on the following: on 500 labeled historical dialogue segments, the precision and recall for retrieving the same entity-state pair are calculated by traversing from 0.7 to 0.95 with a step size of 0.05. The threshold with the highest F1 score is selected, and experiments show that the F1 score reaches 94% at 0.85, so it is set to 0.85. For all historical dialogue segments in the search results, the key entities, state descriptions, and timestamps stored in the historical dialogue segment are parsed. If there are historical dialogue segments with the same key entity (string exact match) and the same state description (string exact match), the key entity-state pair is retrieved. If the number of historical mentions is greater than 1 (i.e., there is at least one historical fragment besides the current fragment), a comprehensive anomaly report marked as having abnormal content is generated. The report includes key entities, status descriptions, first occurrence time, most recent occurrence time, and total number of mentions. Otherwise, a comprehensive anomaly report marked as having no abnormal content is generated. When the vector index is empty on the first execution, the search results are an empty set, and a comprehensive anomaly report marked as having no abnormal content is generated directly. This step, by tracking the repeated abnormal status of the same equipment or work section across time periods, realizes the automatic identification of recurring hidden dangers or unresolved faults and generates comprehensive anomaly reports, effectively solving the shortcomings of traditional methods that cannot associate historical information. Step S4 enables the generation of comprehensive audit tags, ranging from single-sentence violations and dialogue risks to cross-time-period anomalies. The finely tuned Llama 2 language model achieves an accuracy rate of over 90% in single-sentence violation identification; dialogue-level risk decision-making can detect approximately 85% of covert risk decision-making behaviors; cross-time-period entity status tracking is automatically executed every 24 hours, which can identify repeated mentions of the same device and the same abnormal status as persistent anomalies, avoiding missed reports. The final output of the preliminary audit tag set and comprehensive anomaly report provides input for the compliance determination and manual review in step S5.

[0031] Step S5, which involves determining the compliance of the final corrected text, specifically includes: Step S501: Perform semantic role labeling on the final corrected text to obtain candidate operation behavior phrases; Perform keyword matching on candidate action phrases and output action descriptions; A safety procedure knowledge base is constructed, and behavioral descriptions are used as query conditions. Vector semantic retrieval is then used to search the safety procedure knowledge base to obtain safety procedure vector retrieval results. The BM25 keyword was used to search the safety regulations knowledge base to obtain the safety regulations BM25 search results. Take the union of the safety procedure vector retrieval results and the safety procedure BM25 retrieval results, sort them in descending order according to the final comprehensive similarity, and obtain the output retrieval result set; Select the first Q corrected texts from the output search results set as the reference text set; Step S502: Input the behavior description and procedure clause text into the Llama 2 language model to make a compliance judgment, and output the compliance judgment result, severity level and judgment basis text to obtain the compliance audit result; Step S503: Mark the initial audit flags with low confidence levels as pending manual review to obtain the audit list; Entries with a confidence level of medium or high are not marked.

[0032] Step S5 also includes: Step S504: Receive manual review instructions input by security management personnel through the interactive interface for the markers in the pending review list; Manual review instructions include confirmation, correction, or rejection; The confirmation command means to keep the original mark unchanged, the correction command means to modify the original mark to the correction value entered by the user, and the rejection command means to delete the original mark. The tags are manipulated according to the manual review instructions to generate feedback records, resulting in the final audit tag set and feedback records. Step S505: Generate an audit report from the final audit tag set, compliance audit results, and comprehensive anomaly report. The audit report includes a statistical summary, violation cases, and rectification recommendations. Generate a rectification task list based on the audit report, and output the audit report and rectification task list.

[0033] In one embodiment, step S5 performs compliance judgment on the final corrected text and introduces a closed loop of manual review. Operational behaviors are extracted through semantic role labeling and keyword matching, combined with a safety procedure knowledge base for hybrid retrieval, and compliance judgment is achieved using the Llama 2 language model. Simultaneously, low-confidence audit markers are handed over to management personnel for confirmation or correction, ultimately generating a complete audit report and rectification task sheet. This step combines automated intelligent auditing with human experience review, ensuring both audit efficiency and the accuracy of key conclusions, thus realizing a closed loop for information management of chemical safety production. Step S5 specifically includes the following sub-steps:

[0034] Step S501: Semantic role labeling is performed on the final corrected text to obtain candidate operation behavior phrases. Then, keyword matching is performed on the candidate operation behavior phrases to output behavior descriptions. Semantic role labeling specifically includes: inputting the final corrected text into the BiLSTM-Transformer model, outputting the semantic role label corresponding to each predicate in the sentence, such as ARG0 representing the agent, ARG1 representing the patient, and ARGM-LOC representing the location, and extracting the predicates and their associated agent and patient phrases as candidate operation behavior phrases. Keyword matching specifically includes: pre-constructing a chemical operation verb library, which includes verbs such as accelerate, close, open, regulate, and replace. Depressurization and purging; candidate phrases obtained from semantic role labeling are segmented, and verb units are extracted. If the verb unit exists in the chemical operation verb library, a successful match is achieved, and the phrase is marked as an operation behavior phrase; otherwise, the candidate is discarded, and the successfully matched phrase is output as an operation description. This sub-step transforms the operation intent in natural language into a structured behavior description, providing accurate query input for subsequent compliance comparison; and a safety procedure knowledge base is constructed: internal chemical enterprise procedures documents such as "Chemical Safety Operation Procedures," "Emergency Response Cards," and "Job Operation Instructions" are collected, and the clause texts are segmented to establish a procedure clause text library; Sentenc is used... The e-BERT text embedding model encodes each clause text into a semantic vector and builds a FAISS vector index, while retaining the clause number and text content. This forms a safety procedure knowledge base that supports vector retrieval and keyword retrieval. Using behavioral descriptions as query conditions, the model performs searches in the safety procedure knowledge base using both vector semantic retrieval and BM25 keyword retrieval, obtaining safety procedure vector retrieval results and safety procedure BM25 retrieval results. The union of these two results is then taken and sorted in descending order according to the final comprehensive similarity, yielding the output retrieval result set. The calculation method for the final comprehensive similarity specifically includes: for each clause in the safety procedure vector retrieval results... For each result, Min-Max normalization is used to map the cosine similarity to the [0,1] interval to obtain the safety procedure vector normalized score. All results with this score constitute the safety procedure vector normalized result set. During Min-Max normalization, if the maximum value of all cosine similarities is equal to the minimum value, all scores are normalized to 0.5. For each result in the safety procedure BM25 retrieval results, Min-Max normalization is also used to map the BM25 score to the [0,1] interval to obtain the safety procedure BM25 normalized score, which constitutes the safety procedure BM25 normalized result set. If the maximum value of all BM25 scores is equal to the minimum value, all scores are normalized to 0.5. If a result appears only in the vector normalization result set or the BM25 normalization result set of the safety procedure, the normalization score corresponding to that set is directly used as the final comprehensive similarity. If a result appears in both result sets, the larger value between the vector normalization score and the BM25 normalization score is taken as the final comprehensive similarity. The merged result set is sorted in descending order according to the final comprehensive similarity, and the top Q corrected texts after sorting are selected as the reference text set. Q is set to 3 according to the experiment. The experiment process specifically includes: taking Q=1, 3, and 5 respectively on 500 test behavior descriptions, and calculating the accuracy of compliance judgment. When Q=3, the accuracy reaches 96.5%, while when Q=5, it only improves by 0.6% but the retrieval time increases by about 35%. Therefore, Q=3 is chosen. This sub-step ensures that the retrieved procedure clauses are semantically related to and comprehensively covered with the behavior descriptions through hybrid retrieval and normalization fusion, providing a reliable basis for compliance judgment.

[0035] Step S502 involves inputting the behavior description and the retrieved procedural clause text into the Llama 2 language model for compliance judgment, outputting the compliance judgment result, severity level, and judgment basis text to obtain the compliance audit result. The compliance judgment specifically includes: concatenating the behavior description and the retrieved procedural clause text with the task description and output format requirements to form a prompt text; inputting this prompt text into the Llama 2 language model; the Llama 2 language model generating a structured text that meets the requirements; and parsing this text to obtain the compliance judgment result, severity level, and judgment basis text. This sub-step utilizes the reasoning capabilities of the Llama 2 language model to automatically determine whether the operational behavior complies with the procedures, improving the efficiency and consistency of the compliance audit. Step S503: Mark the initial audit markers with low confidence levels as pending manual review to obtain the audit list; do not mark the markers with medium or high confidence levels. This sub-step filters out the results with low confidence levels, avoids the impact of automated audit misjudgments on the final conclusion, and ensures that high-risk or ambiguous items are reviewed manually. Step S504: For the markers in the audit list, receive manual review instructions input by security management personnel through the interactive interface. The manual review instructions include confirmation, correction, or rejection: the confirmation instruction means to keep the original marker unchanged, the correction instruction means to modify the original marker to the correction value input by the user, and the rejection instruction means to delete the original marker. The markers are operated on according to the manual review instructions, and feedback records are generated to obtain the final audit marker set and feedback records. This sub-step realizes the human-machine collaborative audit closed loop, which not only retains the efficiency of automated audit, but also incorporates the professional judgment of management personnel. Step S505 generates an audit report from the final audit mark set, compliance audit results, and comprehensive anomaly report. The audit report includes a statistical summary, violation cases, and rectification recommendations. Based on the audit report, a rectification task sheet is generated, and the audit report and rectification task sheet are output. This sub-step summarizes the multi-source audit results into actionable rectification tasks, directly serving the improvement of safety management in chemical production. Step S5 automates the entire process from behavior extraction, procedure retrieval, compliance assessment to manual review and rectification task generation. BiLSTM-Transformer semantic role labeling combined with keyword matching from a chemical operation verb library achieves an accuracy rate of over 92% in behavior description extraction. Vector and BM25 hybrid retrieval, along with Min-Max normalization fusion, improves the recall rate of procedure clause retrieval by approximately 15% compared to a single method. Llama, a two-language model, performs compliance assessment, achieving an accuracy rate of over 94% in identifying violations. The low-confidence labeling manual review mechanism controls the false positive rate of the final audit results to within 2%. The resulting audit report and rectification task sheet provide chemical enterprises with traceable and implementable safety management data.

[0036] Example 2, refer to Figure 2 In another embodiment of the present invention, which differs from the first embodiment, a chemical safety production information management system based on a large model is provided, including an identification module, a decoding module, a sorting module, a judgment module, and a generation module; The recognition module converts the original audio and then recognizes it to obtain a list of clean speech segments of a single speaker. The decoding module decodes the clean speech segments of a single speaker to obtain the N-best candidate text list and the original acoustic score; The sorting module performs a mixed retrieval on each candidate text in the N-best candidate text list to obtain a reference text set. Based on the reference text set, it calculates the re-sorting score for each candidate text and merges the re-sorting score with the normalized original acoustic score to obtain the final corrected text. The judgment module performs anomaly judgment based on the final error correction text, and obtains a preliminary audit mark set and a comprehensive anomaly report; The generation module performs compliance assessment on the final corrected text to obtain compliance audit results, and manually reviews the preliminary audit tag set to obtain the final audit tag set. Based on the final audit tag set, compliance audit results, and comprehensive anomaly report, an audit report is generated.

[0037] In one embodiment, the recognition module converts the original audio and then recognizes it to obtain a list of clean speech segments for a single speaker, effectively eliminating noise and overlapping speech interference, providing high-quality audio input for subsequent decoding. The decoding module decodes the clean speech segments for a single speaker to obtain an N-best candidate text list and the original acoustic score. The sorting module performs a mixed search on multiple candidate texts to obtain a reference text set, calculates the re-sorting score for each candidate text based on the reference text set, and performs a weighted fusion of the normalized original acoustic score and the re-sorting score, selecting the candidate with the highest final score as the final error-corrected text. This invention overcomes the limitation that acoustic scores alone are insufficient to distinguish between superior and inferior candidates when multiple candidate confidence levels are similar, significantly reducing the terminology recognition error rate. The judgment module automatically detects violations and cross-time-period anomalies based on the final corrected text, performs anomaly judgment based on the final corrected text, obtains a preliminary audit tag set and a comprehensive anomaly report, and achieves comprehensive capture of isolated violations and hidden recurring anomalies. The generation module performs compliance judgment on the final corrected text, obtains compliance audit results, and performs manual review of the preliminary audit tag set to obtain the final audit tag set. Based on the final audit tag set, compliance audit results, and comprehensive anomaly report, an audit report is generated.

[0038] On the same test dataset, the present invention was compared with the comparison method (general Whisper decoding + preset keyword string precise matching method). The test dataset consisted of 2,000 chemical field voice samples, covering walkie-talkie conversations, inspection voice reports and shift handover recordings. The experimental results are shown in Table 1.

[0039] Table 1. Comparison of the effects of the present invention and the comparative method; The comparison method lacks the ability to understand dialogue context and perform cross-time semantic retrieval, thus failing to achieve dialogue-level risk decision-making and cross-time recurring anomaly identification. This invention is the first to implement the above-mentioned auditing function. The comparison method uses the same Whisper-base model as this invention for speech recognition, but does not use chemical terminology database hot word enhancement and hybrid retrieval reordering error correction. The auditing part only relies on a preset list of violation keywords, such as violation verb phrases like "close," "skip," and "not record," and uses a string exact matching method (case-sensitive). Matching any one of them is judged as a violation, without using a large language model for judgment and semantic retrieval compliance judgment. The data in Table 1 shows that the present invention is significantly superior to existing methods in key indicators such as chemical terminology recognition, final error correction accuracy, multi-level violation detection, and cross-period anomaly tracking. In particular, the realization of dialogue-level risk decision-making and cross-period repeated anomaly recognition fills the gaps that traditional technologies cannot handle. At the same time, the terminology error rate is reduced by 56% in complex acoustic environments, and the overall audit false alarm rate is controlled within 2%.

[0040] This invention selects the optimal text solely based on the original acoustic score using existing algorithms. However, when multiple candidate texts have similar acoustic scores, it cannot effectively distinguish between them and lacks utilization of chemical engineering terminology and historical correct corpora, resulting in a high error rate in terminology recognition and limited error correction capabilities. This invention adds positive scores to the professional term path by using a hot word prefix tree in a chemical engineering terminology database, thereby improving the terminology recognition tendency. Furthermore, it integrates acoustic scores with vector-based, BM25 hybrid retrieval, and Cross-Encoder reordering scores, achieving a fundamental leap from purely acoustic optimization to semantic and historically empirical optimization. This effectively overcomes the error propagation problem caused by acoustic ambiguity. This improvement significantly reduces the error rate in recognizing chemical engineering colloquial terms, enhances the robustness of recognition in complex acoustic environments, and provides highly reliable text input for subsequent anomaly detection and compliance auditing, realizing intelligent and precise chemical safety production management.

[0041] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. 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.

[0042] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A chemical safety production information management method based on a large model, characterized in that, Includes the following steps: Step S1: After converting the original audio, perform recognition to obtain a list of clean speech segments of a single speaker; Step S2: Decode the clean speech segment of a single speaker to obtain the N-best candidate text list and the original acoustic score; Step S3: Perform a mixed retrieval on each candidate text in the N-best candidate text list to obtain a reference text set. Calculate the re-ranking score for each candidate text based on the reference text set. Fuse the re-ranking score with the normalized original acoustic score to obtain the final corrected text. Step S4: Based on the final error correction text, perform anomaly detection to obtain a preliminary audit mark set and a comprehensive anomaly report; Step S5: Perform compliance assessment on the final corrected text to obtain compliance audit results, and manually review the preliminary audit mark set to obtain the final audit mark set. Based on the final audit mark set, compliance audit results, and comprehensive anomaly report, generate an audit report.

2. The chemical safety production information management method based on a large model as described in claim 1, characterized in that: Step S1 specifically includes: Step S101: Collect raw audio from the chemical production process; The original audio is converted using audio resampling technology to obtain audio data; Audio data is converted using channel conversion technology to obtain standardized audio data; Step S102: Perform speech recognition on the standardized audio data to obtain a list of clean speech segments from a single speaker.

3. The chemical safety production information management method based on a large model as described in claim 2, characterized in that: Step S2 specifically includes: Build a chemical engineering terminology database; For a single-speaker clean speech segment, the Whisper model is used for decoding. During the decoding process, a beam search algorithm based on hot word enhancement is called to construct a hot word prefix tree from the chemical terminology database to obtain candidate text sequences. When a candidate text sequence matches any hot word path in the hot word prefix tree, a preset positive score is added to the candidate text sequence, and the N-best candidate text list and the original acoustic score corresponding to each candidate text sequence are output.

4. The chemical safety production information management method based on a large model as described in claim 3, characterized in that: Step S3 specifically includes: Step S301: Construct a knowledge base based on the pairing of historical speech recognition text with the corrected text and the explanation of technical terms; Step S302 involves performing a mixed retrieval on each candidate text in the N-best candidate text list, specifically including: For each candidate text in the N-best candidate text list, vector semantic retrieval and BM25 keyword retrieval are used respectively to obtain historical paired vector retrieval results and historical paired BM25 retrieval results; Take the union of historical pairing vector retrieval results and historical pairing BM25 retrieval results, sort them in descending order according to the final similarity, and obtain the sorted merged retrieval result set. The top K corrected texts in the sorted and merged search results set are selected as the reference text set. Where K is greater than or equal to 1; If the number of correct texts in the merged search results after sorting is less than K, placeholder texts are used to make up to K, thus obtaining a set of reference texts for each candidate text. Otherwise, directly take the first K corrected texts as the reference text set.

5. The chemical safety production information management method based on a large model as described in claim 4, characterized in that: Step S3 also includes: Step S303: Process each candidate text and its corresponding set of reference texts according to the re-ranking score calculation rules to obtain the re-ranking score of each candidate text. The specific rules for calculating reordering scores include: The Cross-Encoder model is used to calculate the relevance score of each corrected text that is not a placeholder in the candidate text and the reference text set. The arithmetic mean of all relevance scores is taken as the re-ranking score of the candidate text. If the reference text set consists entirely of placeholder text, then the reordering score for the candidate text is set to 0. Step S304: The original acoustic score corresponding to each candidate text is processed by Softmax normalization to obtain the normalized original acoustic score of each candidate text. Step S305: The re-ranking scores of all candidate texts are processed using Softmax normalization to obtain the normalized re-ranking score of each candidate text. The weighted sum of the normalized original acoustic score and the normalized reordered score for each candidate text is used to obtain the final score. Select the candidate text with the highest final score as the final text to be corrected; The final corrected text is taken directly from the N-best candidate text list.

6. The chemical safety production information management method based on a large model as described in claim 5, characterized in that: The anomaly detection in step S4 specifically includes: Step S401: Based on punctuation marks and line breaks, the final corrected text is segmented into sentences to obtain multiple sentences. For each sentence, the Llama 2 language model is used to judge the violation, and the violation type, confidence level and violation clause number are output to obtain the sentence-level violation mark. The confidence levels include high, medium, and low. Step S402: The final corrected text is used as a dialogue fragment, and the Llama 2 large language model is used for judgment to obtain dialogue-level risk decision tags; Dialogue-level risk decision markers include violation type, confidence level, and violation clause number; The confidence levels include high, medium, and low. A preliminary audit tag set is constructed based on statement-level violation tags and dialogue-level risk decision tags.

7. The chemical safety production information management method based on a large model as described in claim 6, characterized in that: Step S4 also includes: Step S403: Use regular expression matching to extract key entities and state descriptions from the dialogue segments, input the key entities and state descriptions into the Sentence-BERT model for encoding, obtain semantic vectors, and store them in the vector index library. According to a preset cycle, key entities are used as query conditions to retrieve historical dialogue fragments from the vector index library to obtain search results. The search results are input into the Llama 2 language model to determine the same state description corresponding to the same key entity in the search results. Specifically, this includes: If the same state description for the same key entity is mentioned more than once, a comprehensive anomaly report marked as content abnormality will be generated. Otherwise, generate a comprehensive anomaly report marked as having no abnormalities.

8. The chemical safety production information management method based on a large model as described in claim 7, characterized in that: Step S5, which involves determining the compliance of the final corrected text, specifically includes: Step S501: Perform semantic role labeling on the final corrected text to obtain candidate operation behavior phrases; Perform keyword matching on candidate action phrases and output action descriptions; A safety procedure knowledge base is constructed, and behavioral descriptions are used as query conditions. Vector semantic retrieval is then used to search the safety procedure knowledge base to obtain safety procedure vector retrieval results. The BM25 keyword was used to search the safety regulations knowledge base to obtain the safety regulations BM25 search results. Take the union of the safety procedure vector retrieval results and the safety procedure BM25 retrieval results, sort them in descending order according to the final comprehensive similarity, and obtain the output retrieval result set; Select the first Q corrected texts from the output search results set as the reference text set; Step S502: Input the behavior description and procedure clause text into the Llama 2 language model to make a compliance judgment, and output the compliance judgment result, severity level and judgment basis text to obtain the compliance audit result; Step S503: Mark the initial audit flags with low confidence levels as pending manual review to obtain the audit list; Entries with a confidence level of medium or high are not marked.

9. The chemical safety production information management method based on a large model as described in claim 8, characterized in that: Step S5 also includes: Step S504: Receive manual review instructions input by security management personnel through the interactive interface for the markers in the pending review list; Manual review instructions include confirmation, correction, or rejection; The confirmation command means to keep the original mark unchanged, the correction command means to modify the original mark to the correction value entered by the user, and the rejection command means to delete the original mark. The tags are manipulated according to the manual review instructions to generate feedback records, resulting in the final audit tag set and feedback records. Step S505: Generate an audit report from the final audit tag set, compliance audit results, and comprehensive anomaly report. The audit report includes a statistical summary, violation cases, and rectification recommendations. Generate a rectification task list based on the audit report, and output the audit report and rectification task list.

10. A chemical safety production information management system based on a large model, which is applied to the chemical safety production information management method based on a large model as described in any one of claims 1-9, characterized in that, It includes an identification module, a decoding module, a sorting module, a judgment module, and a generation module; The recognition module converts the original audio and then recognizes it to obtain a list of clean speech segments of a single speaker. The decoding module decodes the clean speech segments of a single speaker to obtain the N-best candidate text list and the original acoustic score; The sorting module performs a mixed retrieval on each candidate text in the N-best candidate text list to obtain a reference text set. Based on the reference text set, it calculates the re-sorting score for each candidate text and merges the re-sorting score with the normalized original acoustic score to obtain the final corrected text. The judgment module performs anomaly judgment based on the final error correction text, and obtains a preliminary audit mark set and a comprehensive anomaly report; The generation module performs compliance assessment on the final corrected text to obtain compliance audit results, and manually reviews the preliminary audit tag set to obtain the final audit tag set. Based on the final audit tag set, compliance audit results, and comprehensive anomaly report, an audit report is generated.

Images