A method, system, device, medium and product for monitoring a tour inspection task

Abstract

The application discloses a kind of monitoring methods, systems, equipment, medium and product of tour and inspection task, it is related to tour monitoring technical field, using the preset frozen large language model is fine-tuning training by tour feature set, obtain corresponding target large language model, wherein, target large language model includes information processing network and output network, obtain tour task text, tour task text is carried out semantic analysis by preset tour knowledge base and information processing network, obtain associated feature vector and multiple relevant evidence vectors, risk monitoring is carried out to associated feature vector by output network and each relevant evidence vector, obtain corresponding monitoring result. Overcome the technical problem that the monitoring method of existing tour and inspection task is multiple dependence artificial account or general text analysis tool, lack of depth understanding and dynamic correlation ability, reduce the reliability of the monitoring of tour and inspection task.

Description

Technical Field

[0001] This invention relates to the field of patrol and monitoring technology, and in particular to a monitoring method, system, equipment, medium and product for patrol and inspection tasks. Background Technology

[0002] In the power industry, inspections and audits, as a core supervisory method for identifying problems, preventing risks, and ensuring compliance, play an irreplaceable role in improving corporate governance and mitigating operational risks. However, current inspection and audit work still faces a series of prominent challenges: Inspectors lack sufficient understanding of business operations and policy regulations; report writing and problem characterization lack unified standards, making it difficult to guarantee the depth and standardization of the work. Faced with complex and ever-changing violations, manual review of regulations and policies is inefficient, easily overlooking key clauses and affecting the accuracy of problem judgment. Simultaneously, the lack of domain-specific knowledge bases and intelligent analysis tools results in low utilization rates of historical data, policy documents, and typical cases, leading to insufficient precision and routineness in supervision and inspection. Furthermore, in the rectification supervision phase, the review and implementation of rectification plans lack systematic intelligent support, the tracking and evaluation of rectification effectiveness is limited, and the closed-loop supervision effect is poor. These problems severely restrict the overall effectiveness of inspection and audit work, necessitating the use of digital and intelligent means to build efficient and reliable monitoring methods to support the transformation and upgrading of power industry inspections and audits from "human defense" to "technology defense."

[0003] Currently, existing monitoring methods for inspection and supervision tasks mostly rely on manual ledgers, fixed process management systems, or general text analysis tools. They lack a deep understanding and dynamic correlation capabilities of power industry professional terminology, complex policies and regulations, and historical cases, which reduces the reliability of monitoring inspection and supervision tasks. Summary of the Invention

[0004] This invention provides a monitoring method, system, equipment, medium, and product for inspection and patrol tasks. It solves the technical problem that existing monitoring methods for inspection and patrol tasks rely heavily on manual ledgers, fixed process management systems, or general text analysis tools, which lack a deep understanding and dynamic correlation ability of power industry professional terminology, complex policies and regulations, and historical cases, thus reducing the reliability of inspection and patrol task monitoring.

[0005] The first aspect of this invention provides a monitoring method for patrol and inspection tasks, comprising:

[0006] Multiple training inspection data are acquired, and each training inspection data is preprocessed to obtain the corresponding inspection feature set.

[0007] The patrol feature set is used to fine-tune and train the preset frozen large language model to obtain the corresponding target large language model, wherein the target large language model includes an information processing network and an output network;

[0008] The inspection task text is obtained, and semantic parsing is performed on the inspection task text through a preset inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors.

[0009] The associated feature vectors are monitored for risk using the output network and the relevant evidence vectors to obtain the corresponding monitoring results.

[0010] Optionally, the step of fine-tuning the preset frozen large language model using the patrol feature set to obtain the corresponding target large language model includes:

[0011] The patrol feature set is input into a preset frozen large language model to obtain the corresponding training evidence data;

[0012] Based on a preset weight adjustment function, the sample weights corresponding to the patrol feature set are determined according to the training evidence data.

[0013] The loss function value of the patrol feature set is calculated based on the weights of each sample and the training evidence data.

[0014] When the loss function value is greater than or equal to the preset loss threshold, the incremental parameters of the frozen large language model are adjusted until the loss function value is less than the loss threshold.

[0015] When the loss function value is less than the loss threshold, the target large language model is generated.

[0016] By adopting the above technical solutions and introducing adaptive adjustment of sample weights and incremental parameter optimization mechanisms during the model training phase, the target large language model can be more accurately adapted to the professional tasks of inspection and supervision, effectively improving training efficiency and model reliability. Inputting the inspection feature set into the frozen large language model to generate training evidence data fully utilizes the basic semantic capabilities of the pre-trained frozen large language model, avoiding the high costs and resource consumption of training from scratch. Automatically allocating sample weights based on the weight adjustment function strengthens the training influence of high-value, highly typical, and highly compliant matching samples, improving the model's learning effect on core inspection business features. Simultaneously, calculating the loss function value based on the training evidence data, inspection feature set, and sample weights, and dynamically comparing it with the loss threshold, enables quantitative control and precise iteration of training effects. Incremental parameters are adjusted only when the loss function value is insufficient, minimizing the training parameter scale while maintaining the stability of the main model weights, significantly reducing computing power and storage overhead. When the loss function value meets the requirements, the target large language model is generated, ensuring stable and reliable output and accurate inference, enabling a better understanding of inspection policies, regulations, typical cases, and business logic.

[0017] Optionally, the information processing network includes an embedding layer and an information processing module. The step of semantically parsing the inspection task text using a preset inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors includes:

[0018] Based on the preset number of searches, the inspection task text is used to search the preset inspection knowledge base and match the corresponding relevant evidence vectors.

[0019] The relevant evidence vectors and the inspection task text are vectorized by the embedding layer to obtain the corresponding inspection feature vectors.

[0020] The information processing module performs contextual semantic parsing on the patrol feature vector to obtain the corresponding associated feature vector. The information processing module includes a first freezing module and a second freezing module connected in sequence.

[0021] By adopting the above technical solution and performing a knowledge base search on the inspection task text according to a preset search quantity, relevant evidence vectors highly related to the inspection task can be quickly located, ensuring sufficient and authoritative reasoning basis and avoiding omissions and biases that are prone to occur in manual searches. Through a unified vectorization process of the task text and related evidence vectors at the embedding layer, unstructured text can be converted into standardized feature representations, eliminating format differences and preserving complete business semantics, laying a stable foundation for subsequent in-depth analysis. Furthermore, through an information processing module containing a first and second freezing module, contextual semantic analysis of the inspection feature vectors can be performed, enabling layered completion of basic semantic understanding and business logic matching. While retaining the model's general capabilities, it strengthens the policy alignment, element extraction, and logical association capabilities in the inspection field, ultimately generating accurate and reliable associated feature vectors, providing high-quality input for subsequent risk monitoring.

[0022] Optionally, the step of performing contextual semantic parsing on the patrol feature vector through the information processing module to obtain the corresponding associated feature vector includes:

[0023] The first freezing module performs semantic understanding on the patrol feature vector to obtain the corresponding initial semantic feature vector;

[0024] The second freezing module performs logical matching on the initial semantic feature vector to obtain the corresponding associated feature vector.

[0025] By adopting the above technical solution, the first freezing module performs semantic understanding on the inspection feature vector, which can efficiently extract the core semantics, business elements, and contextual information from the inspection tasks and evidence, and output a stable and reliable initial semantic feature vector, avoiding feature confusion and information loss. Then, the second freezing module performs professional logic matching and policy rule alignment on the initial semantic feature vector, deeply binding the initial semantics with professional knowledge such as inspection regulations, problem characterization, and rectification logic, forming an associated feature vector that can be directly used for decision-making reasoning, thereby reducing computational overhead while improving the professionalism and accuracy of feature expression.

[0026] Optionally, the output network includes a third freezing module and an evidence alignment verification layer. The step of performing risk monitoring on the associated feature vectors through the output network and each of the relevant evidence vectors to obtain the corresponding monitoring results includes:

[0027] The third freezing module performs decision-making reasoning on the associated feature vector to obtain the corresponding semantic feature vector;

[0028] The cosine similarity between the semantic feature vector and each of the relevant evidence vectors is calculated to obtain multiple cosine similarities, and the maximum value among the cosine similarities is selected as the evidence similarity.

[0029] When the similarity of the evidence is greater than or equal to a preset similarity threshold, the semantic feature vector is used as the corresponding monitoring result.

[0030] When the evidence similarity is less than the similarity threshold, the semantic feature vector is corrected by the evidence alignment verification layer based on each relevant evidence vector and the associated feature vector to obtain the corresponding monitoring result.

[0031] By adopting the above technical solution and utilizing the third freezing module to conduct professional decision-making reasoning on the associated feature vectors, stable and reliable semantic feature vectors can be generated based on the inspection business logic, ensuring that risk assessment and problem characterization are in line with domain standards. Furthermore, by calculating the cosine similarity between the semantic feature vectors and each relevant evidence vector, and taking the maximum value as the evidence similarity, the degree of matching between the conclusions and the knowledge base evidence can be quantitatively evaluated, ensuring that the monitoring results have objective basis and auditability. When the evidence similarity reaches the similarity threshold, the monitoring results are directly output, ensuring monitoring efficiency. When the evidence similarity is insufficient, the evidence alignment and verification layer is activated to make decision corrections to the semantic feature vectors, automatically correcting biases and supplementing evidence, avoiding judgment deviations due to insufficient evidence. This ensures that the monitoring results are accurate and based on sufficient evidence, while also meeting the high requirements of inspection and supervision for standardization, rigor, and risk controllability.

[0032] Optionally, the step of performing decision correction on the semantic feature vectors through the evidence alignment verification layer based on each of the relevant evidence vectors and the associated feature vectors to obtain the corresponding monitoring results includes:

[0033] The semantic feature vector and each of the relevant evidence vectors are corrected and guided by the evidence alignment and verification layer to obtain the corresponding correction and guidance vector.

[0034] The modified guiding vector and the associated feature vector are concatenated to obtain a new associated feature vector.

[0035] When the number of times the associated feature vector is concatenated is less than the preset concatenation threshold, the process jumps to the step of performing decision reasoning on the associated feature vector through the third freezing module to obtain the corresponding semantic feature vector.

[0036] When the number of times the associated feature vector is concatenated is greater than or equal to the concatenation threshold, the preset warning signal and the semantic feature vector are used as the corresponding monitoring results.

[0037] By adopting the above technical solution, the evidence alignment and verification layer corrects and guides the semantic feature vector and related evidence vector, accurately locating inference deviations and generating directional correction guidance vectors, making the model output more aligned with the inspection knowledge base. Concatenating the correction guidance vector with the associated feature vector integrates the correction information into the feature space, providing a more accurate feature foundation for the next round of inference. By setting a concatenation threshold to control the number of iterations, infinite loops can be avoided while ensuring the correction effect, balancing monitoring accuracy and execution efficiency. When the number of iterations reaches the threshold but still cannot meet the matching requirements, an early warning signal and semantic feature vector are automatically output, enabling timely alerts and evidence retention for insufficient evidence, high-risk, or abnormal situations. This ensures that the monitoring results are traceable, verifiable, and auditable, improving the monitoring reliability of inspection and supervision tasks.

[0038] A second aspect of the present invention provides a monitoring system for patrol and inspection tasks, comprising:

[0039] The preprocessing module is used to acquire multiple training inspection data, preprocess each training inspection data, and obtain the corresponding inspection feature set.

[0040] The training module is used to fine-tune the preset frozen large language model using the patrol feature set to obtain the corresponding target large language model, wherein the target large language model includes an information processing network and an output network.

[0041] The parsing module is used to acquire the inspection task text, and to perform semantic parsing on the inspection task text through a preset inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors.

[0042] The monitoring module is used to perform risk monitoring on the associated feature vector through the output network and each of the relevant evidence vectors, and obtain the corresponding monitoring results.

[0043] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the monitoring method for patrol and inspection tasks as described above.

[0044] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the monitoring method for patrol and inspection tasks as described above.

[0045] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs a monitoring method for patrol and inspection tasks as described above.

[0046] As can be seen from the above technical solutions, the present invention has the following advantages:

[0047] By acquiring multiple training inspection data sets, preprocessing each set to obtain corresponding inspection feature sets, and then fine-tuning a pre-defined frozen large language model using these feature sets, a corresponding target large language model is obtained. This target model includes an information processing network and an output network. Inspection task text is then acquired, and semantically parsed using a pre-defined inspection knowledge base and the information processing network to obtain associated feature vectors and multiple relevant evidence vectors. The output network and the relevant evidence vectors are then used to monitor the associated feature vectors for risk, yielding corresponding monitoring results. This approach overcomes the technical problem that existing inspection task monitoring methods often rely on manual ledgers, fixed process management systems, or general text analysis tools, lacking a deep understanding and dynamic correlation capability of power industry terminology, complex policies and regulations, and historical cases, thus reducing the reliability of inspection task monitoring. Compared with traditional methods for monitoring inspection and supervision tasks, this invention efficiently fine-tunes a frozen large language model based on inspection feature sets. This significantly reduces training computation and storage costs while enabling the model to accurately learn professional knowledge, policies, regulations, and typical case features of power industry inspections. This avoids the comprehension bias of general models in professional scenarios. Furthermore, by leveraging an inspection knowledge base and information processing network to perform deep semantic analysis on the inspection task text, it can automatically associate and match relevant evidence vectors and generate associated feature vectors. This effectively solves problems such as low efficiency, easy omissions, and inconsistent qualitative analysis in manual review, improving the completeness of the evidence chain and the accuracy of judgment. Finally, by using the output network and various relevant evidence vectors to perform similarity verification and decision correction on the associated feature vectors, the corresponding monitoring results are obtained, thus improving the monitoring reliability of inspection and supervision tasks. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating the steps of a monitoring method for patrol and inspection tasks provided in Embodiment 1 of the present invention;

[0050] Figure 2 This is a flowchart illustrating the steps of a monitoring method for patrol and inspection tasks provided in Embodiment 2 of the present invention.

[0051] Figure 3 This is a structural block diagram of a monitoring system for patrol and inspection tasks provided in Embodiment 3 of the present invention;

[0052] Figure 4This is a structural block diagram of an electronic device provided in Embodiment 4 of the present invention. Detailed Implementation

[0053] This invention provides a monitoring method, system, equipment, medium, and product for inspection and patrol tasks, which addresses the technical problem that existing monitoring methods for inspection and patrol tasks rely heavily on manual ledgers, fixed process management systems, or general text analysis tools, lacking a deep understanding and dynamic correlation ability of power industry professional terminology, complex policies and regulations, and historical cases, thus reducing the reliability of inspection and patrol task monitoring.

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a monitoring method for patrol and inspection tasks provided in Embodiment 1 of the present invention.

[0056] This invention provides a monitoring method for patrol and inspection tasks, comprising:

[0057] Step 101: Obtain multiple training inspection data sets, preprocess each training inspection data set, and obtain the corresponding inspection feature set.

[0058] Training inspection data refers to the original inspection data related to inspections used for model training. It includes various types of business data such as policy and regulatory texts, inspection reports, rectification materials, typical cases, and manual review records. It is the basic material for the model to learn inspection business knowledge and rules.

[0059] Preprocessing refers to a series of operations on the original training and inspection data, such as cleaning, deduplication, standardization, text segmentation, and metadata annotation. The purpose is to remove invalid information, unify the data format, standardize the text expression, and transform the unstructured raw data into structured data that meets the requirements of model training.

[0060] The patrol feature set refers to a standardized dataset containing the core semantics and feature information of patrol operations, extracted from the training patrol data after preprocessing.

[0061] In this embodiment of the invention, multiple training inspection data are acquired, and each training inspection data is cleaned, deduplicated, standardized, and segmented into text. Invalid information and redundant content are removed, and the text format and expression specifications are unified. The data is split into fixed-length text segments according to the semantic integrity of the inspection business. At the same time, metadata information such as source, type, time, and version is labeled for each segment to complete the structured and standardized transformation of the data. Finally, an inspection feature set that can be used for model training is extracted.

[0062] Step 102: Use the patrol feature set to fine-tune the preset frozen large language model to obtain the corresponding target large language model, wherein the target large language model includes an information processing network and an output network.

[0063] In this embodiment of the invention, the patrol feature set is input into a preset frozen large language model to obtain corresponding training evidence data. Based on a preset weight adjustment function, the sample weights corresponding to the patrol feature set are determined using the training evidence data. Each sample weight, the training evidence data, and the patrol feature set are input into a preset loss function to obtain corresponding loss function values. The loss function values ​​are compared with a preset loss threshold. If the loss function value is greater than or equal to the loss threshold, the incremental parameters of the frozen large language model are adjusted until the loss function value is less than the loss threshold. If the loss function value is less than the loss threshold, a target large language model is generated, wherein the target large language model includes an information processing network and an output network.

[0064] Step 103: Obtain the inspection task text, and perform semantic parsing on the inspection task text through a preset inspection knowledge base and information processing network to obtain associated feature vectors and multiple related evidence vectors.

[0065] The inspection task text refers to the inspection and supervision business input text to be parsed, which includes specific business content such as policy consultation, problem verification, report writing, and rectification evaluation.

[0066] The inspection knowledge base refers to a professional knowledge base built for the field of inspection and supervision, which stores vectorized business knowledge such as legal provisions, typical cases, rectification standards, and key points of supervision.

[0067] In this embodiment of the invention, the inspection task text is obtained, and then formatted and anonymized (i.e., information that is not suitable for direct exposure, such as unit names, personnel names, positions, classified information numbers, sensitive projects, financial data, and internal classified clauses, is masked, replaced, or anonymized in the inspection task text to eliminate the risk of sensitive information leakage without destroying the semantics and business logic of the text, and to ensure the compliance and security of data processing). Then, using the inspection task text as the query object, vector retrieval is performed in a preset inspection knowledge base to match relevant evidence vectors highly related to the task content. The inspection task text and each relevant evidence vector are semantically parsed through an information processing network to obtain associated feature vectors.

[0068] Step 104: Perform risk monitoring on the associated feature vectors by outputting the network and various relevant evidence vectors to obtain the corresponding monitoring results.

[0069] In this embodiment of the invention, various relevant evidence vectors and associated feature vectors are input to the output network. Semantic feature vectors are generated by first performing decision reasoning on the associated feature vectors. The cosine similarity between the semantic feature vector and each relevant evidence vector is calculated, resulting in multiple cosine similarities, and the maximum value is selected as the evidence similarity. The evidence similarity is compared with a preset similarity threshold. If the evidence similarity is greater than or equal to the threshold, the semantic feature vector is used as the corresponding monitoring result. If the evidence similarity does not reach the threshold, an evidence alignment verification layer is activated for decision correction and guidance. If multiple matches still fail to meet the threshold, a warning signal and the corresponding semantic feature vector are output as the corresponding monitoring result.

[0070] In this embodiment of the invention, multiple training inspection data are acquired, and each training inspection data is preprocessed to obtain a corresponding inspection feature set. The inspection feature set is then used to fine-tune and train a pre-defined frozen large language model to obtain a corresponding target large language model. The target large language model includes an information processing network and an output network. Inspection task text is acquired, and semantic parsing of the inspection task text is performed through a pre-defined inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors. The associated feature vectors are then monitored for risk using the output network and the various related evidence vectors to obtain corresponding monitoring results. This overcomes the technical problem that existing inspection task monitoring methods rely heavily on manual ledgers, fixed process management systems, or general text analysis tools, lacking a deep understanding and dynamic correlation capability of power industry professional terminology, complex policies and regulations, and historical cases, thus reducing the reliability of inspection task monitoring. Compared with traditional methods for monitoring inspection and supervision tasks, this invention efficiently fine-tunes a frozen large language model based on inspection feature sets. This significantly reduces training computation and storage costs while enabling the model to accurately learn professional knowledge, policies, regulations, and typical case features of power industry inspections. This avoids the comprehension bias of general models in professional scenarios. Furthermore, by leveraging an inspection knowledge base and information processing network to perform deep semantic analysis on the inspection task text, it can automatically associate and match relevant evidence vectors and generate associated feature vectors. This effectively solves problems such as low efficiency, easy omissions, and inconsistent qualitative analysis in manual review, improving the completeness of the evidence chain and the accuracy of judgment. Finally, by using the output network and various relevant evidence vectors to perform similarity verification and decision correction on the associated feature vectors, the corresponding monitoring results are obtained, thus improving the monitoring reliability of inspection and supervision tasks.

[0071] Please see Figure 2 , Figure 2 This is a flowchart illustrating the steps of a monitoring method for patrol and inspection tasks provided in Embodiment 2 of the present invention.

[0072] This invention provides a monitoring method for patrol and inspection tasks, comprising:

[0073] Step 201: Obtain multiple training inspection data sets, preprocess each training inspection data set, and obtain the corresponding inspection feature set.

[0074] In this embodiment of the invention, multiple training inspection data are acquired, and text cleaning, deduplication, format unification and semantic segmentation are carried out on each training inspection data in sequence. Invalid characters are filtered, expression deviations are corrected, and structured conversion and annotation are completed in accordance with the inspection business specifications to form an inspection feature set suitable for model training.

[0075] Step 202: Input the patrol feature set into the preset frozen large language model to obtain the corresponding training evidence data.

[0076] Frozen large language models refer to pre-trained large language models in which the weights of the main model are fixed and only a small number of incremental parameters are allowed to be trained. This preserves the basic semantic capabilities and significantly reduces the computational cost of training.

[0077] Training evidence data refers to standardized text and vector data generated by the frozen large language model based on the inspection feature set through semantic reasoning and knowledge association, covering problem clues, policy basis, qualitative conclusions, and rectification directions, and is used for incremental parameter optimization of the model and calibration of business capabilities.

[0078] In this embodiment of the invention, the inspection feature set is input into a preset frozen large language model. Based on the frozen basic semantic weights and hierarchical adaptation incremental parameters, the inspection feature set is subjected to contextual semantic encoding, professional feature extraction and business logic reasoning. It automatically associates core knowledge such as policy clauses, typical problems and rectification norms in the field of inspection and supervision, and generates training evidence data containing problem clues, policy basis, qualitative references and rectification directions.

[0079] Step 203: Based on the preset weight adjustment function, determine the sample weights corresponding to the patrol feature set according to the training evidence data.

[0080] Sample weights refer to the numerical coefficients assigned to each sample in the inspection feature set. They are used to control the influence of the sample in model training. The higher the weight, the greater the impact on model optimization.

[0081] In this embodiment of the invention, training evidence data and patrol feature set are input into a preset weight adjustment function to obtain the sample weights corresponding to each sample in the patrol feature set.

[0082] It should be noted that the weight adjustment function is as follows:

[0083] ;

[0084] in, Let i be the sample weight of the i-th sample. The first adaptive weight coefficient, The second adaptive weighting coefficient, The third adaptive weighting coefficient, Here, represents the fourth adaptive weight coefficient, and i is the index of the sample. Score the completeness of evidence for the i-th sample. Let represent the number of core inspection elements actually included in the training evidence data corresponding to the i-th sample. Let be the total number of complete elements of inspection evidence in the inspection feature set corresponding to the i-th sample. Score the semantic matching degree of the evidence for the i-th sample. Let be the semantic feature vector of the i-th sample. Let i be the standard semantic feature vector corresponding to the i-th sample. Score the business compliance fit of the i-th sample. The number of inspection and supervision regulations, policies, and normative clauses matched in the i-th sample. Let i be the total number of clauses involved in the i-th sample. To score the business typicality of the i-th sample, Let i be the frequency of occurrence of the most frequent problem corresponding to the i-th sample. Let be the frequency of occurrence of the problem type corresponding to the i-th sample in the historical inspection data.

[0085] Step 204: Calculate the loss function value of the patrol feature set based on the weights of each sample and the training evidence data.

[0086] In this embodiment of the invention, the weights of each sample, training evidence data, and patrol feature set are input into a preset loss function to obtain the corresponding loss function value.

[0087] It should be noted that the loss function is as follows:

[0088] ;

[0089] in, The value of the loss function. The total number of samples, Let the training evidence data labels be the i-th sample. This is the predicted probability output by the model for the i-th sample.

[0090] Step 205: When the loss function value is greater than or equal to the preset loss threshold, adjust the incremental parameters of the frozen large language model until the loss function value is less than the loss threshold.

[0091] The loss threshold refers to the critical loss value set in advance according to the accuracy requirements of the inspection and patrol task, which is used to determine whether the model training meets the standards.

[0092] Incremental parameters refer to the first, second, and third incremental parameters injected into the model via LoRA. The first incremental parameter is a small-rank LoRA (r=4), the second incremental parameter is a medium-rank LoRA (r=8), and the third incremental parameter is a large-rank LoRA (r=8).

[0093] In this embodiment of the invention, when the loss function value is greater than or equal to the preset loss threshold, it indicates that the inference output of the current frozen large language model still has a large deviation from the standard training evidence data in the inspection feature set, which cannot meet the accuracy requirements of the inspection business. Based on the backpropagation algorithm, the incremental parameters of the frozen large language model are adaptively iteratively updated. Under the premise of keeping the main model weight completely frozen, the error between the prediction result and the standard evidence is gradually reduced. After each parameter update, the loss function value is recalculated and compared with the loss threshold. The optimization is continuously iterated until the loss function value is less than the preset loss threshold.

[0094] It should be noted that the backpropagation algorithm refers to the core machine learning algorithm used to calculate the gradient of parameters based on the loss function value and guide the optimization and updating of model parameters.

[0095] Step 206: When the loss function value is less than the loss threshold, a target large language model is generated, which includes an information processing network and an output network.

[0096] In this embodiment of the invention, when the loss function value is less than the loss threshold, it indicates that the model has completed the accurate learning of the characteristics of inspection and supervision business, policies and regulations and evidence logic, and the inference output meets the requirements of compliance, accuracy and typicality. The parameter iteration is stopped and the currently trained model is saved to generate a target large language model for inspection and supervision monitoring tasks.

[0097] It's worth noting that once the target large language model is obtained, optimizations such as quantization and distillation can be performed as needed to reduce inference latency and resource consumption. Quantization refers to converting high-precision floating-point parameters (e.g., FP16 / FP32) in the large language model into low-precision integers (e.g., INT8 / INT4) through linear mapping, clustering, or dynamic scaling. This model compression and acceleration technique can significantly reduce memory usage, increase computation speed, and reduce hardware power consumption with almost no loss of inference accuracy, making it suitable for deployment on edge nodes and in low-computing-power environments. Distillation transfers the knowledge and inference logic learned by the large, high-performing teacher model to a smaller, faster student model through knowledge transfer training. This significantly reduces the number of model parameters and computational load while allowing the lightweight student model to retain near-large model performance and generalization ability, achieving efficient and lightweight deployment through model compression and optimization.

[0098] It should be noted that by introducing adaptive adjustment of sample weights and incremental parameter optimization mechanisms during the model training phase, the target large language model can be more accurately adapted to the professional tasks of inspection and supervision, effectively improving training efficiency and model reliability. Inputting the inspection feature set into the frozen large language model to generate training evidence data fully utilizes the basic semantic capabilities of the pre-trained frozen large language model, avoiding the high costs and resource consumption of training from scratch. Automatically allocating sample weights based on the weight adjustment function strengthens the training influence of high-value, highly typical, and highly compliant matching samples, improving the model's learning effect on core inspection business features. Simultaneously, calculating the loss function value based on the training evidence data, inspection feature set, and sample weights, and dynamically comparing it with the loss threshold, enables quantitative control and precise iteration of training effects. Incremental parameters are adjusted only when the loss function value is insufficient, minimizing the training parameter scale while maintaining the stability of the main model weights, significantly reducing computing power and storage overhead. When the loss function value meets the requirements, the target large language model is generated, ensuring stable and reliable output and accurate inference, enabling a better understanding of inspection policies, regulations, typical cases, and business logic.

[0099] Step 207: Obtain the inspection task text, and perform semantic parsing on the inspection task text through a preset inspection knowledge base and information processing network to obtain associated feature vectors and multiple related evidence vectors.

[0100] Furthermore, the information processing network includes an embedding layer and an information processing module, and step 207 includes the following sub-steps:

[0101] S11. Based on the preset number of searches, the inspection task text is used to search the preset inspection knowledge base and match the corresponding relevant evidence vectors.

[0102] The number of searches refers to the maximum number of similar evidence vectors that need to be returned in the inspection knowledge base, which is preset based on the actual needs of the inspection business, computing resources, and matching accuracy requirements.

[0103] Relevant evidence vectors refer to high-dimensional numerical vectors corresponding to evidence data that are highly similar to the inspection task text in terms of semantic content, business logic, and policy norms, and are matched from the inspection knowledge base through semantic retrieval technology. They are used to characterize the semantic features of the evidence.

[0104] In this embodiment of the invention, the inspection task text is semantically vectorized to generate a query vector with contextual semantic features. The query vector is then used to perform a high-dimensional semantic similarity search in a pre-defined inspection knowledge base. The results are sorted from highest to lowest similarity score, and a pre-defined number of optimal results are selected. This allows for precise matching of relevant evidence vectors that are highly correlated with the inspection task in terms of policy basis, problem characterization, and rectification standards.

[0105] S12. Vectorize the relevant evidence vectors and inspection task texts through the embedding layer to obtain the corresponding inspection feature vectors.

[0106] The embedding layer refers to the network layer in the target large language model that is responsible for mapping discrete data such as text and symbols into continuous dense vectors, and is used to unify semantic representation.

[0107] In this embodiment of the invention, the relevant evidence vectors and inspection task texts are vectorized by the embedding layer, and the relevant evidence vectors and inspection task texts are uniformly mapped to dense numerical features of the same dimension and computable. This completes the standardized conversion from unstructured text to structured vectors, eliminates text format differences and retains complete business semantics, and finally obtains an inspection feature vector that integrates task semantics and evidence features.

[0108] S13. The information processing module performs contextual semantic parsing on the patrol feature vector to obtain the corresponding associated feature vector. The information processing module includes a first freezing module and a second freezing module connected in sequence.

[0109] Furthermore, S13 includes the following sub-steps:

[0110] S131. The first freezing module performs semantic understanding on the patrol feature vector to obtain the corresponding initial semantic feature vector.

[0111] The initial semantic feature vector refers to the basic semantic representation vector output by the first freezing module, which carries the core semantic information of the task and evidence.

[0112] In this embodiment of the invention, the first freezing module performs layer-by-layer semantic encoding and contextual understanding on the inspection feature vector. By extracting the basic semantics, sentence structure and core business elements of the inspection task text and evidence vector, redundant features are filtered out and key semantic information is retained to obtain an initial semantic feature vector with global semantic representation capabilities.

[0113] It should be noted that the first freezing module comprises ten sequentially connected first freezing blocks. Each first freezing block is a TransformerBlock (i.e., the encoder block of the Transformer) that has had its first incremental parameter injected into the key linear mapping layer. The key linear mapping layer refers to the core linear layer in the Transformer architecture responsible for performing Query, Key, and Value projections, as well as the output transformation of the feedforward network. Specifically, it includes the Q / K / V projection layer, the output linear layer, and the upper and lower linear layers of the feedforward network FFN. It is the core computational unit for semantic feature transformation, attention calculation, and high-level inference in large language models.

[0114] S132. Logically match the initial semantic feature vector through the second freezing module to obtain the corresponding associated feature vector.

[0115] The associated feature vector refers to a structured feature vector that integrates semantic information and business logic and can be directly used to output network inference decisions.

[0116] In this embodiment of the invention, the initial semantic feature vector is inspected by the second freezing module to match business logic and align with policy rules. Based on the solidified domain knowledge, the basic semantic features are associated with clauses, classified as issues, and verified for compliance. The scattered semantic information is integrated into structured features with business logic association, resulting in an associated feature vector that can be directly used for risk decision-making.

[0117] It should be noted that the second freeze module includes ten sequentially connected second freeze blocks, wherein the second freeze block is a TransformerBlock that has injected a second incremental parameter into the key linear mapping layer.

[0118] It should be noted that by performing semantic understanding on the inspection feature vector through the first freezing module, the core semantics, business elements, and contextual information in the inspection tasks and evidence can be efficiently extracted, outputting a stable and reliable initial semantic feature vector, avoiding feature confusion and information loss. Then, the second freezing module performs professional logic matching and policy rule alignment on the initial semantic feature vector, deeply binding the initial semantics with professional knowledge such as inspection regulations, problem characterization, and rectification logic, forming an associated feature vector that can be directly used for decision-making reasoning. This reduces computational overhead while improving the professionalism and accuracy of feature expression.

[0119] It should be noted that, by performing a knowledge base search on the inspection task text according to a preset search volume, the system can quickly locate relevant evidence vectors highly related to the inspection task, ensuring sufficient and authoritative reasoning and avoiding omissions and biases that are prone to occur in manual searches. Through a unified vectorization process using an embedding layer, the task text and related evidence vectors are converted from unstructured text into standardized feature representations, eliminating format differences and preserving complete business semantics, laying a stable foundation for subsequent in-depth analysis. Furthermore, the information processing module, which includes a first and second freezing module, performs contextual semantic analysis on the inspection feature vectors, enabling layered completion of basic semantic understanding and business logic matching. While retaining the model's general capabilities, it strengthens policy alignment, element extraction, and logical association capabilities in the inspection field, ultimately generating accurate and reliable associated feature vectors, providing high-quality input for subsequent risk monitoring.

[0120] Step 208: Perform risk monitoring on the associated feature vectors by outputting the network and various relevant evidence vectors to obtain the corresponding monitoring results.

[0121] Furthermore, the output network includes a third freeze module and an evidence alignment verification layer, and step 208 includes the following sub-steps:

[0122] S21. The third freezing module performs decision reasoning on the associated feature vectors to obtain the corresponding semantic feature vectors.

[0123] A semantic feature vector is a decision-level feature vector generated after decision-making reasoning, which contains the final judgment conclusion and risk information.

[0124] In this embodiment of the invention, the third freezing module performs decision reasoning on the associated feature vectors, completes the problem characterization, compliance judgment and risk level assessment based on the fixed business rules and evidence logic, and transforms the business-related features into decision-level semantic features that can be directly output, thereby obtaining the corresponding semantic feature vectors.

[0125] It should be noted that the third freeze module includes ten sequentially connected third freeze blocks, where each third freeze block is a TransformerBlock that has injected a third incremental parameter into the key linear mapping layer.

[0126] S22. Calculate the cosine similarity between the semantic feature vector and each relevant evidence vector to obtain multiple cosine similarities, and select the maximum value among the cosine similarities as the evidence similarity.

[0127] Evidence similarity refers to the maximum value selected from the cosine similarity of multiple matching pieces of evidence, representing the degree of matching between the model's output conclusion and the best evidence.

[0128] In this embodiment of the invention, multiple cosine similarities are obtained by calculating the cosine similarity between the semantic feature vector and each relevant evidence vector one by one, and the maximum value among the cosine similarities is selected as the evidence similarity.

[0129] S23. When the evidence similarity is greater than or equal to the preset similarity threshold, the semantic feature vector will be used as the corresponding monitoring result.

[0130] The similarity threshold refers to the maximum cosine similarity value calculated between the semantic feature vector and the relevant evidence vector, which is used to measure the degree of matching between the conclusion and the knowledge base.

[0131] In this embodiment of the invention, when the evidence similarity is greater than or equal to a preset similarity threshold, it indicates that the semantic feature vector generated by the model is highly matched with the optimal evidence in the inspection knowledge base, the conclusion is compliant, the basis is sufficient, and it meets the accuracy requirements of the inspection business. The semantic feature vector is then used as the corresponding monitoring result.

[0132] S24. When the evidence similarity is less than the similarity threshold, the semantic feature vector is corrected by the evidence alignment verification layer based on each relevant evidence vector and associated feature vector to obtain the corresponding monitoring result.

[0133] Furthermore, S24 includes the following sub-steps:

[0134] S241. The semantic feature vector and each relevant evidence vector are corrected and guided through the evidence alignment and verification layer to obtain the corresponding corrected guidance vector.

[0135] The revised guiding vector refers to the guiding feature vector output by the evidence alignment and verification layer, which is used to correct semantic features and make the conclusion more consistent with the evidence.

[0136] In this embodiment of the invention, the mean of each relevant evidence vector is compared with the semantic feature vector to obtain the corresponding deviation vector. An evidence alignment verification layer then performs feature extraction, deviation correction, and evidence alignment processing on the deviation vector, outputting a correction guidance vector that guides the semantic features toward the optimal evidence direction, improving the compliance and matching degree of the conclusion.

[0137] It should be noted that the evidence alignment verification layer consists of a first fully connected layer (which is a fully connected layer with embedded layer normalization (i.e., LayerNorm) and GELU activation function), a Dropout layer, and a fully connected layer connected in sequence. The first fully connected layer is used to normalize and stabilize the bias vector and perform nonlinear feature transformation to extract effective feature components related to evidence alignment from the bias. The Dropout layer randomly suppresses the response of some neurons during feature transmission to avoid the model overfitting to local bias features. Finally, the fully connected layer maps the corrected bias features to a correction guide vector of a specified dimension.

[0138] S242. The modified guiding vector and the associated feature vector are concatenated to obtain a new associated feature vector.

[0139] In this embodiment of the invention, the modified guiding vector and the associated feature vector are channel-joined and fused along the feature dimension to obtain a new associated feature vector.

[0140] It should be noted that the corrected guiding vector and the associated feature vector are concatenated to obtain a new associated feature vector. This newly generated associated feature vector simultaneously carries the original business logic features and evidence alignment correction information, further enhancing the fit between the features and the evidence in the inspection knowledge base.

[0141] S243. When the number of times the associated feature vector is concatenated is less than the preset concatenation threshold, the process jumps to the step of performing decision reasoning on the associated feature vector through the third freezing module to obtain the corresponding semantic feature vector.

[0142] The number of splicing operations refers to the cumulative number of times the associated feature vector and the modified guiding vector are spliced ​​and fused along the feature dimension. It is used to control the number of loop iterations for decision correction.

[0143] The splicing threshold refers to the maximum number of splicing operations preset according to the monitoring accuracy requirements of the inspection task and the inference efficiency of the model. It is usually set to 2 or 3.

[0144] In this embodiment of the invention, when the number of times the associated feature vector is concatenated is less than a preset concatenation threshold, it indicates that there is still room for optimization in the matching degree between the current semantic feature vector and the relevant evidence vector. Based on the corrected associated feature vector, the process can continue to jump to the step of performing decision reasoning on the associated feature vector through the third freezing module to obtain the corresponding semantic feature vector. Decision reasoning and evidence similarity verification are then performed again until the evidence similarity meets the standard or reaches the concatenation threshold, thereby ensuring the accuracy and reliability of the monitoring results.

[0145] S244. When the number of times the associated feature vector is spliced ​​is greater than or equal to the splicing threshold, the preset warning signal and evidence vector will be used as the corresponding monitoring result.

[0146] Warning signals refer to the warning signs output by the model when the risk is unclear, the evidence is insufficient, or the risk exceeds the normal reasoning range.

[0147] In this embodiment of the invention, when the number of times the associated feature vector is concatenated is greater than or equal to the concatenation threshold, it indicates that after multiple iterations and corrections, the semantic feature vector generated by the model reasoning still cannot reach the preset matching standard with the relevant evidence vector in the inspection knowledge base. At this time, it is determined that the current inspection task has a high risk or insufficient basis, and the preset warning signal and the corresponding semantic feature vector are directly output as the final monitoring result to achieve timely risk warning and evidence retention, and ensure the rigor and traceability of the inspection and monitoring work.

[0148] It is worth mentioning that when the monitoring results contain warning signals, the monitoring results can be manually reviewed by technicians.

[0149] It should be noted that by utilizing the evidence alignment and verification layer to correct and guide the semantic feature vector and related evidence vector, reasoning deviations can be accurately located and directional correction guidance vectors can be generated, making the model output more aligned with the inspection knowledge base. Concatenating the correction guidance vector with the associated feature vector integrates the correction information into the feature space, providing a more accurate feature foundation for the next round of reasoning. By setting a concatenation threshold to control the number of iterations, infinite loops can be avoided while ensuring the correction effect, balancing monitoring accuracy and execution efficiency. When the number of iterations reaches the threshold but still cannot meet the matching requirements, an early warning signal and semantic feature vector are automatically output, enabling timely alerts and evidence retention for insufficient evidence, high-risk, or abnormal situations. This ensures that the monitoring results are traceable, verifiable, and auditable, improving the monitoring reliability of inspection and supervision tasks.

[0150] It should be noted that by utilizing the third freezing module to conduct professional decision-making reasoning on the associated feature vectors, stable and reliable semantic feature vectors can be generated based on the inspection business logic, ensuring that risk assessment and problem characterization align with domain standards. Furthermore, by calculating the cosine similarity between the semantic feature vectors and each relevant evidence vector, and taking the maximum value as the evidence similarity, the degree of matching between the conclusions and the knowledge base evidence can be quantitatively evaluated, ensuring that the monitoring results have objective basis and auditability. When the evidence similarity reaches the similarity threshold, the monitoring results are directly output, ensuring monitoring efficiency. When the evidence similarity is insufficient, the evidence alignment and verification layer is activated to correct the semantic feature vectors, automatically correcting biases and supplementing evidence to avoid judgment deviations due to insufficient evidence. This ensures both accurate monitoring results and sufficient evidence, while also meeting the high requirements of inspection and supervision for standardization, rigor, and risk controllability.

[0151] In this embodiment of the invention, multiple training inspection data are acquired, and each training inspection data is preprocessed to obtain a corresponding inspection feature set. The inspection feature set is then used to fine-tune and train a pre-defined frozen large language model to obtain a corresponding target large language model. The target large language model includes an information processing network and an output network. Inspection task text is acquired, and semantic parsing of the inspection task text is performed through a pre-defined inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors. The associated feature vectors are then monitored for risk using the output network and the various related evidence vectors to obtain corresponding monitoring results. This overcomes the technical problem that existing inspection task monitoring methods rely heavily on manual ledgers, fixed process management systems, or general text analysis tools, lacking a deep understanding and dynamic correlation capability of power industry professional terminology, complex policies and regulations, and historical cases, thus reducing the reliability of inspection task monitoring. Compared with traditional methods for monitoring inspection and supervision tasks, this invention efficiently fine-tunes a frozen large language model based on inspection feature sets. This significantly reduces training computation and storage costs while enabling the model to accurately learn professional knowledge, policies, regulations, and typical case features of power industry inspections. This avoids the comprehension bias of general models in professional scenarios. Furthermore, by leveraging an inspection knowledge base and information processing network to perform deep semantic analysis on the inspection task text, it can automatically associate and match relevant evidence vectors and generate associated feature vectors. This effectively solves problems such as low efficiency, easy omissions, and inconsistent qualitative analysis in manual review, improving the completeness of the evidence chain and the accuracy of judgment. Finally, by using the output network and various relevant evidence vectors to perform similarity verification and decision correction on the associated feature vectors, the corresponding monitoring results are obtained, thus improving the monitoring reliability of inspection and supervision tasks.

[0152] Please see Figure 3 , Figure 3 This is a structural block diagram of a monitoring system for patrol and inspection tasks provided in Embodiment 3 of the present invention.

[0153] This invention provides a monitoring system for patrol and inspection tasks, comprising:

[0154] The preprocessing module 301 is used to acquire multiple training inspection data, preprocess each training inspection data, and obtain the corresponding inspection feature set.

[0155] Training module 302 is used to fine-tune the preset frozen large language model using the patrol feature set to obtain the corresponding target large language model, wherein the target large language model includes an information processing network and an output network.

[0156] The parsing module 303 is used to obtain the inspection task text, and to perform semantic parsing on the inspection task text through a preset inspection knowledge base and information processing network to obtain the associated feature vector and multiple related evidence vectors.

[0157] The monitoring module 304 is used to perform risk monitoring on the associated feature vectors by outputting the network and various relevant evidence vectors, and obtain the corresponding monitoring results.

[0158] Furthermore, training module 302 includes:

[0159] The training submodule is used to input the patrol feature set into the preset frozen large language model to obtain the corresponding training evidence data.

[0160] The first adjustment submodule is used to determine the sample weights corresponding to the patrol feature set based on the training evidence data according to the preset weight adjustment function.

[0161] The loss function value of the patrol feature set is calculated based on the weights of each sample and the training evidence data;

[0162] The second adjustment submodule is used to adjust the incremental parameters of the frozen large language model when the loss function value is greater than or equal to the preset loss threshold, until the loss function value is less than the loss threshold.

[0163] When the loss function value is less than the loss threshold, the target large language model is generated.

[0164] Furthermore, the information processing network includes an embedding layer and an information processing module, and the parsing module 303 includes:

[0165] The matching submodule is used to search the pre-defined inspection knowledge base using the inspection task text based on a preset number of searches, and match the corresponding relevant evidence vectors.

[0166] The vectorization submodule is used to vectorize the relevant evidence vectors and inspection task texts through the embedding layer to obtain the corresponding inspection feature vectors.

[0167] The semantic parsing submodule is used to perform contextual semantic parsing on the patrol feature vector through the information processing module to obtain the corresponding associated feature vector. The information processing module includes a first freezing module and a second freezing module connected in sequence.

[0168] Furthermore, the semantic parsing submodule includes:

[0169] The first freezing unit is used to perform semantic understanding on the patrol feature vector through the first freezing module to obtain the corresponding initial semantic feature vector;

[0170] The second freezing unit is used to perform logical matching on the initial semantic feature vector through the second freezing module to obtain the corresponding associated feature vector.

[0171] Furthermore, the output network includes a third freezing module and an evidence alignment verification layer, and the monitoring module 304 includes:

[0172] The reasoning submodule is used to perform decision reasoning on the associated feature vectors through the third freezing module to obtain the corresponding semantic feature vectors;

[0173] The similarity submodule is used to calculate the cosine similarity between the semantic feature vector and each relevant evidence vector, obtain multiple cosine similarities, and select the maximum value among the cosine similarities as the evidence similarity.

[0174] The first analysis submodule is used to take the semantic feature vector as the corresponding monitoring result when the evidence similarity is greater than or equal to the preset similarity threshold.

[0175] The second analysis submodule is used to make decision corrections to the semantic feature vectors based on each relevant evidence vector and associated feature vector through the evidence alignment verification layer when the evidence similarity is less than the similarity threshold, so as to obtain the corresponding monitoring results.

[0176] Furthermore, the second analysis submodule includes:

[0177] The correction unit is used to correct and guide the semantic feature vector and each relevant evidence vector through the evidence alignment and verification layer to obtain the corresponding correction and guidance vector.

[0178] The concatenation unit is used to concatenate the modified guiding vector and the associated feature vector to obtain a new associated feature vector;

[0179] The jump unit is used to jump to the step of performing decision reasoning on the associated feature vector through the third freezing module to obtain the corresponding semantic feature vector when the number of times the associated feature vector is concatenated is less than the preset concatenation threshold.

[0180] When the number of times the associated feature vector is concatenated is greater than or equal to the concatenation threshold, the preset warning signal and semantic feature vector will be used as the corresponding monitoring result.

[0181] Please see Figure 4 , Figure 4 This is a structural block diagram of an electronic device provided in Embodiment 4 of the present invention.

[0182] An electronic device according to an embodiment of the present invention includes: a memory 401 and a processor 402. The memory 401 stores a computer program. When the computer program is executed by the processor 402, the processor 402 performs a monitoring method for patrol and inspection tasks as described in any of the above embodiments.

[0183] Memory 401 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 401 has storage space 403 for program code 413 for performing any of the method steps described above. For example, storage space 403 for program code may include individual program codes 413 for implementing the various steps in the methods described above. This program code may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. This program code may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When this code is run by a computing device, it causes the computing device to perform the various steps in the monitoring method for the inspection and patrol task described above.

[0184] Embodiment 5 of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a monitoring method for patrol and inspection tasks as described in any of the above embodiments.

[0185] Embodiment 6 of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs a monitoring method for patrol and inspection tasks as described in any of the above embodiments.

[0186] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0187] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0188] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0189] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0190] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A monitoring method for patrol and inspection tasks, characterized in that, include: Multiple training inspection data are acquired, and each training inspection data is preprocessed to obtain the corresponding inspection feature set. The patrol feature set is used to fine-tune and train the preset frozen large language model to obtain the corresponding target large language model, wherein the target large language model includes an information processing network and an output network; The inspection task text is obtained, and semantic parsing is performed on the inspection task text through a preset inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors. The associated feature vectors are monitored for risk using the output network and the relevant evidence vectors to obtain the corresponding monitoring results.

2. The monitoring method for patrol and inspection tasks according to claim 1, characterized in that, The step of fine-tuning and training the preset frozen large language model using the patrol feature set to obtain the corresponding target large language model includes: The patrol feature set is input into a preset frozen large language model to obtain the corresponding training evidence data; Based on a preset weight adjustment function, the sample weights corresponding to the patrol feature set are determined according to the training evidence data. The loss function value of the patrol feature set is calculated based on the weights of each sample and the training evidence data. When the loss function value is greater than or equal to the preset loss threshold, the incremental parameters of the frozen large language model are adjusted until the loss function value is less than the loss threshold. When the loss function value is less than the loss threshold, the target large language model is generated.

3. The monitoring method for patrol and inspection tasks according to claim 1, characterized in that, The information processing network includes an embedding layer and an information processing module. The step of semantically parsing the inspection task text using a preset inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors includes: Based on the preset number of searches, the inspection task text is used to search the preset inspection knowledge base and match the corresponding relevant evidence vectors. The relevant evidence vectors and the inspection task text are vectorized by the embedding layer to obtain the corresponding inspection feature vectors. The information processing module performs contextual semantic parsing on the patrol feature vector to obtain the corresponding associated feature vector. The information processing module includes a first freezing module and a second freezing module connected in sequence.

4. The monitoring method for patrol and inspection tasks according to claim 3, characterized in that, The step of performing contextual semantic parsing on the patrol feature vector through the information processing module to obtain the corresponding associated feature vector includes: The first freezing module performs semantic understanding on the patrol feature vector to obtain the corresponding initial semantic feature vector; The second freezing module performs logical matching on the initial semantic feature vector to obtain the corresponding associated feature vector.

5. The monitoring method for patrol and inspection tasks according to claim 1, characterized in that, The output network includes a third freezing module and an evidence alignment and verification layer. The step of performing risk monitoring on the associated feature vectors through the output network and each of the relevant evidence vectors to obtain the corresponding monitoring results includes: The third freezing module performs decision-making reasoning on the associated feature vector to obtain the corresponding semantic feature vector; The cosine similarity between the semantic feature vector and each of the relevant evidence vectors is calculated to obtain multiple cosine similarities, and the maximum value among the cosine similarities is selected as the evidence similarity. When the similarity of the evidence is greater than or equal to a preset similarity threshold, the semantic feature vector is used as the corresponding monitoring result. When the evidence similarity is less than the similarity threshold, the semantic feature vector is corrected by the evidence alignment verification layer based on each relevant evidence vector and the associated feature vector to obtain the corresponding monitoring result.

6. The monitoring method for patrol and inspection tasks according to claim 5, characterized in that, The step of performing decision correction on the semantic feature vectors through the evidence alignment verification layer based on each of the relevant evidence vectors and the associated feature vectors to obtain the corresponding monitoring results includes: The semantic feature vector and each of the relevant evidence vectors are corrected and guided by the evidence alignment and verification layer to obtain the corresponding correction and guidance vector. The modified guiding vector and the associated feature vector are concatenated to obtain a new associated feature vector. When the number of times the associated feature vector is concatenated is less than the preset concatenation threshold, the process jumps to the step of performing decision reasoning on the associated feature vector through the third freezing module to obtain the corresponding semantic feature vector. When the number of times the associated feature vector is concatenated is greater than or equal to the concatenation threshold, the preset warning signal and the semantic feature vector are used as the corresponding monitoring results.

7. A monitoring system for patrol and inspection tasks, characterized in that, include: The preprocessing module is used to acquire multiple training inspection data, preprocess each training inspection data, and obtain the corresponding inspection feature set. The training module is used to fine-tune the preset frozen large language model using the patrol feature set to obtain the corresponding target large language model, wherein the target large language model includes an information processing network and an output network. The parsing module is used to acquire the inspection task text, and to perform semantic parsing on the inspection task text through a preset inspection knowledge base and the information processing network to obtain associated feature vectors and multiple related evidence vectors. The monitoring module is used to perform risk monitoring on the associated feature vector through the output network and each of the relevant evidence vectors, and obtain the corresponding monitoring results.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the monitoring method for patrol and inspection tasks as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the monitoring method for patrol and inspection tasks as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the monitoring method for the patrol and inspection task as described in any one of claims 1-6.

Images