Operation risk identification analysis method, storage medium and system

Abstract

The invention discloses an operation risk identification and analysis method, a storage medium and a system, which realize intelligent identification of operation risks, mainly realize intelligent monitoring and intelligent analysis of unsafe behaviors of personnel, and realize qualitative and quantitative analysis of the operation risks. The intelligent monitoring of the unsafe behaviors of the personnel mainly realizes a video AI identification function through a camera and edge calculation; according to the intelligent analysis, a large amount of data generated in the using process is analyzed, such as operation trend analysis, operation part risk statistical analysis and operation violation analysis, an intelligent system automatically recognizes risks existing in the operation process for identification, intelligent early warning and alarm monitoring are conducted on the operation process, and the operation safety is improved. The occurrence rate of various safety accidents in the field operation process is reduced, and safe, stable, long-term, full and excellent operation of enterprises is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, data representation and prediction technology, and relates to a method for establishing, evaluating and predicting an operation risk identification model, and in particular to an operation risk identification and analysis method, storage medium and system. Background Art

[0002] The petrochemical industry has the characteristics of high temperature and high pressure, flammable and explosive, toxic and harmful, continuous operation, and long and wide chain. For refining enterprises, the tasks of engineering construction and inspection and maintenance are very heavy, which concentrates the potential safety risks of construction and production equipment operation. According to the statistical analysis of accidents in the domestic petrochemical industry, accidents occurred in the construction stage and inspection and maintenance sites accounted for more than 46%, and the task of safety supervision and management is very heavy. The group company summarized the industry experience and aimed to build a HSE management system with Chinese petroleum characteristics. It has successively issued more than dozens of new HSE standards, focusing on the formulation of detailed safety regulations and requirements for various unconventional operations, focusing on accident prevention and standardizing on-site safety operation behaviors. According to statistics, accidents in special operation links are frequent, mainly due to the inadequate implementation of the special operation management system of the enterprise, unclear identification of operation risks, and inadequate risk control during the operation process.

[0003] The application of large models has become very common and new large models are constantly being developed, which has promoted the continuous advancement of AI technology and the continuous expansion of application scenarios. This has led to the widespread use of large models not only in the field of natural language processing, but also in computer vision, speech recognition, operational risk identification and other fields. The present invention relies on large computing power and big data to process massive amounts of data and information and realize a variety of complex computing tasks, such as artificial intelligence, scientific simulation and digital twins.

[0004] The so-called big computing power refers to large-scale computing capabilities, and the so-called big data refers to the fact that the global data growth in the past five years has almost doubled every year, providing more support for the continuous development and application of large operational risk models.

[0005] CN115099586A discloses a method and device for identifying operational risks. By establishing a model for risk identification, it is possible to generate items for risk identification corresponding to different operation tickets, thereby improving the pertinence of risk items, and then based on the risk items, the accuracy of risk assessment can be improved, solving the technical problems of high error rate, low efficiency, and lack of pertinence caused by manual or fixed template-specified methods in related technologies. The specific technical means adopted include: after obtaining historical operation ticket data, the operation ticket data is cleaned according to preset data cleaning rules; the cleaned operation ticket data is preprocessed to extract feature content based on the preprocessed data; the extracted feature content is grouped and clustered to determine the items for risk prediction based on the data of each group after clustering, and a model for identifying operational risks is obtained.

[0006] CN116757467A discloses a method for identifying and evaluating the hidden danger risks in the production and operation of marine oil and gas fields. The invention belongs to the field of petroleum engineering, and specifically relates to a method for identifying and evaluating the hidden danger risks in the production and operation of marine oil and gas fields. The method for identifying and evaluating the hidden danger risks in the production and operation of marine oil and gas fields comprises five steps: analyzing and determining the risk characteristic item system of hidden danger data, extracting hidden danger data characteristic items, classifying hidden danger data characteristic items, determining the risk value calculation characteristic items of hidden danger data, and calculating the risk value. Among them, the extraction of hidden danger data characteristic items is divided into four steps: labeling hidden danger characteristic items, establishing characteristic item extraction model, model training, and data prediction. The classification of hidden danger data characteristic items is divided into five steps: establishing the classification system of hidden danger data characteristic items, labeling hidden danger characteristic item categories, establishing hidden danger characteristic item classification model, training hidden danger characteristic item classification model, and predicting hidden danger characteristic item classification. The determination of the risk value calculation characteristic items of hidden danger data is divided into two steps: calculating the occurrence probability characteristic items and calculating the consequence evaluation characteristic items.

[0007] CN117709715A discloses a tunnel engineering construction risk assessment method, system, terminal and medium. The tunnel construction parameters are obtained and preprocessed; the pre-trained risk assessment neural network model is retrieved, and the pre-processed tunnel construction parameters are input into the risk assessment neural network model to obtain the risk assessment result; the pre-trained GPT large language model is retrieved, and the risk assessment result is input into the GPT large language model to obtain a reference risk solution; the risk assessment result and the risk solution are displayed on the human-computer interaction interface. The present invention evaluates the risk through the neural network model based on the construction parameters, and the GPT large language model gives a reference solution according to the evaluation result. The evaluation efficiency is high, and all construction parameters are considered to improve the evaluation accuracy. At the same time, a reference solution is given in time to guide on-site construction, reduce the impact on construction, and improve construction efficiency.

[0008] By analyzing the above-mentioned existing technologies, it can be seen that their shortcomings include: insufficient dynamics of data processing, insufficient accuracy and efficiency of model tuning, imperfect detailed processing of risk levels, poor adaptability to cross-industry applications, and application limitations of large language models in different scenarios. This may lead to insufficient accuracy and efficiency of risk identification in rapidly changing or complex operating environments, making it difficult to meet the needs of multiple industries and scenarios. Summary of the invention

[0009] A method for identifying and analyzing operational risks comprises the following steps: establishing an operational risk database; training an established operational risk identification model based on the operational risk database, wherein the operational risk identification model is established based on a neural network and a Transformer; obtaining training results of the operational risk identification model and evaluating the training results; obtaining evaluation results of the training results, and tuning and verifying the operational risk identification model based on the evaluation results; applying the tuned and verified operational risk identification model to operational risk identification, predicting risk behaviors of staff members, and providing early warning prompts for risk behaviors.

[0010] Furthermore, the process of establishing an operation risk database is as follows: obtaining an operator violation behavior data set, the operator violation behavior data set including video data, image data and text data; performing data cleaning and preprocessing on the operator violation behavior data set, and performing data integration and conversion on the operator violation behavior data set after data cleaning and preprocessing; and assembling the integrated and converted operator violation behavior data set into an operation risk database.

[0011] Furthermore, the process of training the established operation risk identification model based on the operation risk database is as follows: obtain the operator violation behavior data set from the operation risk database for sample extraction; extract features from the sample data and input it into the operation risk identification model for pre-training to obtain pre-training results; fine-tune the operation risk identification model based on the pre-training results; extract features from the sample data and input it into the fine-tuned operation risk identification model for training; perform secondary sample extraction from the operator violation behavior data set to obtain verification data; extract features from the verification data and input it into the trained operation risk identification model to obtain training results.

[0012] Furthermore, the process of obtaining a data set of operator violation behaviors from an operation risk database for sample extraction is as follows: the obtained data set of operator violation behaviors is graded according to the set risk level, and the risk level includes low risk, medium risk and high risk; random samples are extracted from the graded data sets of operator violation behaviors of each level to obtain a low-risk sample data set, a medium-risk sample data set and a high-risk sample data set; each sample in the low-risk sample data set, the medium-risk sample data set and the high-risk sample data set is labeled, including the risk level, business type and timestamp.

[0013] Furthermore, the pre-training results include the pre-training success rate of each risk level and the successful identification time of each risk level; the process of fine-tuning the operation risk identification model based on the pre-training results is as follows: obtaining the expected pre-training results, the expected pre-training results include the expected pre-training success rate of each risk level and the expected successful identification time of each risk level; fusing the pre-training results with the expected pre-training results to obtain a pre-training evaluation index; if the pre-training evaluation index is greater than or equal to the set fine-tuning threshold, fine-tuning the operation risk identification model, and the fine-tuning algorithms used include Fine-tuning, Feature augmentation and Transfer learning; if the pre-training evaluation index is less than the set fine-tuning threshold, the operation risk identification model is not fine-tuned.

[0014] Furthermore, the training results include prediction accuracy, recall rate and precision of samples of each risk level; the process of evaluating the training results to obtain evaluation results is as follows: obtaining expected training results, the expected training results include expected prediction accuracy, expected recall rate and expected precision of samples of each risk level; fusing and comparing the training results with the expected training results to obtain a training result evaluation value, the training result evaluation value is used to represent the evaluation result of the training result, and characterize the training effect of the job risk identification model.

[0015] Furthermore, the process of obtaining an evaluation result of the training result, and tuning and verifying the job risk identification model based on the evaluation result is as follows: obtaining a training result evaluation value, and determining whether the training result evaluation value is greater than or equal to a set tuning threshold; if the training result evaluation value is greater than or equal to the set tuning threshold, obtaining a training deviation result, and matching a hyperparameter from a tuning parameter database based on the training deviation result, wherein the training deviation result is obtained by performing a difference analysis between the training result and the expected training result; applying the matched hyperparameter to the job risk identification model to be tuned, retraining and verifying, and obtaining a training result evaluation value; if the training result evaluation value is less than the set tuning threshold, no optimization is performed.

[0016] Furthermore, the training deviation results include prediction accuracy deviation values, recall deviation values ​​and precision deviation values ​​for samples of each risk level; the process of matching and obtaining hyperparameters from the tuning parameter database based on the training deviation results is as follows: the training deviation results are compared with each matching deviation data in the tuning parameter database to obtain a comparison index, and each matching deviation data includes a prediction accuracy matching value, a recall matching value and a precision matching value for samples of each risk level; the obtained comparison indexes are arranged in order from large to small, and the hyperparameters corresponding to the matching deviation data corresponding to the smallest comparison index are obtained, and the hyperparameters include learning rate, number of iterations and regularization strength.

[0017] Furthermore, the calculation formula of the comparison index is:

[0018]

[0019] Wherein, BdZ is the comparison index, Dx is the prediction accuracy deviation value of low-risk samples in the prediction accuracy deviation value of each risk level sample, Dy is the prediction accuracy matching value of low-risk samples in the prediction accuracy matching value of each risk level sample, Zx is the prediction accuracy deviation value of medium-risk samples in the prediction accuracy deviation value of each risk level sample, Zy is the prediction accuracy matching value of medium-risk samples in the prediction accuracy matching value of each risk level sample, Gx is the prediction accuracy deviation value of high-risk samples in the prediction accuracy deviation value of each risk level sample, Gy is the prediction accuracy matching value of high-risk samples in the prediction accuracy matching value of each risk level sample, α1 is the deviation weight factor of the risk level sample prediction accuracy, h x is the recall rate deviation value, hy is the recall rate matching value, px is the precision deviation value, py is the precision matching value, and α2 is the comprehensive weight factor of recall rate and precision.

[0020] Furthermore, the optimized and verified operational risk identification model is applied to operational risk identification to predict the risky behavior of staff and to provide early warning for risky behavior. The process is as follows: real-time work data of staff is obtained, and the real-time work data is input into the operational risk identification model for identification to determine whether there is a risk. If there is no risk, no early warning is issued; if there is a risk, an early warning is issued.

[0021] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for identifying and analyzing operational risks described in the present invention.

[0022] A work risk identification and analysis system includes a computer and a camera connected to the computer and located at a work site for acquiring real-time work images, and also includes a computer-readable storage medium described in the present invention. The camera acquires real-time work images and inputs them into the computer to determine whether there is a risk. If there is no risk, no early warning is given; if there is a risk, an early warning is given.

[0023] The beneficial effects of the present invention include:

[0024] The operation risk identification and analysis method of the present invention establishes an operation risk database, training models, evaluation results, and applies them to operation risk identification, so that the system can comprehensively identify the risk behaviors of workers at different risk levels and types, improves the ability to identify potential risks, and improves the accuracy and reliability of the risk identification model through model training, evaluation and tuning. It can accurately warn of potential risks in real-time applications and effectively reduce the possibility of accidents. Through evaluation results and real-time warning prompts, it can provide intelligent decision-making support for relevant personnel, helping them to take necessary measures in a timely manner to prevent potential risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the establishment, evaluation and prediction of a model of an operation risk identification method of the present invention.

[0026] Figure 2 It is a schematic diagram of neurons in a neural network, where: x is the input, θ is the weight, b is the threshold of the current neuron, the output is h, and the activation function is f.

[0027] Figure 3 It is a diagram of the Transformer model. DETAILED DESCRIPTION

[0028] The specific implementation of the method for establishing, evaluating and predicting an operation risk identification model of the present invention is described below in conjunction with the accompanying drawings and embodiments.

[0029] The embodiments of the present application improve the accuracy and reliability of the risk identification model through the establishment, evaluation and prediction methods of the operation risk identification model, can accurately warn of potential risks in real-time applications, and effectively reduce the possibility of accidents.

[0030] The problems in the embodiments of this application are generally solved as follows:

[0031] Collect and organize historical operation data, including work environment, operation type, operator information, possible risk factors, etc. Use deep learning technologies such as neural networks and Transformer to train models based on the established operation risk database. The input of the model can be the operation environment, operation type, operator information, etc., and the output is the prediction of operation risk. Evaluate the trained operation risk identification model, including the calculation of indicators such as accuracy, recall rate, F1 score, etc., to ensure the performance and reliability of the model.

[0032] According to the evaluation results, the operation risk identification model is tuned. Possible methods include adjusting model parameters, increasing training data, improving feature extraction, etc., and then verifying the effect of the model through validation sets or cross-validation. The optimized and verified operation risk identification model is applied to actual operation scenarios, real-time monitoring of the behavior of operators, and early warning prompts for possible risk behaviors based on the model prediction results. The early warning prompts are integrated with the alarm system. When potential risk behaviors are detected, an alarm is triggered to notify relevant personnel or departments in a timely manner so that necessary measures can be taken to prevent accidents.

[0033] See also Figure 2 , a neural network is a mathematical model or computational model that imitates the structure and function of biological neural networks and is used to estimate or approximate functions. The basic principle is to process and transform the input data through a series of neurons to finally obtain the output result. In a neural network, each node is also called an artificial neuron, which is connected to another node with associated weights and thresholds. If the output of any single node is higher than the specified threshold, the node is activated and the data is sent to the next layer of the network. Neural networks rely on training data to learn and improve their accuracy over time. Each neuron receives input signals from other neurons, calculates according to the strength of the input signal and the connection weight, generates an output signal, and transmits the signal to other neurons. By continuously learning and adjusting the connection weights between neurons, the neural network can gradually optimize its decision-making process and improve the accuracy and stability of the output results. The neural network algorithm is a computational model that simulates the structure of neurons in the human brain, and learns and makes decisions by simulating the connection and signal transmission process between neurons. It has broad application prospects in many fields such as job risk identification and analysis, job risk intelligent early warning and alarm.

[0034] The neural network model used in the present invention corresponds to the specific implementation of the preprocessing part and can also be used for feature extraction before model training. The main steps are as follows:

[0035] First layer of pooling: max pooling 2x2 kernel, stride 2;

[0036] The second convolution layer: 64 convolution kernels of size 5x5, with a stride of 2;

[0037] Second layer of pooling: max pooling 2x2 kernel, stride 2;

[0038] The third layer is a fully connected layer: 512 neural units, dropout is 0.6; it can also be 0.3, 0.5 or 0.8;

[0039] The fourth layer is the sub-network output layer;

[0040] Use the Softmax activation function to map the output to the probability space;

[0041] Except for the last layer, the activation functions of other layers use ReLU function;

[0042] The weight W is initialized with a Gaussian cutoff of μ = 0 and std = 0.1;

[0043] The learning rate when training the model is 0.001.

[0044] The neural network model can realize back propagation, gradient descent and regularization; by continuously iterating and optimizing the parameters in the neural network, the neural network can better adapt to different operational risk identification tasks and scenarios, and improve the performance and efficiency of operational risk identification. The specific implementation of the Transformer model, such as Figure 3 As shown, the Transformer model includes an input layer, a hidden layer and an output layer. Through the self-attention mechanism and positional encoding, the self-attention mechanism allows the model to pay attention to information at different positions when processing input data, while positional encoding solves the problem of position information encoding in traditional models. The self-attention mechanism and parallel computing are introduced. The self-attention mechanism compares each element in the sequence with other elements for the input data that constitutes the sequence, so as to correctly process each element in different contexts to generate the final output result, which is higher than the existing models and algorithms in terms of accuracy and performance. The self-attention mechanism includes two types: single-head self-attention mechanism and multi-head self-attention mechanism.

[0045] In the process of large model training, the neural network is used to learn how to extract features from raw data, automatically extracting features such as edges and textures in the data; Figure 3 Thinking (cognitive process); then, through the combination of large-scale convolutional layers and pooling layers, higher-level feature representations are gradually abstracted; corresponding Figure 3The machine (computation process) shown in Figure 1; finally, the fully connected layer is used for classification.

[0046] In the Transformer model, the input data is sent to the multi-head self-attention sublayer and the feedforward neural network sublayer for processing. The multi-head self-attention sublayer decomposes the input data into the attention distribution of multiple heads and allows the model to exchange information between different heads; the feedforward neural network sublayer adopts a feedforward neural network structure, which improves the expression, generalization and nonlinear expression capabilities of the model.

[0047] Self-Attention Mechanism: This mechanism is the core part of the Transformer model, which allows the model to focus on different parts when processing the input sequence. Specifically, given an input sequence X = [x1, x2, ..., xn], the model calculates the weight distribution of each input element and uses the weighted average method to obtain the representation vector of each position.

[0048] Positional Encoding: Since the Transformer model does not contain components with temporal dependencies such as recurrent neural networks (RNNs) or convolutional neural networks (CNNs), positional encoding needs to be added to ensure that the model can understand the position information of elements in the sequence.

[0049] Multi-Head Attention: To enhance the expressiveness of the model, the Transformer model uses a multi-head attention mechanism. It divides the input sequence into multiple subsequences and independently calculates the weight distribution and representation vector of each subsequence. These representation vectors are then concatenated and the final output vector is obtained through linear transformation.

[0050] Feed-Forward Neural Network: To further improve the expressiveness of the model, the Transformer model adds a feed-forward neural network after the multi-head attention mechanism. The network consists of two linear layers and uses the ReLU activation function.

[0051] During the training process, the Transformer model uses an optimizer to update model parameters to minimize prediction errors; batch normalization is also used to speed up the training process and improve the stability of the model.

[0052] See also Figure 1 The embodiment of the present invention provides a technical solution: a method for establishing, evaluating and predicting an operation risk identification model, comprising the following steps:

[0053] Establish an operation risk database; train the established operation risk identification model based on the operation risk database, wherein the operation risk identification model is established based on a neural network and a Transformer; obtain the training results of the operation risk identification model, and evaluate the training results; obtain the evaluation results of the training results, and tune and verify the operation risk identification model based on the evaluation results; apply the tuned and verified operation risk identification model to operation risk identification, predict the risk behavior of staff, and provide early warning prompts for risk behavior.

[0054] Specifically, the process of establishing an operation risk database is as follows: obtaining an operator violation behavior data set, the operator violation behavior data set including video data, image data and text data; performing data cleaning and preprocessing on the operator violation behavior data set, and performing data integration and conversion on the operator violation behavior data set after data cleaning and preprocessing; and assembling the integrated and converted operator violation behavior data set into an operation risk database.

[0055] In this embodiment, data cleaning involves detecting and repairing erroneous, missing, abnormal or inconsistent data in a data set. For video data, it is necessary to detect and repair video file corruption or format incompatibility issues; for image data, it is necessary to deal with noise, blur or deformation in the image; for text data, it is necessary to deal with spelling errors, grammatical errors or missing values. Cleaning data can be done using automated tools or manually to ensure the accuracy and completeness of the data.

[0056] After data cleaning and preprocessing, data from different sources, formats or types need to be integrated into a unified dataset and converted as necessary to meet modeling needs. This may combine video, image and text data, convert formats or encode data as needed to facilitate subsequent model training and analysis. The integrated and converted dataset has a consistent structure and standardized representation to facilitate subsequent data analysis and mining.

[0057] Specifically, the process of training the established operation risk identification model based on the operation risk database is as follows: obtain the operator violation behavior data set from the operation risk database for sample extraction; extract features from the sample data and input it into the operation risk identification model for pre-training to obtain pre-training results; fine-tune the operation risk identification model based on the pre-training results; extract features from the sample data and input it into the fine-tuned operation risk identification model for training; perform secondary sample extraction from the operator violation behavior data set to obtain verification data; extract features from the verification data and input it into the trained operation risk identification model to obtain training results.

[0058] The process of obtaining a data set of operator violation behaviors from an operation risk database for sample extraction is as follows: the obtained data set of operator violation behaviors is divided into levels according to the set risk levels, and the risk levels include low risk, medium risk and high risk; random samples are extracted from the data sets of operator violation behaviors of each level after the level division to obtain a low-risk sample data set, a medium-risk sample data set and a high-risk sample data set; each sample in the low-risk sample data set, the medium-risk sample data set and the high-risk sample data set is labeled, including the risk level, business type and timestamp.

[0059] In this embodiment, the data of operators' illegal behaviors are divided according to risk levels, which helps to focus on and deal with behaviors of different risk levels. By randomly selecting samples, the representativeness and randomness of the samples can be guaranteed, thereby avoiding sample selection bias and ensuring the generalization ability of the model. The feature extraction process can convert the original data into a more representative and information-rich feature representation, which helps the model better capture the key features of operational risk behaviors. Feature extraction can reduce the dimension and complexity of the data and reduce the computational cost and resource consumption of model training.

[0060] In the pre-training stage, large-scale data sets can be used to train the model in advance, so that the model has certain prior knowledge, which helps to accelerate model convergence and improve performance. The fine-tuning process can adjust the model parameters according to the needs of specific tasks, making it more suitable for operational risk identification tasks and improving the accuracy and generalization ability of the model in actual scenarios. The verification process can evaluate the performance of the model on unseen data, verify the generalization ability and robustness of the model, and improve the reliability and practicality of the model. Labeling the verification data can help further optimize and adjust the model, and can also provide a reference for subsequent model improvements.

[0061] Specifically, the pre-training results include the pre-training success rate of each risk level and the successful identification time of each risk level; the process of fine-tuning the operation risk identification model based on the pre-training results is as follows: obtaining the expected pre-training results, the expected pre-training results include the expected pre-training success rate of each risk level and the expected successful identification time of each risk level; fusing the pre-training results with the expected pre-training results for analysis to obtain a pre-training evaluation index; if the pre-training evaluation index is greater than or equal to the set fine-tuning threshold, fine-tuning the operation risk identification model, and the fine-tuning algorithms used include Fine-tuning, Feature augmentation, and Transfer learning; if the pre-training evaluation index is less than the set fine-tuning threshold, the operation risk identification model is not fine-tuned.

[0062] Large model fine-tuning refers to fine-tuning (FineTune) based on a given pre-trained model; compared with training a model from scratch, fine-tuning can save a lot of computing resources and computing time, improve computing efficiency, and even improve accuracy; the large model fine-tuning uses a small amount of labeled data to retrain the pre-trained model to adapt to a specific task. In this process, the parameters of the model are adjusted according to the new data distribution. This method takes advantage of the powerful capabilities of the pre-trained model while also adapting to the new data distribution. Therefore, model fine-tuning can improve the generalization ability of the model and reduce overfitting.

[0063] Common large model fine-tuning methods include Fine-tuning, Feature augmentation, and Transfer learning; Fine-tuning, by adding a new classification layer to the last layer of the pre-trained model, and then fine-tuning it according to the new dataset; Feature augmentation: by adding some artificial features to the data to enhance the performance of the model, these features can be manually designed or generated by automatic feature generation technology; Transfer learning: this method uses a model trained on one task as the starting point for a new task, and then fine-tunes the model's parameters to adapt to the new task.

[0064] By fusing the pre-training results with the expected results for analysis, the performance of the pre-trained model on specific tasks can be evaluated more comprehensively. By fusing the pre-training results with the expected results for analysis, the expected success rate and recognition success time of the model at different risk levels can be comprehensively considered to derive a comprehensive evaluation index as a basis for whether to perform fine-tuning.

[0065] Setting a fine-tuning threshold helps automate the decision-making fine-tuning process, improves the intelligence and adaptability of the system, and avoids unnecessary human intervention. If the evaluation index is higher than the threshold, it means that the pre-trained model has adapted to the task relatively well and fine-tuning can be skipped; otherwise, fine-tuning is required to further optimize the model.

[0066] In addition to obtaining the pre-training evaluation index through statistical analysis, the pre-training evaluation index can also be calculated using the following calculation formula:

[0067]

[0068] Wherein, Xz is the pre-training evaluation index, dc is the pre-training success rate of low risk level in the pre-training success rate of each risk level, dy is the expected pre-training success rate of low risk level in the expected pre-training success rate of each risk level, tdc is the low risk level recognition success time in the recognition success time of each risk level, tdy is the expected recognition success time of low risk level in the expected recognition success time of each risk level, zc is the pre-training success rate of medium risk level in the pre-training success rate of each risk level, zy is the expected pre-training success rate of medium risk level in the expected pre-training success rate of each risk level, tzc is the medium risk level recognition success time in the recognition success time of each risk level, tzy is the expected recognition success time of medium risk level in the expected recognition success time of each risk level, gc is the pre-training success rate of high risk level in the pre-training success rate of each risk level, gy is the expected pre-training success rate of high risk level in the expected pre-training success rate of each risk level, tgc is the high risk level recognition success time in the recognition success time of each risk level, tgy is the expected recognition success time of high risk level in the expected recognition success time of each risk level, and e is a natural constant.

[0069] The example values ​​of the parameters in the pre-training evaluation index are as follows:

[0070] Table 1 Example values ​​of parameters in the pre-training evaluation index

[0071] dc tdc z tzc gc tgc XZ 0.815 0.404 0.524 0.296 0.121 0.199 6.4 0.859 0.18 0.464 0.367 0.196 0.296 7.19 0.874 0.34 0.64 0.135 0.172 0.255 7.2

[0072] Among them, dy is 0.9, tdy is 0.4, zy is 0.75, tzy is 0.3, gy is 0.13, and tgy is 0.2.

[0073] In this embodiment, the formula calculates the difference between the actual pre-training success rate and the recognition success time and the expected value. This comparison can comprehensively evaluate the performance of the pre-training model at different risk levels, taking into account not only the pre-training success rate but also the recognition success time, thereby more comprehensively evaluating the performance of the model. For each risk level, the difference between the pre-training success rate and the recognition success time is calculated and weighted and summed. The purpose of this is to comprehensively consider the differences between different risk levels and derive a comprehensive evaluation index for subsequent decision-making.

[0074] The pre-training success rate of each risk level and the recognition success time of each risk level can be obtained through data statistics in actual applications. For example, during the model pre-training and testing process, the success rate and recognition time of each risk level are recorded, and then statistical analysis is performed to obtain the corresponding values.

[0075] Specifically, the training results include the prediction accuracy, recall rate and precision of samples of each risk level; the process of evaluating the training results to obtain the evaluation results is as follows: obtaining the expected training results, the expected training results include the expected prediction accuracy, expected recall rate and expected precision of samples of each risk level; fusing the training results with the expected training results for comparison and analysis to obtain the training result evaluation value, the training result evaluation value is used to represent the evaluation result of the training result, and characterize the training effect of the job risk identification model.

[0076] In addition to statistical analysis or manual comparison, the training result evaluation value can also be obtained through the following calculation formula, and before the calculation, the training result and the expected training result are processed quantitatively:

[0077]

[0078] In the formula, XgZ is the evaluation value of the training result, Db is the prediction accuracy of low-risk level samples in the prediction accuracy of samples of each risk level, Dc is the expected prediction accuracy of low-risk level samples in the expected prediction accuracy of samples of each risk level, Zb is the prediction accuracy of medium-risk level samples in the prediction accuracy of samples of each risk level, Zc is the expected prediction accuracy of medium-risk level samples in the expected prediction accuracy of samples of each risk level, Gb is the prediction accuracy of high-risk level samples in the prediction accuracy of samples of each risk level, Gc is the expected prediction accuracy of high-risk level samples in the expected prediction accuracy of samples of each risk level, ΔPC is the allowable deviation of accuracy, hx is the recall rate, hy is the expected recall rate, px is the precision, and py is the expected py.

[0079] In this embodiment, before calculation, the training results and the expected training results are processed in scalar form to ensure that each indicator has a unified dimension and range for easy comparison and analysis. The formula calculates the difference between the training results and the expected training results, including accuracy, recall and precision. This comparison can comprehensively evaluate the performance of the model at different risk levels, taking into account not only accuracy, but also recall and precision, thereby more comprehensively evaluating the training effect of the model.

[0080] The formula uses the method of summing absolute values ​​to quantitatively compare the differences between various indicators, making the evaluation values ​​comparable and quantitative.

[0081] Sample prediction accuracy, recall rate and precision can be obtained by comparing the model's prediction results on the training set or validation set with the true label. Usually, the model predicts each sample and compares it with the actual label to obtain indicators such as accuracy, recall rate and precision.

[0082] Specifically, the process of obtaining the evaluation results of the training results, and tuning and verifying the job risk identification model based on the evaluation results is as follows: obtaining the training result evaluation value, and determining whether the training result evaluation value is greater than or equal to the set tuning threshold; if the training result evaluation value is greater than or equal to the set tuning threshold, obtaining the training deviation result, and matching the hyperparameters from the tuning parameter database based on the training deviation result, wherein the training deviation result is obtained by performing a difference analysis between the training results and the expected training results; applying the matched hyperparameters to the job risk identification model to be tuned, retraining and verifying, and obtaining the training result evaluation value; if the training result evaluation value is less than the set tuning threshold, no optimization is performed.

[0083] The training deviation results include prediction accuracy deviation values, recall deviation values ​​and precision deviation values ​​for samples of each risk level; the process of matching and obtaining hyperparameters from the tuning parameter database based on the training deviation results is as follows: the training deviation results are compared with the matching deviation data in the tuning parameter database to obtain a comparison index, and each matching deviation data includes a prediction accuracy matching value, a recall matching value and a precision matching value for samples of each risk level; the obtained comparison indexes are arranged in order from large to small, and the hyperparameters corresponding to the matching deviation data corresponding to the smallest comparison index are obtained, and the hyperparameters include a learning rate, a number of iterations and a regularization strength.

[0084] In addition to statistical analysis of data or simulation evaluation of the model, the comparison index can also be obtained through the following calculation formula, and the training deviation results and matching deviation data are processed before calculation:

[0085]

[0086] In the formula, BdZ is the comparison index, Dx is the prediction accuracy deviation value of low risk level samples in the prediction accuracy deviation value of each risk level sample, Dy is the prediction accuracy matching value of low risk level samples in the prediction accuracy matching value of each risk level sample, Zx is the prediction accuracy deviation value of medium risk level samples in the prediction accuracy deviation value of each risk level sample, Zy is the prediction accuracy matching value of medium risk level samples in the prediction accuracy matching value of each risk level sample, Gx is the prediction accuracy deviation value of high risk level samples in the prediction accuracy deviation value of each risk level sample, Gy is the prediction accuracy matching value of high risk level samples in the prediction accuracy matching value of each risk level sample, α1 is the deviation weight factor of risk level sample prediction accuracy, hx is the recall rate deviation value, hy is the recall rate matching value, px is the precision deviation value, py is the precision matching value, and α2 is the comprehensive weight factor of recall and precision.

[0087] In this embodiment, first, the training deviation results are compared with the deviation data in the tuning parameter database to determine which hyperparameters best match the training deviation results. Secondly, according to the calculation formula of the comparison index, the similarity between each set of deviation data and the training deviation results is calculated to evaluate their matching degree. Finally, they are arranged in order of size according to the comparison index, and the hyperparameters corresponding to the deviation data with the highest matching degree are selected as the final selection.

[0088] The formula comprehensively considers the deviation values ​​of prediction accuracy, recall rate and precision of samples of each risk level, and takes into account the weight factors of different indicators to comprehensively evaluate the difference between the training results and the expected results. α1 and α2 in the formula represent the weights of accuracy deviation and comprehensive indicators (recall rate and precision), respectively. By adjusting these weight factors, the importance of different indicators can be adjusted according to actual needs, thereby better reflecting the performance of the model.

[0089] The sample prediction accuracy, recall rate and precision deviation values ​​are obtained by comparing the prediction results of the model on the training set or validation set with the true label, indicating the deviation of the model on different indicators. The matching value is obtained from the tuning parameter database, representing the matching value of the prediction accuracy, recall rate and precision of samples of each risk level, which is usually the expected value obtained based on prior knowledge or experiments.

[0090] Specifically, the process of applying the tuned and verified operation risk identification model to operation risk identification, predicting the risk behavior of staff, and issuing early warning prompts for risky behavior is as follows: obtaining the real-time work data of staff, inputting the real-time work data into the operation risk identification model for identification, and determining whether there is a risk. If there is no risk, no early warning prompt is issued; if there is a risk, an early warning prompt is issued.

[0091] In this embodiment, by inputting the real-time work data of the staff into the operational risk identification model for identification, real-time monitoring and identification of the staff's risky behavior can be achieved. When a risk is identified, the system will immediately issue an early warning prompt to remind relevant personnel to pay attention and take necessary measures.

[0092] A computer-readable storage medium is used to store a program, which, when executed by a processor, implements the method for establishing, evaluating and predicting an operation risk identification model as described above.

[0093] A work risk identification and analysis system includes a computer and a camera connected to the computer and located at a work site for acquiring real-time work images, and also includes a computer-readable storage medium described in the present invention. The camera acquires real-time work images and inputs them into the computer to determine whether there is a risk. If there is no risk, no early warning is given; if there is a risk, an early warning is given.

[0094] In summary, the embodiments of the present application have at least the following effects:

[0095] The establishment of an operational risk database, training models, evaluation results, and real-time applications enables the system to comprehensively identify risk behaviors of workers at different risk levels and types, improves the ability to identify potential risks, and improves the accuracy and reliability of the risk identification model through model training, evaluation, and tuning. In real-time applications, it can accurately warn of potential risks and effectively reduce the possibility of accidents. Through evaluation results and real-time warning prompts, it can provide intelligent decision-making support for relevant personnel, helping them to take necessary measures in a timely manner to prevent potential risks.

[0096] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0100] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0101] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for identifying and analyzing operational risks, characterized in that: The following steps are involved: (1) Establishing an operational risk database; (2) training an established operation risk identification model based on an operation risk database, wherein the operation risk identification model is established based on a neural network and a Transformer; (3) obtaining training results of the operation risk identification model and evaluating the training results; the training results include prediction accuracy, recall rate and precision of samples at each risk level; the evaluation of the training results includes obtaining expected training results, and the expected training results include expected prediction accuracy, expected recall rate and expected precision of samples at each risk level; (4) Obtaining evaluation results of the training results, and tuning and verifying the operational risk identification model based on the evaluation results; The process of obtaining the evaluation result includes: fusing and comparing the training result with the expected training result to obtain a training result evaluation value, wherein the training result evaluation value is used to represent the evaluation result of the training result and characterize the training effect of the operation risk identification model; (5) Apply the optimized and verified operational risk identification model to operational risk identification, predict the risk behaviors of staff members, and provide early warning for risk behaviors.

2. The method according to claim 1, characterized in that: The step (1) comprises: Acquire a data set of operator violation behaviors, wherein the data set of operator violation behaviors includes video data, image data and text data; perform data cleaning and preprocessing on the data set of operator violation behaviors, and perform data integration and conversion on the data set of operator violation behaviors after data cleaning and preprocessing; and construct the integrated and converted data set of operator violation behaviors into an operation risk database.

3. The method according to claim 1, characterized in that The step (2) comprises: A data set of operator violation behaviors is obtained from an operation risk database for sample extraction; after feature extraction of the sample data, the sample data is input into an operation risk identification model for pre-training to obtain pre-training results; based on the pre-training results, the operation risk identification model is fine-tuned; after feature extraction of the sample data, the sample data is input into the fine-tuned operation risk identification model for training; a second sample extraction is performed from the operator violation behavior data set to obtain verification data; after feature extraction of the verification data, the verification data is input into the trained operation risk identification model to obtain training results; the pre-training results include the pre-training success rate of each risk level and the successful identification time of each risk level.

4. The method according to claim 2, characterized in that: The process of obtaining the operator violation behavior data set from the operation risk database for sample extraction in step (2) includes: The obtained operator violation behavior data set is divided into levels according to the set risk levels, and the risk levels include low risk, medium risk and high risk; random samples are extracted from the operator violation behavior data sets of each level after the level division to obtain a low-risk sample data set, a medium-risk sample data set and a high-risk sample data set; each sample in the low-risk sample data set, the medium-risk sample data set and the high-risk sample data set is labeled, including the risk level, business type and timestamp.

5. The method according to claim 4, characterized in that Step (4) also includes: Obtain the training result evaluation value, and determine whether the training result evaluation value is greater than or equal to the set tuning threshold; if the training result evaluation value is greater than or equal to the set tuning threshold, obtain the training deviation result, match the hyperparameter from the tuning parameter database based on the training deviation result, and the training deviation result is obtained by performing a difference analysis between the training result and the training expected result; apply the matched hyperparameter to the job risk identification model to be tuned, perform retraining and verification, and obtain the training result evaluation value; if the training result evaluation value is less than the set tuning threshold, do not perform optimization; The training deviation results include prediction accuracy deviation values, recall deviation values ​​and precision deviation values ​​for samples of each risk level.

6. The method according to claim 5, characterized in that The process of tuning the operation risk identification model based on the pre-training results includes: Obtain expected pre-training results, including expected pre-training success rates for each risk level and expected identification success times for each risk level; fuse and analyze the pre-training results with the expected pre-training results to obtain a pre-training evaluation index; if the pre-training evaluation index is greater than or equal to a set fine-tuning threshold, fine-tune the operation risk identification model, using fine-tuning algorithms such as Fine-tuning, Feature augmentation, and Transfer learning; if the pre-training evaluation index is less than the set fine-tuning threshold, do not fine-tune the operation risk identification model.

7. The method according to claim 5, characterized in that The process of matching and obtaining hyperparameters from the tuning parameter database based on the training deviation results includes: Compare the training deviation results with the matching deviation data in the tuning parameter database to obtain a comparison index, wherein each matching deviation data includes a prediction accuracy matching value, a recall matching value, and a precision matching value for samples of each risk level; arrange the obtained comparison indexes in descending order, and obtain the hyperparameters corresponding to the matching deviation data corresponding to the smallest comparison index, wherein the hyperparameters include a learning rate, a number of iterations, and a regularization strength; The calculation formula of the comparison index is: In the formula, BdZ is the comparison index, Dx is the prediction accuracy deviation value of low risk level samples in the prediction accuracy deviation value of each risk level sample, Dy is the prediction accuracy matching value of low risk level samples in the prediction accuracy matching value of each risk level sample, Zx is the prediction accuracy deviation value of medium risk level samples in the prediction accuracy deviation value of each risk level sample, Zy is the prediction accuracy matching value of medium risk level samples in the prediction accuracy matching value of each risk level sample, Gx is the prediction accuracy deviation value of high risk level samples in the prediction accuracy deviation value of each risk level sample, Gy is the prediction accuracy matching value of high risk level samples in the prediction accuracy matching value of each risk level sample, α1 is the deviation weight factor of risk level sample prediction accuracy, hx is the recall rate deviation value, hy is the recall rate matching value, px is the precision deviation value, py is the precision matching value, and α2 is the comprehensive weight factor of recall and precision.

8. The method according to claim 1, characterized in that The step (5) further comprises: Obtain the real-time work data of the staff, input the real-time work data into the operation risk identification model for identification, and determine whether there is a risk. If there is no risk, no early warning will be issued. If there is a risk, an early warning will be issued.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a method for identifying and analyzing operational risks as described in any one of claims 1 to 8 are implemented.

10. A system for identifying and analyzing operational risks, comprising a computer and a camera connected to the computer and arranged at an operation site for acquiring real-time operational images, characterized in that: It also includes a computer-readable storage medium as described in claim 9, wherein the camera obtains real-time working images and inputs them into the computer to determine whether there is a risk. If there is no risk, no early warning prompt is issued; if there is a risk, an early warning prompt is issued.

Images