Communication engineering whole-process safety production risk dynamic assessment method

Abstract

The invention relates to a communication engineering full-process safety production risk dynamic assessment method, in particular to the technical field of communication engineering, and effectively overcomes the hysteresis quality and one-sidedness of a traditional static assessment method. According to the method, real-time data stream processing and an online learning mechanism are fused, continuous perception and quantitative evaluation of safety production risks are realized, short-term sudden risk response and long-term law research and judgment can be intelligently balanced, the timeliness of risk early warning and the accuracy of decision support are remarkably improved, the engineering field safety management and control capability is enhanced, and the engineering field safety management and control efficiency is improved. The occurrence probability of safety accidents is reduced.

Description

Technical Field

[0001] This invention relates to the field of communication engineering technology, and more specifically, to a method for dynamic assessment of safety production risks throughout the entire communication engineering process. Background Technology

[0002] Safety in telecommunications engineering involves multiple continuous and interconnected stages, including surveying and design, construction and installation, equipment commissioning, and operation and maintenance. The composition of risks exhibits significant dynamic evolution. In the initial surveying and design phase, geological conditions and the rationality of the design scheme are the primary risks. During the construction and installation phase, the standardization of personnel operations, the condition of large machinery and equipment, and complex on-site environmental factors become key risks. In the operation and maintenance phase, equipment fatigue wear, external environmental corrosion, and changes in network load become the dominant risk sources. Furthermore, changes in internal management processes, updates to external policies and regulations, and disturbances from unforeseen environmental events further exacerbate the complexity of the risk environment. The relative importance of these multi-dimensional risk factors throughout the entire process is not static but dynamically changes with project progress, technical conditions, and shifts in management priorities, placing extremely high demands on the timeliness and adaptability of risk assessment.

[0003] Currently, knowledge-driven models based on expert experience are widely used in the field of safety production risk assessment in telecommunications engineering, with the combined application of the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method being particularly common. The core of this technical solution lies in organizing domain experts to conduct back-to-back questionnaire surveys, assigning relative importance weights to multi-dimensional risk indicators such as personnel, equipment, environment, and management based on their subjective experience, thereby constructing an assessment model with a fixed weight set. This method is effective in dealing with static and steady-state risk assessment scenarios. However, its inherent limitation is that once the weight system is determined, it remains unchanged throughout the entire project cycle, failing to respond to feedback from actual project data. As the project progresses from one phase to the next, or encounters unforeseen changes in internal and external conditions, the fixed weights initially preset by experts will gradually become out of touch with reality, leading to inaccurate model assessments. For example, during peak construction periods, the model may still use the weight allocation from the design phase, which focused on scheme review, failing to dynamically increase attention to real-time risks such as high-altitude operations and electrical safety, resulting in the underestimation or omission of major risk sources. The lag and bias in assessments caused by rigid weighting not only weaken the scientific nature of risk management decisions but also potentially create serious safety hazards, leading to resource misallocation and economic losses. Therefore, developing an assessment method that can autonomously adjust weights based on real-time data and feedback, adapting to dynamic risk changes throughout the entire process, has become a pressing technological bottleneck in this field. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a dynamic assessment method for safety production risks throughout the entire communication engineering process, thereby resolving the issues raised in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for dynamic assessment of safety production risks throughout the entire process of communication engineering, comprising the following steps:

[0006] S1. Receive multi-source heterogeneous data streams from the communication engineering site in real time, extract risk indicators from multiple dimensions based on the multi-source heterogeneous data streams; preprocess the multi-source heterogeneous data streams to eliminate time inconsistencies and data missing issues; on this basis, based on a preset rolling time window, calculate the original values ​​of each risk indicator, the trend characteristics of value changes, and the real-time correlation characteristics between different risk indicators, and combine these characteristics to generate a standardized risk state vector at the current moment.

[0007] S2. Input the current risk state vector into the online learning model. The online learning model generates an instant reward signal based on the real-time risk event monitoring results, and adaptively updates the model parameters based on the instant reward signal, outputting a short-term weight vector.

[0008] S3. Input the time series containing the risk state vectors of the current moment and the historical moment into the time series analysis model to extract the time series pattern features; at the same time, adopt the reinforcement learning framework, with the communication engineering stage as the state space and the weight allocation scheme as the action space, and combine the safety performance score of the engineering milestone as the long-term reward signal, and output the long-term weight vector through the collaborative learning of the time series analysis model and the reinforcement learning framework.

[0009] S4. Calculate the short-term confidence level of the short-term weight vector and the long-term confidence level of the long-term weight vector. Make a decision based on the preset confidence threshold, selectively fuse the short-term weight vector and the long-term weight vector, and output the final weight vector for dynamic risk assessment.

[0010] In a preferred embodiment, in step S1, the multi-source heterogeneous data stream includes equipment status monitoring data, personnel operation record data, and environmental sensor data;

[0011] The specific steps for preprocessing multi-source heterogeneous data streams to eliminate timing inconsistencies and missing data are as follows:

[0012] Timestamp alignment and missing value imputation are performed on multi-source heterogeneous data streams. First, timestamp alignment is performed. A reference time granularity with a preset duration of ten milliseconds to one second is set as a unified time slot. The equipment status monitoring data stream, personnel operation record data stream, and environmental sensor data stream are aligned according to the unified time slot. Multiple data points existing in the same unified time slot are aggregated by arithmetic mean to generate an aligned unified data sequence.

[0013] Subsequently, missing value imputation is performed based on the spatiotemporal correlation graph of the communication engineering site. The spatiotemporal correlation graph uses all monitoring devices as vertices and the spatial adjacency between monitoring devices as edges. The edge weight is calculated using the following formula: the edge weight equals the functional similarity between devices divided by one plus the product of the distance attenuation coefficient and the physical distance between devices. When a monitoring device is found to have missing data at a specific time, data from the same time is obtained from the three to ten nearest neighbor monitoring devices determined by the edge weights in the spatiotemporal correlation graph. Imputed values ​​are generated through a graph attention network, where the graph attention network calculates attention coefficients based on the feature vectors of the nearest neighbor monitoring devices. The attention coefficients are obtained by normalization using the Softmax function.

[0014] In a preferred embodiment, in step S1, the specific operation of calculating the original values, value change trend characteristics, and real-time correlation characteristics between different risk indicators based on a preset rolling time window, and combining these characteristics to generate a standardized risk state vector at the current moment, is as follows:

[0015] The numerical trend characteristics are calculated based on a rolling time window with a preset duration of one minute to one hour. The weighted least squares method is used to calculate the numerical trend characteristics of each risk indicator. The weight matrix is ​​a diagonal matrix, and the weight values ​​on the diagonal are set according to the following rules: the weight values ​​are positively correlated with the exponential decay function values ​​corresponding to the time points, and the decay factor of the exponential decay function is preset to 0.01 to 0.1.

[0016] Real-time correlation feature calculation, based on the same rolling time window, uses the symbolic dynamic entropy method to calculate the real-time correlation features between different risk indicators, including: first, converting the time series of each risk indicator into a symbolic sequence, the symbolization method is to map continuous values ​​into a sequence based on first-order difference symbols, then calculating the joint probability distribution of the symbolic sequences of multiple risk indicators, and finally calculating the symbolic dynamic entropy value based on the joint probability distribution;

[0017] Feature combination involves concatenating the original values ​​of standardized risk indicators, the calculated trend features of numerical changes, and real-time correlation features into vectors to generate a standardized risk state vector for the current moment.

[0018] In a preferred embodiment, the specific operation of the online learning model generating an instant reward signal based on the real-time risk event monitoring results in step S2 is as follows:

[0019] First, the current risk state vector is input into an online learning model. The online learning model maps the risk state to a comprehensive risk score. The comprehensive risk score is calculated by performing a dot product operation between the current short-term weight vector and the current risk state vector, and adding a bias term.

[0020] Subsequently, an instant reward signal is generated. When the system detects a risk event of a preset level, a negative instant reward signal is automatically generated. The value of the negative instant reward signal is equal to the severity level of the negative risk event divided by the severity baseline value. The severity levels of the risk event range from level one to level ten, and the severity baseline value is five. When the system does not report any risk event of a preset level for six consecutive hours, a positive instant reward signal is automatically generated. The value of the positive instant reward signal is equal to the actual safe duration divided by the maximum monitoring window of 24 hours.

[0021] In a preferred embodiment, step S2, which involves adaptively updating the model parameters based on the immediate reward signal and outputting the short-term weight vector, specifically involves:

[0022] First, a loss function is constructed, which consists of three parts. The first part is half of the squared error between the predicted comprehensive risk score and the actual risk label. The actual risk label is 1 when there is a risk event and 0 when there is no risk event. The second part is a regularization term, which is composed of the regularization coefficient multiplied by half of the square of the Euclidean norm of the short-term weight vector. The regularization coefficient is preset to 0.01. The third part is a reward influence term, which is composed of the reward influence coefficient multiplied by the immediate reward signal. The reward influence coefficient is preset to 0.1.

[0023] Then, the short-term weight vector is updated using an online gradient descent algorithm. The update amount is the learning rate multiplied by the negative of the gradient of the loss function with respect to the short-term weight vector. The learning rate is preset to 0.005. The gradient of the loss function with respect to the short-term weight vector includes a prediction error term, a regularization term, and a reward signal gradient term. The reward signal gradient term is simplified to zero in actual calculation. The updated short-term weight vector is output as the new short-term weight vector.

[0024] In a preferred embodiment, step S3, which involves inputting the time series containing risk state vectors from the current and historical times into the time series analysis model to extract time series pattern features, includes the following specific operations:

[0025] First, a time series network based on a multi-head attention mechanism is constructed. The input of the time series network is a risk state vector sequence from step S1, which includes the current time and historical time. The number of consecutive time points in the risk state vector sequence is preset to 30-100.

[0026] Secondly, the multi-head attention mechanism is configured with 4-16 attention heads, and the feature dimension of each attention head is set to an integer between 64 and 512. By calculating the query matrix, key matrix and value matrix, the input risk state vector sequence is subjected to feature transformation and weighted fusion.

[0027] Finally, a multilayer perceptron is used to perform a nonlinear transformation on the weighted and fused features through the attention mechanism, and output a temporal pattern feature vector. The multilayer perceptron contains 1-3 hidden layers, and the number of neurons in each layer is preset to 128-512. The temporal pattern feature vector is used as the feature input for the subsequent reinforcement learning framework.

[0028] In a preferred embodiment, in step S3, a reinforcement learning framework is adopted, with the communication engineering stage as the state space, the weight allocation scheme as the action space, and the safety performance score of the engineering milestone as the long-term reward signal. The specific operation of outputting the long-term weight vector through the collaborative learning of the time series analysis model and the reinforcement learning framework is as follows:

[0029] An Actor-Critic reinforcement learning framework is adopted. First, the entire process of communication engineering is divided into 5-10 discrete stages to define the state space, with each stage corresponding to a unique state identifier.

[0030] Secondly, different long-term weight vector allocation schemes are defined as action spaces. The dimension of the long-term weight vector is consistent with the number of risk indicators. The value range of each component is 0-1, which represents the importance weight of the risk indicator. The action selection is based on the output of the policy network.

[0031] Furthermore, a safety performance score is calculated based on engineering milestones as a long-term reward signal. The safety performance score is obtained by weighted summation of three items: accident incidence rate score, number of violations score, and equipment failure duration score. The weight ratio of the three items is preset to 1:2:1.

[0032] Finally, the probability distribution of actions is output through the Actor network, and the value function of the state is evaluated through the Critic network. Both the Actor network and the Critic network adopt a 3-layer fully connected neural network structure with 256 neurons in each layer. The reinforcement learning framework uses long-term reward signals and temporal pattern feature vectors for co-training and outputs long-term weight vectors.

[0033] In a preferred embodiment, the specific operation of calculating the short-term confidence level for the short-term weight vector in step S4 includes:

[0034] The accuracy of risk event prediction based on the online learning model within the most recent preset time window is calculated to obtain the basic confidence level.

[0035] The accuracy calculation uses a time decay weighting method, which assigns higher weight to prediction results that are closer to the current time.

[0036] Meanwhile, a correction factor based on the statistical characteristics of prediction error is introduced to correct the base confidence level in order to output the final short-term confidence level.

[0037] In a preferred embodiment, the specific operation of calculating the long-term confidence for the long-term weight vector in step S4 includes:

[0038] The basic confidence level is calculated based on the information entropy of the probability distribution of the current project stage judgment output by the time series network.

[0039] It also introduces a correction term based on the consistency between the current judgment and the historical judgment to enhance the basic confidence level, so as to output the final long-term confidence level.

[0040] In a preferred embodiment, in step S4, the specific operation of making a decision based on a preset confidence threshold, selectively fusing short-term and long-term weight vectors, and outputting the final weight vector for dynamic risk assessment is as follows:

[0041] The weighted fusion decision adopts a three-level decision mechanism based on confidence thresholds, specifically including:

[0042] When the short-term confidence level is higher than the first threshold, the short-term weight vector is directly used as the final weight vector output.

[0043] When the short-term confidence level is lower than the second threshold and the long-term confidence level is higher than the third threshold, the long-term weight vector is directly used as the final weight vector output.

[0044] When neither of the above two conditions is met, the short-term weight vector and the long-term weight vector are weighted and summed according to an adaptively calculated fusion weight, and the summation result is output as the final weight vector.

[0045] The beneficial effects of this invention are: it effectively overcomes the lag and one-sidedness of traditional static evaluation methods, and by constructing a multi-dimensional dynamic risk indicator system covering all stages of the entire life cycle of communication engineering, and integrating real-time data stream processing and online learning mechanisms, it realizes continuous perception and quantitative evaluation of safety production risks. This method can intelligently balance short-term emergency risk response and long-term pattern judgment, significantly improve the timeliness of risk warning and the accuracy of decision support, thereby enhancing the on-site safety management and control capabilities of engineering projects and reducing the probability of safety accidents. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0049] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0050] This embodiment provides, for example Figure 1 The method for dynamic assessment of safety production risks throughout the entire process of communication engineering, as shown, includes the following steps:

[0051] S1. A multi-source heterogeneous data stream from the communication engineering site is received in real time via a stream processing platform. This multi-source heterogeneous data stream includes equipment status monitoring data, personnel operation record data, and environmental sensor data. Based on this multi-source heterogeneous data stream, multiple risk indicators are extracted and formed. These risk indicators include at least the personnel operation compliance rate, equipment operation stability index, and environmental safety coefficient. The multi-source heterogeneous data stream is then processed with timestamp alignment and missing value imputation. Based on a preset rolling time window, the original values ​​of each risk indicator, the value change trend characteristics based on the rolling time window, and the real-time correlation characteristics between different risk indicators are calculated. Subsequently, the calculated original values, the value change trend characteristics, and the real-time correlation characteristics are combined to construct a standardized risk state vector for the current moment.

[0052] S2. Input the current risk state vector into an online learning model; when the system detects a risk event of a preset level, it automatically generates a negative immediate reward signal; when the system does not report any risk event of a preset level within a preset time period, it automatically generates a positive immediate reward signal; the online learning model updates the internal weight parameters of the model using an online gradient descent algorithm based on the immediate reward signal, and outputs the updated internal weight parameters of the model as a short-term weight vector;

[0053] S3. A time series containing risk state vectors at the current and historical moments is input into an attention-based time series network to extract temporal pattern features from the time series. Simultaneously, an Actor-Critic reinforcement learning framework is employed, defining different stages of the communication engineering project as the state space, different weight allocation schemes as the action space, and using safety performance scores calculated in units of engineering milestones as long-term reward signals. The time series network and the reinforcement learning framework work together to learn and output a long-term weight vector.

[0054] S4. Calculate a short-term confidence score for the short-term weight vector obtained in S2. This short-term confidence score is calculated based on the accuracy of the online learning model in predicting risk events within the most recent preset time window, where the accuracy is the proportion of correct predictions to the total number of predictions. Calculate a long-term confidence score for the long-term weight vector obtained in S3. This long-term confidence score is calculated by calculating the information entropy of the probability distribution of the time series network's judgment output for the current project stage, and the difference between 1 and the information entropy is used as the long-term confidence score. If the short-term confidence score is higher than a preset first threshold, the short-term weight vector is directly output as the final weight vector. Otherwise, the short-term weight vector and the long-term weight vector are weighted and summed according to their respective confidence scores, and the sum is used as the final weight vector.

[0055] In this embodiment, it is specifically necessary to explain the specific operations of timestamp alignment and missing value interpolation for multi-source heterogeneous data streams in step S1:

[0056] First, timestamp alignment is performed. A preset time granularity of ten milliseconds to one second is set as a unified time slot. The equipment status monitoring data stream, personnel operation record data stream, and environmental sensor data stream are aligned according to the unified time slot. Multiple data points existing in the same unified time slot are aggregated by arithmetic mean to generate an aligned unified data sequence.

[0057] The specific implementation details of timestamp alignment are as follows:

[0058] Set the reference time granularity A time slot of 100 milliseconds (this value references the industrial IoT real-time data processing standard, striking a balance between timeliness and computational overhead) is used as a unified time slot to process the device status monitoring data stream. Personnel operation record data stream and environmental sensor data stream Alignment is performed according to this time slot. Multiple data points existing within the same time slot are aggregated using the arithmetic mean to generate an aligned unified data sequence. The expression for aggregation using the arithmetic mean is as follows:

[0059] ;

[0060] in, It represents a single data value output after aggregation, and is a measure of the average level of all data points in a dataset; This represents the total number of data points from the same data source within that specific time slot; Represents the first time slot within that time slot One set of original data values; the final average value calculated using a formula. As the final valid value of this data source in that time slot, this process is repeated for all time slots, ultimately generating an aligned, uniform data sequence with equal time intervals and no duplicate data points for each data source. This operation ensures data synchronization in time, providing a foundation for subsequent processing.

[0061] Subsequently, missing value imputation was performed based on the spatiotemporal correlation graph of the communication engineering site. The spatiotemporal correlation graph uses all monitoring devices as vertices and the spatial adjacency between monitoring devices as edges. The weight of the edge is calculated using the following formula: the weight equals the functional similarity between devices divided by one plus the product of the distance attenuation coefficient and the physical distance between devices. The specific implementation details of the missing value imputation process are as follows:

[0062] Spatiotemporal correlation diagram constructed based on communication engineering site , where: vertex set This represents all monitoring equipment (such as tower crane sensors, environmental monitors); edge set Represents the spatial adjacency between devices (e.g., physical distance less than 50 meters is considered adjacent); edge weights. Calculated using the following formula:

[0063] ;

[0064] in, Indicates equipment and equipment The correlation weight between two devices is a dimensionless quantity that takes positive values ​​(usually between 0 and 1). It represents the strength of the correlation between the two devices in risk assessment. The higher the weight, the stronger the correlation between the two devices. For equipment The greater the impact, the more it is used for data interpolation in graph attention networks in implementations to quantify spatiotemporal dependencies between devices; Indicates device and equipment Functional similarity between them is a dimensionless scalar with a value range of [0, 1], which is preset by the equipment type (e.g., two tower cranes). =1.0, while the tower crane and temperature sensor =0.2), its function is to measure the degree of similarity between devices in terms of function, the larger the value, the closer the functions are; This represents the distance attenuation coefficient, a dimensionless positive real number, which is set to 0.5 in this embodiment (determined through pre-experimental grid search). Its function is to adjust the degree of influence of physical distance on the weights; a larger value indicates better performance. This will make the effect of distance more significant; Indicates equipment and equipment The physical distance between them, measured in meters (m), is obtained from the on-site equipment layout drawings. It represents the degree of spatial separation between the equipment; the larger the value, the greater the weight. The smaller the value, the better; this formula is used to calculate the correlation weight between field monitoring equipment in communication engineering. This weight is used as the edge weight in the spatiotemporal correlation graph and supports the imputation of missing data.

[0065] When a monitoring device is detected At a specific moment When data is missing, the monitoring device's position in the spatiotemporal correlation diagram can be used to identify the missing data. A set of nearest neighbor monitoring devices Data from the same time point is obtained, and interpolated values ​​are generated using a graph attention network. The expression for generating interpolated values ​​using a graph attention network is:

[0066] ;

[0067] in, Indicates equipment In time The interpolated values ​​serve to provide continuous risk indicator inputs when data is missing; Indicates equipment The identifier belongs to the device. The neighbor set is an index variable representing a node in the spatiotemporal correlation graph; equipment of Nearest neighbor set, among which It is a positive integer, and in this embodiment, it is taken as 5 (based on empirical values ​​to balance computational complexity and accuracy). This set includes devices... closest One device (based on a spatiotemporal correlation graph) is used for local data borrowing; Represented as attention coefficient, indicating the device For equipment The influence weight is used to quantify the relative importance of neighboring devices in interpolation; Indicates equipment In time The actual observed values, whose units depend on risk indicators (such as equipment temperature, number of operations, etc.), are real-time measurements from the raw data stream. This formula imputes missing data by weighted summation of observations from neighboring devices, with weights determined by the attention coefficient. This decision reflects the idea of ​​graph neural networks, which utilizes local correlations to improve interpolation accuracy and avoid the biases of traditional methods (such as mean interpolation).

[0068] The graph attention network calculates the attention coefficient based on the feature vectors of nearest neighbor monitoring devices. Attention coefficient The expression is obtained by normalization using the Softmax function:

[0069] ;

[0070] in, Represented as the attention coefficient, defined as above; It is represented as a learnable parameter vector, with dimensions matching the concatenated vector after feature transformation, used to calculate attention scores, and optimized through training to capture inter-device correlation patterns; Represented as a learnable weight matrix, it is used to perform linear transformations on device features. Its dimension depends on the dimension of the input features. Its function is to map the original features to a high-dimensional space, thereby enhancing expressiveness. equipment and The feature vectors, which contain device state data, are the input to the graph attention network; This is represented as a concatenation operation of feature vectors, used to compute the joint representation of device pairs; It is an activation function whose function is to introduce nonlinearity to prevent gradient vanishing; The exponential function converts the score into a positive number, which facilitates softmax normalization. The formula calculates the attention score of the device pair using learnable parameters and normalizes it using softmax. This enables the model to adaptively focus on important neighbors, overcomes the limitations of the fixed weight method, and improves the interpolation accuracy.

[0071] In step S1, the specific operations for calculating the original values ​​of each risk indicator, the value change trend characteristics based on the rolling time window, and the real-time correlation characteristics between different risk indicators, based on a preset rolling time window, include:

[0072] The numerical trend characteristics are calculated based on a rolling time window with a preset duration of one minute to one hour. The weighted least squares method is used to calculate the numerical trend characteristics of each risk indicator. The weight matrix is ​​a diagonal matrix, and the weight values ​​on the diagonal are set according to the following rules: the weight values ​​are positively correlated with the exponential decay function values ​​corresponding to the time points, and the decay factor of the exponential decay function is preset to 0.01 to 0.1.

[0073] The expression for calculating the numerical trend characteristics is:

[0074]

[0075] This formula is used to calculate the time decay weight in weighted least squares. This assigns greater importance to recent data for trend feature extraction; among them, For time points The weight is a dimensionless positive real number. Its function is to weight data points in trend calculation. The larger the value, the greater the influence of that point in time on the current trend. The decay factor is a dimensionless positive real number, set to 0.05 in this embodiment (based on the median of a common range in deep learning), which controls the rate of weight decay; a larger value... This will cause the weight of historical data to decay more quickly; For reference time units ( =1min), used for dimensional normalization; The duration of the rolling time window is in minutes (min). In this embodiment, it is 30 minutes (the risk change cycle is determined through pre-experimentation). It defines the time range for trend calculation. This is the offset of the current time point relative to the start of the window;

[0076] Real-time correlation feature calculation, based on the same rolling time window, uses the symbolic dynamic entropy method to calculate the real-time correlation features between different risk indicators, including: first, converting the time series of each risk indicator into a symbolic sequence, the symbolization method is to map continuous values ​​into a sequence based on first-order difference symbols, then calculating the joint probability distribution of the symbolic sequences of multiple risk indicators, and finally calculating the symbolic dynamic entropy value based on the joint probability distribution;

[0077] The expression for real-time correlation feature calculation is:

[0078] ;

[0079] in, Information entropy is a dimensionless scalar; a larger value indicates higher uncertainty or complexity. In implementation, it is used to quantify the randomness of dynamic correlations among multiple risk indicators; a high value indicates a complex correlation pattern. Indicates the first The probability of a symbol or event occurring; The number of possible symbols is a positive integer; It is a logarithmic function, usually using the natural logarithm (base e) or the logarithm to base 2; the meaning of this formula is: entropy measures the degree of dispersion of the probability distribution. In risk assessment, a high entropy value may indicate a non-linear or unpredictable correlation between risk factors, thereby enhancing the model's ability to capture complex risks and replacing the traditional linear correlation coefficient.

[0080] Feature combination involves concatenating the original values ​​of each risk indicator after standardization (using Z-score standardization), the calculated numerical trend features, and real-time correlation features into a vector to generate a standardized risk state vector for the current moment. The formula for constructing the state vector is as follows:

[0081] ;

[0082] in, Indicates time The risk state vector is a comprehensive feature vector with a dimension equal to the sum of the dimensions of each sub-vector. It serves as the input to the evaluation model and encapsulates multi-dimensional information about the current security state. The standardized original numerical vector contains the Z-score standardized values ​​of various risk indicators (such as personnel operation standardization rate and equipment stability index). It is dimensionless and its function is to preserve the instantaneous state of the original data. The numerical trend feature vector is derived from the numerical trend estimated by the weighted least squares method, which represents the direction and magnitude of change of each risk indicator within the time window. To capture the dynamic dependencies between risk indicators by associating feature vectors in real time, this formula integrates raw data, trends, and correlation features to form a comprehensive state representation. This concatenation ensures that the risk assessment model can simultaneously consider instantaneous values, temporal evolution, and interaction effects, improving the accuracy and robustness of dynamic assessment.

[0083] In this embodiment, it is specifically necessary to explain the following steps in step S2: inputting the current risk state vector into an online learning model and generating an instant reward signal.

[0084] First, the risk state vector at the current moment. The input is fed into an online learning model, which maps the risk state to a comprehensive risk score. Comprehensive risk score The calculation method is as follows: perform a dot product operation between the current short-term weight vector and the current risk state vector, and add a bias term. The expression is:

[0085] ;

[0086] in, This is the risk state vector at the current moment, derived from the output of step S1; This is the short-term weight vector at the current moment, with dimensions matching the risk state vector, and each element representing a corresponding risk feature (from...). The importance (weight) of the current overall risk; This is a bias term used to adjust the baseline output of the model; The comprehensive risk score represents the model's quantitative assessment of the current risk level. This formula aggregates multi-dimensional risk features into a single score through linear combination, providing a quantitative basis for subsequent risk assessment. Its advantage lies in its high computational efficiency and ability to meet real-time requirements.

[0087] Then, an instant reward signal is given. When the system detects a risk event of a preset level, it automatically generates a negative instant reward signal. The value of this negative instant reward signal is equal to the severity level of the negative risk event divided by a severity baseline value, where the risk event severity levels range from level one to ten, and the severity baseline value is five. The calculation formula for triggering the negative reward is as follows:

[0088] ;

[0089] in, (Risk event severity level) is an integer ranging from level one to level ten. The severity baseline is fixed at five; when the system detects a risk event of a preset level, this formula generates a negative reward signal as a penalty.

[0090] When the system reports no risk events of a preset level for six consecutive hours, it automatically generates a positive immediate reward signal. The value of this positive immediate reward signal is equal to the actual safe duration divided by the maximum monitoring window of twenty-four hours. The calculation formula for generating the positive immediate reward signal is as follows:

[0091] ;

[0092] in, Set the maximum monitoring window to 24 hours; The actual safe duration refers to the length of time during which the system has not experienced a risk event of the preset level for an extended period. This formula generates a positive reward signal as a reward when the system is running safely for an extended period.

[0093] The design principle of the reward signal is as follows: negative rewards are proportional to the severity of the event to ensure that major risk events receive sufficient attention; positive rewards increase with the duration of safety to encourage continuous safe production; and the two-way adjustment mechanism enables the model to adapt to changes in on-site risks.

[0094] In step S2, the online learning model updates its internal weight parameters using the online gradient descent algorithm based on the immediate reward signal, specifically as follows:

[0095] First, a loss function is constructed, which consists of three parts. The first part is half of the squared error between the predicted comprehensive risk score and the actual risk label. The actual risk label is 1 when there is a risk event and 0 when there is no risk event. The second part is a regularization term, which is composed of the regularization coefficient multiplied by half of the square of the Euclidean norm of the short-term weight vector. The regularization coefficient is preset to 0.01. The third part is a reward influence term, which is composed of the reward influence coefficient multiplied by the immediate reward signal. The reward influence coefficient is preset to 0.1.

[0096] Specifically, construct a composite loss function. Construct a composite loss function It consists of three parts:

[0097] ;

[0098] The first term, prediction error, measures the model's predicted values. Compared with actual risk labels Differences ( Values: 1 for risky events, 0 for risk-free events); Second term: L2 regularization term. (Regularization coefficient) is preset to 0.01 to prevent overfitting; Third item: Reward impact item. The (reward impact coefficient) is preset to 0.1, introducing a risk feedback signal.

[0099] Then, the short-term weight vector is updated using an online gradient descent algorithm. The update amount is the learning rate multiplied by the negative of the gradient of the loss function with respect to the short-term weight vector. The learning rate is preset to 0.005. The gradient of the loss function with respect to the short-term weight vector includes a prediction error term, a regularization term, and a reward signal gradient term. The reward signal gradient term is simplified to zero in actual calculation. The updated short-term weight vector is output as the new short-term weight vector.

[0100] Specifically, the short-term weight vector is updated using an online gradient descent algorithm, and the update formula is as follows:

[0101] ;

[0102] in, For the next moment The short-term weight vector is the updated weight; For the current moment The short-term weight vector is the old weight before the update; The learning rate is preset to 0.005, which controls the update step size. loss function The gradient is a vector whose direction points in the direction in which the loss function grows the fastest.

[0103] The gradient term Expanded to:

[0104] ;

[0105] in, The gradient of the prediction error guides the adjustment of weights toward the true label. The gradient of the regularization term pushes the weight values ​​toward zero to prevent them from becoming too large; To reward the gradient of the signal item, real-time risk feedback is incorporated into the weight update; As a reward signal Weights The partial derivatives;

[0106] This update mechanism achieves triple optimization: 1. Ensuring model accuracy through the prediction error term; 2. Maintaining model stability through the regularization term; 3. Introducing on-site risk feedback through the reward term, making weight adjustments more targeted.

[0107] In this embodiment, it should be specifically explained that the specific operation in step S3, which involves inputting a time series containing risk state vectors at the current and historical times into a time series network based on an attention mechanism to extract temporal pattern features from the time series, includes:

[0108] First, a time series network based on a multi-head attention mechanism is constructed. The input of the time series network is a risk state vector sequence from step S1, which includes the current time and historical time. The number of consecutive time points in the risk state vector sequence is preset to 30-100. Each time point corresponds to a risk state vector. The risk state vector contains standardized values ​​of multiple risk indicators, such as equipment status and personnel operation standardization rate.

[0109] Secondly, when constructing a multi-head attention mechanism, it is configured with 4-16 attention heads, each with different feature dimensions. Set to an integer between 64 and 512, the query matrix, key matrix, and value matrix are used to perform feature transformation on the input risk state vector sequence. Key matrix Sum matrix The calculation formula is:

[0110] ;

[0111] Among them, RSVS is a risk state vector sequence, which is a sequence of standardized risk state vectors output from step S1 arranged in chronological order, and includes multi-dimensional risk information such as equipment status, personnel operation, and environmental conditions. , and It is a linear transformation function used to map the input to a query, key, and value space; , and The transformed matrix is ​​used to calculate the attention weights; this transformation projects the input sequence into different subspaces, enabling the model to capture temporal dependencies from multiple perspectives.

[0112] Next, the transformed features are weighted and fused using an attention mechanism, calculated as follows:

[0113] ;

[0114] in, For the feature dimension of the attention head, The function is used to convert attention scores into a probability distribution;

[0115] Finally, a multilayer perceptron is used to perform a nonlinear transformation on the attention-weighted fused features, outputting a temporal pattern feature vector. The calculation formula for the multilayer perceptron is as follows:

[0116] ;

[0117] in, The input feature vector, specifically the feature vector after weighted fusion through the attention mechanism, has the same dimension as the output dimension of the attention module. , This is the weight matrix, used for linear transformations; , This is the bias vector, used in conjunction with the weight matrix to increase the flexibility of the model; it is usually initialized as an all-zero vector. The activation function introduces nonlinear capabilities into the network. The output layer activation function determines the form of the output based on the final goal of the task.

[0118] The multilayer perceptron contains 1-3 hidden layers, with each layer having a preset number of 128-512 neurons. The temporal pattern feature vector is used as the feature input for the subsequent reinforcement learning framework.

[0119] The specific operations of using an Actor-Critic reinforcement learning framework in step S3 also include:

[0120] First, the entire communication engineering process is divided into 5-10 discrete stages to define the state space. Each stage corresponds to a unique state identifier. ,in Total number of stages;

[0121] Secondly, different long-term weight vector allocation schemes are defined as action spaces. Long-term weight vector Dimensions and number of risk indicators Consistent, each of its components The value of is in the range of 0-1, indicating that the is the . The importance weight of risk indicators, among which ;

[0122] Furthermore, safety performance scores are calculated based on engineering milestones as a long-term reward signal. The calculation formula is as follows:

[0123] ;

[0124] in, , , These are weighting coefficients, with the ratio preset to 1:2:1; Score the accident incidence rate; Scoring is given based on the number of violations. Rate the duration of equipment downtime;

[0125] The safety performance score is obtained by weighted summation of three components: accident incidence rate score, number of violations score, and equipment failure duration score. The weight ratio of the three components is preset to 1:2:1. Long-term reward signals are calculated based on the safety performance score. The calculation formula is:

[0126] ;

[0127] in, The long-term reward signal is the optimization objective of the reinforcement learning framework, used to evaluate the overall security performance at a certain engineering stage or milestone. The calculation result is a scalar value, and the theoretical range is usually between [−1, 1]. A safety performance score is a comprehensive safety indicator that quantifies safety performance over a specific period of time (such as during a project milestone). The highest safety performance score is used to assess... Normalize it so that it falls within the interval [0, 1]; Loss cost refers to the depreciated value of the total economic losses caused by safety accidents, equipment failures, etc. during the assessment period; The maximum acceptable cost of loss is used for... Normalization is the "red line" or "budget" for cost control.

[0128] Finally, the probability distribution of actions is output through the Actor network. The value function of the state is evaluated through a Critic network and optimized using a policy gradient algorithm. The objective function is:

[0129] ;

[0130] in, Let the policy objective function be the function that is maximized through gradient ascent. The expectation operator indicates that the current policy should be followed. Calculate the average of all possible situations (states, action sequences); The policy function is defined by the following parameters: The Actor network represents the output of a given state. Choose each possible action below The probability distribution; The advantage function is the advantage estimate calculated by the Critic network, which is used to evaluate the relative merit of actions; These are policy parameters, typically referring to the set of parameters for the entire policy model. The parameters of the Actor network are calculated by applying the objective function. The gradient is used to update the Actor network.

[0131] Both the Actor and Critic networks employ a 3-layer fully connected neural network structure, with 256 neurons per layer. The reinforcement learning framework utilizes long-term reward signals. and temporal pattern feature vector Perform collaborative training to output long-term weight vectors .

[0132] The specific operations for calculating the short-term confidence score for the short-term weight vector in step S4 include:

[0133] Based on the online learning model, within the most recent preset time window (the time window is denoted as...) The accuracy of risk event predictions within a preset timeframe of 6-48 hours is calculated to obtain a baseline confidence level. The accuracy calculation employs a time-decay weighted method, assigning higher weights to predictions more recent to the current time. The specific calculation formula is as follows:

[0134] ;

[0135] in, This is the short-term confidence level, the output value, which represents the reliability of the model's recent predictions; This is the current time step, used to define the end point of the calculation window; The length of the time window determines how much historical data to consider. The preset time is usually 6-24 hours (adjusted according to the application scenario); For example, a time decay weighting function assigns higher weights to recent data. Attenuation factor The reference value is 0.05-0.2; As an indicator function, when the predicted value equal to the actual value The value is 1 if it is true, and 0 otherwise. For at a certain point in time The predicted values ​​come from the output of the online learning model; For at a certain point in time The true value comes from actual observation data; The standard deviation of the prediction error is calculated over a time window to measure the volatility of the error. The mean of the prediction error is calculated over the time window to measure the average error level. To be a small constant, to prevent the denominator from being zero (in (in Chinese), the reference value is usually... .

[0136] Simultaneously, a correction factor based on the statistical characteristics of prediction error (i.e., the exponential term in the formula) is introduced to correct the base confidence level, thereby outputting the final short-term confidence level. .

[0137] The specific operations for calculating the long-term confidence of the long-term weight vector in step S4 include:

[0138] The probability distribution of the output based on the current project stage determined by the time series network (denoted as...) ,in The base confidence level for calculating information entropy is set at 5-10 stages (pre-defined for the total number of stages). The specific calculation formula is as follows:

[0139] ;

[0140] in, The long-term confidence output value is the result of the formula, which represents the reliability assessment score of the long-term weight vector. The reference value is in the range of [0, 1], and the closer it is to 1, the higher the confidence. Information entropy is a measure of the uncertainty of a probability distribution. The higher the entropy value, the higher the uncertainty; the value is ≥0, and the maximum value is... ; The probability distribution for the current stage: a vector output by the time series network. This represents the probability that the current stage belongs to any particular phase of the project. This represents the total number of project phases. The maximum entropy is used to normalize the information entropy, scaling the entropy value to the [0, 1] interval; This is the historical consistency weighting coefficient, which adjusts the weight of the historical consistency item in the total confidence level, with a reference range of 0.1-0.5; The maximum probability value is the maximum value in the probability distribution, reflecting the degree of certainty in the judgment at the current stage. This represents the historical probability distribution, specifically the moving average of the probability distribution within a historical window. It reflects historical judgment patterns, with larger values ​​indicating greater differences.

[0141] Furthermore, a correction term based on the consistency between the current judgment and historical judgments (i.e., the third term in the formula) is introduced to enhance the basic confidence level, thereby outputting the final long-term confidence level. .

[0142] The weighted fusion decision in step S4 adopts a three-level decision mechanism based on confidence thresholds, specifically including:

[0143] When the short-term confidence level is higher than the first threshold In this case, the short-term weight vector is directly used as the final weight vector output;

[0144] When the short-term confidence level is below the second threshold And the long-term confidence level is higher than the third threshold. In this case, the long-term weight vector is directly used as the final weight vector output;

[0145] When neither of the above two conditions is met, the short-term weight vector and the long-term weight vector are weighted and summed according to an adaptively calculated fusion weight, and the summation result is output as the final weight vector.

[0146] Specifically, a three-level decision-making mechanism is established, and the final weight vector is calculated according to the following rules:

[0147] ;

[0148] In this formula, based on short-term confidence level and long-term confidence With preset threshold Based on the comparison results, choose one of the three decision paths:

[0149] when Very high (exceeding) When using short-term weights, directly apply them. This indicates that the short-term model is highly reliable.

[0150] when Very low (below) )and High enough (above) When using long-term weights, directly apply them. This indicates that the short-term model is unreliable, but the long-term model is stable.

[0151] Otherwise (i.e.) (For those falling within the middle range or whose conditions are not met), a weighted average with short-term and long-term weights is used, where... It is a hybrid weight used for smooth transitions.

[0152] The formula achieves dynamic weight fusion, controls the decision boundary through threshold, avoids the limitations of a single model, and improves the adaptability and robustness of the system.

[0153] Among them, adaptive fusion weights Calculated using the following formula:

[0154] ;

[0155] In the formula, function, Scaling factor This formula is an adaptive controller that transforms the confidence differences between short-term and long-term models into a smooth, continuous mixed weight, which is the core of achieving intelligent weight fusion.

[0156] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0158] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0159] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0160] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A communication engineering full-process safety production risk dynamic assessment method, characterized in that, The method comprises the following steps: S1, real-time receiving a multi-source heterogeneous data stream from a communication engineering site, and extracting multiple dimensions of risk indicators based on the multi-source heterogeneous data stream; The multi-source heterogeneous data stream is preprocessed to eliminate timing inconsistency and data missing problems; on this basis, based on a preset rolling time window, the original value, value change trend characteristics of each risk indicator and real-time correlation characteristics between different risk indicators are calculated, and a standardized current time risk state vector is generated by combining these characteristics; S2, inputting the current time risk state vector into an online learning model, and outputting a short-term weight vector based on an instant reward signal generated by the online learning model according to a real-time risk event monitoring result and adaptive updating of model parameters based on the instant reward signal; S3, inputting a time sequence containing risk state vectors of the current time and historical times into a time sequence analysis model to extract time sequence pattern features; at the same time, using a reinforcement learning framework, taking a communication engineering stage as a state space, a weight distribution scheme as an action space, and combining a safety performance score of an engineering milestone as a long-term reward signal, a long-term weight vector is outputted through collaborative learning of the time sequence analysis model and the reinforcement learning framework; S4, calculating a short-term confidence of the short-term weight vector and a long-term confidence of the long-term weight vector, making a decision based on a preset confidence threshold, selectively fusing the short-term weight vector and the long-term weight vector, and outputting a final weight vector for risk dynamic assessment.

2. The communication engineering full-process safety production risk dynamic assessment method according to claim 1, characterized in that: In step S1, the multi-source heterogeneous data stream includes device state monitoring data, personnel operation record data and environmental sensor data; The specific operation of preprocessing the multi-source heterogeneous data stream to eliminate timing inconsistency and data missing problems is as follows: The multi-source heterogeneous data stream is timestamped and missing value interpolated, first, timestamp alignment processing is performed, a preset time length of ten milliseconds to one second is set as a reference time granularity as a unified time slot, the device state monitoring data stream, the personnel operation record data stream and the environmental sensor data stream are aligned according to the unified time slot, and the arithmetic mean value is used for aggregation of multiple data points existing in the same unified time slot to generate a unified data sequence after alignment; Subsequently, the missing value interpolation processing is performed, the missing data is interpolated based on a space-time correlation graph of the communication engineering site, the space-time correlation graph takes all monitoring devices as graph vertices, the spatial adjacent relationship between the monitoring devices as the edges of the graph, and the weight of the edge is calculated according to the following formula: the weight of the edge is equal to the functional similarity between the devices divided by one plus the product of the distance attenuation coefficient and the physical distance between the devices; when it is detected that a certain monitoring device has data missing at a certain time, the data at the same time is obtained from three to ten nearest neighbor monitoring devices of the monitoring device in the space-time correlation graph based on the weight of the edge, and the interpolation value is generated through a graph attention network, wherein the graph attention network calculates an attention coefficient based on the feature vector of the nearest neighbor monitoring device, and the attention coefficient is normalized by a Softmax function.

3. The communication engineering full-process safety production risk dynamic assessment method according to claim 2, characterized in that: In step S1, based on a preset rolling time window, the original values of each risk indicator, the value change trend characteristics, and the real-time association characteristics between different risk indicators are calculated, and these characteristics are combined to generate a standardized current moment risk state vector. The specific operation is as follows: The value change trend characteristic calculation is based on a preset rolling time window with a length of one minute to one hour. The weighted least squares method is used to calculate the value change trend characteristic of each risk indicator. The weight matrix is a diagonal matrix. The weight values on the diagonal are set according to the following rules: the weight value is positively correlated with the exponential decay function value corresponding to the time point, and the decay factor of the exponential decay function is preset to 0.01 to 0.

1. The real-time association characteristic calculation is based on the same rolling time window. The symbolic dynamics entropy method is used to calculate the real-time association characteristics between different risk indicators, including: first, the time series of each risk indicator is converted into a symbolic sequence. The symbolic method is to map the continuous numerical value into a sequence based on the first-order difference symbol. Then, the joint probability distribution of multiple risk indicator symbolic sequences is calculated. Finally, the symbolic dynamics entropy value is calculated based on the joint probability distribution. The feature combination is to concatenate the standardized original values of each risk indicator, the calculated value change trend characteristics, and the real-time association characteristics to generate a standardized current moment risk state vector.

4. The communication engineering full-process safety production risk dynamic assessment method according to claim 3, characterized in that: In step S2, the online learning model generates an instant reward signal based on the real-time risk event monitoring result. The specific operation is as follows: First, input the current moment risk state vector into an online learning model. The online learning model maps the risk state into a comprehensive risk score. The calculation method of the comprehensive risk score is: perform dot product operation on the current moment short-term weight vector and the current moment risk state vector, and add a bias term. Subsequently, the instant reward signal is generated. When the system detects the occurrence of a preset level of risk event, an instant negative reward signal is automatically generated. The value of the negative instant reward signal is equal to the negative risk event severity level divided by the severity reference value. The risk event severity level is from level one to level ten, and the severity reference value is five. When the system has no risk event report of any preset level for six consecutive hours, a positive instant reward signal is automatically generated. The value of the positive instant reward signal is equal to the actual safety duration divided by the maximum monitoring window of 24 hours.

5. The communication engineering full-process safety production risk dynamic assessment method according to claim 4, characterized in that: In step S2, the model parameters are updated adaptively based on the instant reward signal, and the short-term weight vector is output. The specific operation is as follows: First, construct a loss function. The loss function includes three parts. The first part is half of the squared error between the predicted comprehensive risk score and the actual risk label. The actual risk label is 1 when there is a risk event and 0 when there is no risk event. The second part is a regularization term, which is composed of a regularization coefficient multiplied by half of the squared Euclidean norm of the short-term weight vector. The regularization coefficient is preset to 0.

01. The third part is the reward influence term, which is composed of a reward influence coefficient multiplied by the instant reward signal. The reward influence coefficient is preset to 0.

1. Then, the short-term weight vector is updated by using an online gradient descent algorithm, and the update amount is the learning rate multiplied by the negative value of the gradient of the loss function with respect to the short-term weight vector, the learning rate is preset as 0.005, wherein the gradient of the loss function with respect to the short-term weight vector includes a prediction error term, a regularization term and a reward signal gradient term, the reward signal gradient term is simplified to zero in actual calculation, and the updated short-term weight vector is output as a new short-term weight vector.

6. The communication engineering full-process safety production risk dynamic assessment method according to claim 5, characterized in that: In step S3, the time sequence including the risk state vectors of the current moment and the historical moments is input into the time sequence analysis model to extract the time sequence pattern features, and the specific operation includes: Firstly, a time sequence network based on a multi-head attention mechanism is constructed, the input of the time sequence network is the risk state vector sequence including the current moment and the historical moments from step S1, and the number of continuous time points of the risk state vector sequence is preset as 30-100; Secondly, the multi-head attention mechanism is configured with 4-16 attention heads, the feature dimension of each attention head is set as an integer between 64 and 512, the input risk state vector sequence is subjected to feature transformation and weighted fusion by calculating a query matrix, a key matrix and a value matrix; Finally, the features weighted and fused by the attention mechanism are subjected to nonlinear transformation by a multilayer perceptron to output a time sequence pattern feature vector, wherein the multilayer perceptron includes 1-3 hidden layers, the number of neurons in each layer is preset as 128-512, and the time sequence pattern feature vector is used for feature input of a subsequent reinforcement learning framework.

7. The communication engineering full-process safety production risk dynamic assessment method according to claim 6, characterized in that: In step S3, the reinforcement learning framework is used to take the communication engineering phase as the state space, the weight allocation scheme as the action space, and the safety performance score of the engineering milestone as the long-term reward signal, and the specific operation of outputting the long-term weight vector through the collaborative learning of the time sequence analysis model and the reinforcement learning framework is as follows: An Actor-Critic reinforcement learning framework is used, firstly, the whole process of the communication engineering is divided into 5-10 discrete phases to define the state space, and each phase corresponds to a unique state identifier; Secondly, different long-term weight vector allocation schemes are defined as the action space, the dimension of the long-term weight vector is consistent with the number of risk indicators, and the value range of each component of the long-term weight vector is 0-1, which represents the importance weight of the risk indicator, and the action selection is based on the output of the policy network; Thirdly, the safety performance score calculated based on the engineering milestone is used as the long-term reward signal, and the safety performance score is obtained by weighted summation of the accident occurrence rate score, the number of illegal operation scores and the equipment failure time length score, and the weight proportion of the three is preset as 1:2:1; Finally, the Actor network outputs the probability distribution of the action, and the Critic network evaluates the value function of the state, wherein the Actor network and the Critic network both adopt a 3-layer fully connected neural network structure, and the number of neurons in each layer is 256, the reinforcement learning framework is collaboratively trained by using the long-term reward signal and the time sequence pattern feature vector to output the long-term weight vector.

8. The communication engineering full-process safety production risk dynamic assessment method according to claim 7, characterized in that: In step S4, the specific operation of calculating the short-term confidence for the short-term weight vector includes: The accuracy rate of the risk event prediction by the online learning model in a preset time window of a recent time length is calculated to obtain a basic confidence; The calculation of the accuracy rate adopts a time decay weighting method, so that the prediction result closer to the current time is given a higher weight; Meanwhile, a correction factor based on the statistical characteristics of the prediction error is introduced to correct the basic confidence to output a final short-term confidence.

9. The communication engineering full-process safety production risk dynamic assessment method according to claim 8, characterized in that: The specific operation of calculating the long-term confidence for the long-term weight vector in step S4 includes: The information entropy of the probability distribution output by the time series network for the current engineering stage is calculated to obtain the basic confidence; And a correction term based on the consistency of the current judgment and the historical judgment is introduced to enhance the basic confidence to output the final long-term confidence.

10. The communication engineering full-process safety production risk dynamic assessment method according to claim 9, characterized in that: In step S4, based on the preset confidence threshold, the short-term weight vector and the long-term weight vector are selectively fused to output the final weight vector for risk dynamic assessment, and the specific operation is: The weight fusion decision adopts a three-level decision mechanism based on the confidence threshold, which specifically includes: When the short-term confidence is higher than the first threshold, the short-term weight vector is directly used as the final weight vector output; When the short-term confidence is lower than the second threshold and the long-term confidence is higher than the third threshold, the long-term weight vector is directly used as the final weight vector output; When the above two conditions are not met, the short-term weight vector and the long-term weight vector are weighted and summed according to an adaptively calculated fusion weight, and the summation result is taken as the final weight vector output.

Images