Urban underground comprehensive pipe gallery electrical fire early warning method

Abstract

This invention discloses a method for early warning of electrical fires in urban underground utility tunnels, comprising: Step 1, collecting monitoring data for electrical fire early warning through sensing devices installed at key nodes of the underground utility tunnel; Step 2, preprocessing the monitoring data to obtain a standardized time-series dataset; Step 3, constructing a basic risk assessment model based on a Bayesian network and calculating the basic risk probability; Step 4, constructing a trend risk prediction model based on an LSTM neural network and performing model training and optimization; Step 5, concatenating the basic risk probability with the standardized time-series data and inputting it into the trained trend risk prediction model to output the trend risk probability; Step 6, weighted fusion of the basic risk probability and the trend risk probability to obtain the final fire prediction probability; Step 7, setting a risk probability grading threshold; determining the early warning level based on the final fire prediction probability and generating corresponding early warning information.

Description

Technical Field

[0001] This invention belongs to the field of electrical fire monitoring and early warning technology, and specifically relates to an early warning method for electrical fires in urban underground utility tunnels. Background Technology

[0002] Urban underground utility tunnels, as underground infrastructure integrating various pipelines such as electricity, communications, and water supply and drainage, are the "lifeline" ensuring the normal operation of a city. Among them, the electrical system, as the core of the tunnel's power supply, directly affects the overall safety of the tunnel. Due to the characteristics of underground utility tunnels, such as enclosed spaces, poor ventilation, dense equipment, and humid and dusty environments, electrical circuits are constantly under complex operating conditions, making them prone to faults such as insulation aging, poor contact, overload, and short circuits, which can lead to electrical fires. Once a fire occurs, the fire spreads rapidly, smoke is difficult to disperse, rescue is extremely difficult, and it will cause serious casualties and property damage.

[0003] Existing electrical fire monitoring and early warning methods mostly employ fixed threshold alarm mechanisms, such as setting alarm thresholds based on national standards for residual current of 300-500mA and temperature of 70℃. However, the electrical equipment in underground utility tunnels is diverse, with large fluctuations in operating loads and complex and variable environmental parameters. A single fixed threshold is insufficient to meet the risk identification needs under different operating conditions: raising the threshold easily leads to missed alarms, ignoring potential fire hazards; lowering the threshold easily triggers frequent false alarms, interfering with operation and maintenance work. In addition, traditional monitoring methods often rely on alarms based on single characteristic parameters, failing to consider the multi-factor coupling characteristics of electrical fire occurrence, making it difficult to accurately capture the fault evolution process, and resulting in insufficient reliability of early warning. Summary of the Invention

[0004] The purpose of this invention is to provide an early warning method for electrical fires in urban underground utility tunnels, which can achieve accurate early warning of electrical fire risks and reduce the false alarm and missed alarm rates.

[0005] The technical solution provided by this invention is as follows:

[0006] A method for early warning of electrical fires in urban underground utility tunnels includes:

[0007] Step 1: Collect monitoring data for electrical fire early warning by installing sensing devices at key nodes of the underground utility tunnel;

[0008] Step 2: Preprocess the monitoring data to obtain a standardized time-series dataset;

[0009] Step 3: Construct a basic risk assessment model based on Bayesian networks and calculate the basic risk probability;

[0010] Step 4: Construct a trend risk prediction model based on an LSTM neural network, and train and optimize the model.

[0011] Step 5: Concatenate the basic risk probability with the standardized time series data, input it into the trained trend risk prediction model, and output the trend risk probability.

[0012] Step 6: Weight and fuse the basic risk probability and the trend risk probability to obtain the final fire prediction probability;

[0013] Step 7: Set risk probability classification thresholds; determine the warning level based on the final fire prediction probability, and generate corresponding warning information.

[0014] Preferably, the sensing device includes: a current sensor, a voltage sensor, a temperature sensor, a fault arc detector, and a humidity sensor; the monitoring data includes: operating current, voltage, line temperature, fault arc signal, and ambient humidity data.

[0015] Preferably, the preprocessing of the monitoring data includes: missing data completion, data cleaning, normalization, and feature extraction.

[0016] Preferably, the specific process of constructing a basic risk assessment model based on Bayesian networks is as follows: based on expert weighting and historical electrical fire monitoring data statistics, the prior probability, conditional probability, and feature weights of the basic risk are obtained; based on the full sample data of historical electrical fires, the marginal probability of the features is obtained; and through the prior probability, conditional probability, feature weights, and marginal probability of the features, a formula for calculating the basic risk probability is constructed.

[0017] Preferably, the formula for calculating the basic risk probability is:

[0018] ,

[0019] In the formula, Basic risk probability; The prior probability of basic risk; For conditional probability; For the first Marginal probabilities of class features; For feature weights; This represents the basic risk status in the event of an electrical fire. The product symbol is used. The feature sequence is the standardized time-series dataset. For the first Class features; For a specific moment; This represents the total number of feature categories.

[0020] Preferably, the trend risk prediction model based on the LSTM neural network includes an input layer, an LSTM hidden layer, and an output layer.

[0021] Preferably, in step four, the trend risk prediction model is trained and optimized using a reinforcement learning algorithm. The specific steps are as follows:

[0022] Step 1: Using the trend risk prediction model as the agent, the electrical time series dataset as the interaction environment, the hidden layer state of the LSTM neural network at the previous moment, the cell state, the basic risk probability at the current moment, and the standardized time series data as the state, the weight and bias update of the LSTM neural network as the action, and the trend risk prediction error as the core to construct the reward function, and construct the current evaluation network and the target evaluation network.

[0023] Step 2: Initialize the parameters of the LSTM neural network, the current evaluation network parameters of the deep Q network, and the target evaluation network parameters using a random normal distribution, and synchronize the initialized current evaluation network parameters to the target evaluation network parameters;

[0024] Step 3: The intelligent agent interacts with the environment and stores the state, action, reward and the state of the next time step as samples in the experience replay pool. When the number of samples in the experience replay pool reaches the preset sample size, a batch of samples is randomly selected for training.

[0025] Step 4: Calculate the target Q-value of the evaluation network based on the extracted batch samples and construct the loss function; update the LSTM neural network parameters using the gradient descent algorithm;

[0026] Step 5: Every preset number of steps, update the target evaluation network parameters to the current evaluation network parameters, and repeat steps 3 to 4 until the cumulative reward is stable, the loss function value is lower than the preset threshold, or the number of iterations reaches the preset upper limit. At this point, the parameters are considered to have converged, the iteration is stopped, the optimal parameters are output, and the training and optimization of the trend risk prediction model is completed.

[0027] Preferably, the formula for calculating the trend risk probability is:

[0028] P ( R dynamic , t + 1 ) = σ [ W h * ⋅ h t + b h * ] ,

[0029] In the formula, This represents the probability of trend risk. This represents the trend risk status in the event of an electrical fire. Use the Sigmoid activation function; These are the optimal weights for the LSTM neural network; This represents the optimal bias for the LSTM neural network. This is the output of the hidden layer.

[0030] Preferably, the formula for calculating the final fire prediction probability is:

[0031] ,

[0032] In the formula, The final fire prediction probability; Basic risk weights; This represents the trend risk weight.

[0033] 10. The method for early warning of electrical fires in urban underground utility tunnels according to claim 9, wherein the risk probability classification threshold includes:

[0034] when At that time, it was determined to be in a normal state with no fire hazard;

[0035] when When the situation is deemed to be in a warning state, indicating a potential electrical fire hazard, staff should be alerted to conduct an inspection.

[0036] when When the alarm is triggered, it is determined to be an emergency alarm state, indicating a serious electrical fire hazard with an immediate risk of ignition. It is necessary to trigger an audible and visual alarm and cut off the relevant power supply.

[0037] The beneficial effects of this invention are:

[0038] The electrical fire early warning method for urban underground integrated pipe corridors provided by this invention can achieve accurate early warning of electrical fire risks and reduce the false alarm and missed alarm rates. Attached Figure Description

[0039] Figure 1 This is a flowchart of the electrical fire early warning method for urban underground utility tunnels according to the present invention.

[0040] Figure 2 This is an alarm probability diagram comparing the urban underground integrated pipe gallery electrical fire early warning method described in this invention with traditional threshold alarms.

[0041] Figure 3 This diagram illustrates the alarm frequency of the urban underground integrated pipe gallery electrical fire early warning method described in this invention compared to traditional threshold alarms in practical applications. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0043] like Figure 1 As shown, this invention provides a method for early warning of electrical fires in urban underground utility tunnels, and the specific implementation process is as follows:

[0044] Step 1: Deploy multiple sensing devices at key electrical equipment nodes in the underground utility tunnel. These key nodes include cable joints, distribution boxes, and transformers. The sensing devices include current sensors, voltage sensors, temperature sensors, arc fault detectors, and humidity sensors. These devices collect real-time monitoring data for electrical fire early warning within continuous time steps. The monitoring data includes operating current, voltage, line temperature, arc fault signals, and ambient humidity data. Simultaneously, basic data for electrical fire early warning is collected as supplementary information. This basic data includes equipment installation location, system configuration, unit information, expert weighting, historical electrical fire monitoring data, and a full sample of historical electrical fire data. This basic data can be improved through manual annotation.

[0045] Step 2: Preprocess the monitoring data, performing missing data completion, data cleaning, normalization, and feature extraction operations in sequence to obtain a standardized time-series dataset. The specific steps are as follows:

[0046] Step 1: Complete the monitoring data to restore its integrity; specific operations include: filling missing monitoring types with linear interpolation or -1; and completing time series data with the value from the previous time step.

[0047] Step 2: Clean the monitoring data to ensure its validity; specific operations include: removing noisy data, abnormal jumps, and illegal data.

[0048] Step 3: Normalize the monitoring data to eliminate the dimensional differences between different dimensions, ensuring the data is uniformly mapped to a reasonable range; the normalization formula is:

[0049] ,

[0050] In the formula, For the first Class features in the first The normalized value at time; For a moment, ; This represents the total number of steps in time. For the first Class features, ; The total number of feature categories; For the first Class features in the first The original data collected at that moment; For the first The minimum range of the class feature, i.e. the minimum measurable value of the sensor; For the first The maximum safe value of a class of characteristics, i.e. the upper limit of electrical safe operation, is determined by industry standards, equipment ratings, or safety specifications.

[0051] Step 4, the operation of feature extraction of the monitoring data includes: extracting electrical fire characteristic parameters such as current, voltage, temperature, power, and electric arc according to the time window.

[0052] Step 5: Construct a standardized time-series dataset, which is as follows:

[0053] ,

[0054] In the formula, The feature sequence is the standardized time-series dataset. In the first The first type of eigenvalue at time t; In the The first moment Class feature values.

[0055] Step 3: Construct a basic risk assessment model based on Bayesian networks and calculate the real-time basic risk probability. The specific steps are as follows:

[0056] Step 1: Based on expert weighting and statistical analysis of historical electrical fire monitoring data, obtain the basic risk prior probability, conditional probability, and feature weights; wherein, the basic risk prior probability is the initial risk probability of an electrical system fire when there is no feature input; the conditional probability is the probability of a fire occurring in a given static fire risk state. The probability of a class feature occurring; the feature weight is the influence weight of each type of feature on electrical fires, and the total weight is 1.

[0057] Step 2: Based on the full sample data of historical electrical fires, obtain the marginal probability of the feature; the marginal probability of the feature is the probability of the feature occurring in all operating conditions.

[0058] Step 3: Construct the calculation formula for the real-time basic risk probability using the basic risk prior probability, the conditional probability, the feature weights, and the marginal probabilities of the features:

[0059] ,

[0060] In the formula, For the first The basic risk probability at any given moment; The prior probability of basic risk; For conditional probability; For the first Marginal probabilities of class features; For feature weights; This represents the basic risk status in the event of an electrical fire. The product symbol is used.

[0061] The standardized time-series dataset is input into the calculation formula for the real-time basic risk probability to obtain the first... The basic risk probability at any given moment.

[0062] Step 4: Verify that the value range of the basic risk probability at each time step is within... [ 0 , 1 ] Within the interval, after correcting for outliers, the basic risk probability sequence is obtained as follows:

[0063] ,

[0064] In the formula, In order to be in The basic risk probability at any given moment; In order to be in The basic risk probability at any given moment; In order to be in The basic risk probability at any given moment.

[0065] Step 4: Construct a trend risk prediction model based on an LSTM neural network, and train and optimize the trend risk prediction model using a reinforcement learning algorithm; wherein, in this invention, the reinforcement learning algorithm adopts a deep Q-network (DQN); the trend risk prediction model constructed based on the LSTM neural network includes: an input layer, an LSTM hidden layer, and an output layer.

[0066] The trend risk prediction model is trained and optimized using a deep Q-network. The specific steps are as follows:

[0067] Step 1: Using the trend risk prediction model as the agent, the electrical time-series dataset as the interaction environment, the previous hidden layer state and cell state of the LSTM neural network, the current basic risk probability, and the standardized time-series data as the states, the weight and bias update of the LSTM neural network as the actions, and the trend risk prediction error as the core to construct a reward function, thereby constructing the current evaluation network and the target evaluation network. The electrical time-series dataset is composed of the standardized time-series dataset and the basic risk probability sequence.

[0068] The state is as follows:

[0069] S t = [ h t − 1 , C t − 1 , P ( R static , t ) , F 1 , t , ⋯ F n , t ] ,

[0070] In the formula, For the first The state at any given moment; The features output by the hidden layer at the previous time step; This represents the cell state at the previous moment.

[0071] The action is as follows:

[0072] A t = [ Δ W f , Δ W i , Δ W c , Δ W o , Δ W h , Δ b f , Δ b i , Δ b c , Δ b o , Δ b h ] ,

[0073] In the formula, For the first Actions at any given moment; This represents the weight update amount for the LSTM neural network. This represents the bias update amount for the LSTM neural network.

[0074] The reward function is:

[0075] R t =− | P ( R dynamic , t + 1 pred ) − P ( R dynamic , t + 1 true ) | + α ⋅ E [ P ( R dynamic , t + 1 pred ) ] ,

[0076] In the formula, For the first Momentary rewards; The probability of trend risk predicted by the LSTM network; This represents the true trend risk probability marked using historical data. This represents the trend risk status in the event of an electrical fire. This is the regularization coefficient, usually taken as 0.1; Let be the indicator function, if the predicted trend risk probability The value is set to 1 in the range [0,1], otherwise it is set to -10, to avoid the predicted trend risk probability deviating from a reasonable range.

[0077] Step 2: Initialize the parameters of the LSTM neural network, the current evaluation network parameters of the deep Q network, and the target evaluation network parameters using a random normal distribution, and synchronize the initialized current evaluation network parameters to the target evaluation network parameters.

[0078] The parameters of the initialized LSTM neural network are:

[0079] θ 0 = [ W f 0 , W i 0 , W c 0 , W o 0 , W h 0 , b f 0 , b i 0 , b c 0 , b o 0 , b h 0 ] ,

[0080] In the formula, The parameters for initializing the LSTM neural network; The weights of the initialized LSTM neural network; The bias for initializing the LSTM neural network.

[0081] The operation to synchronize the initialized current evaluation network parameters to the target evaluation network parameters is as follows:

[0082] ,

[0083] In the formula, For the current evaluation of network parameters; Evaluate network parameters for the target.

[0084] Step 3: The agent interacts with the environment. In each time step, the agent determines the interaction based on the current state. and Greedy strategy for choosing actions ;in, This is a probability parameter, typically starting at 1.0 and gradually decreasing to 0.01 or 0.1 to balance exploration and exploitation; [This refers to the action to be performed.] Rewards for receiving environmental feedback and the state in the next moment The sample consists of the state, action, reward at each time step, and the state at the next time step. The samples are stored in the experience replay pool to avoid the influence of sample correlation on the optimization effect; when the number of samples in the experience replay pool reaches the preset sample size, a batch of samples are randomly selected for DQN training.

[0085] Step 4: Calculate the target Q value of the evaluation network based on the extracted batch samples, construct the loss function, evaluate the difference between the Q value and the target Q value, and update the LSTM neural network parameters using the gradient descent algorithm.

[0086] The objective Q-value of the evaluation network is calculated as the optimization objective, and the calculation formula is as follows:

[0087] ,

[0088] In the formula, The target Q value; A function for calculating the Q value; As a discount factor, ; The state at the next moment; For the action in the next moment; The state at the next moment The maximum Q value is calculated from the parameters of the target evaluation network.

[0089] The loss function is:

[0090] L ( θ ) = 1 B ∑ m = 1 B [ y m − Q ( S m , A m ; θ ) ] 2 ,

[0091] In the formula, The loss function; The number of samples to be drawn in a batch is typically between 32 and 128. For the first The target Q value for each sample; For the first One sample; For the first The predicted Q value for each sample is calculated using the current evaluation network parameters;

[0092] Calculate the loss function For parameters The gradient is:

[0093] ∇ θ L ( θ ) = 2 B ∑ m = 1 B [ y m − Q ( S m , A m ; θ ) ] ∇ θ Q ( S m , A m ; θ ) ,

[0094] In the formula, This is the gradient operator, used to calculate the derivative of a function at a point. The target Q value;

[0095] Update parameters using gradient descent algorithm :

[0096] ,

[0097] In the formula, The learning rate is typically set to 0.01, which controls the step size of parameter updates and avoids oscillations.

[0098] Step 5: Every preset number of steps, update the target evaluation network parameters to the current evaluation network parameters. Repeat steps 3 to 4 until the cumulative reward is stable, the loss function value is lower than the preset threshold, or the number of iterations reaches the preset upper limit. At this point, the parameters are considered to have converged, the iteration is stopped, the optimal parameters of the LSTM neural network are output, and the training and optimization of the trend risk prediction model is completed.

[0099] The preset number of steps is usually 100. The preset threshold is usually 0.001. The cumulative reward is:

[0100] ,

[0101] In the formula, For cumulative reward function; In order to be in Instant rewards earned at any time; Index of time steps for future rewards.

[0102] The optimal parameters for the output LSTM neural network are:

[0103] θ * = [ W f * , W i * , W c * , W o * , W h * , b f * , b i * , b c * , b o * , b h * ] ,

[0104] In the formula, These are the optimal parameters for the LSTM neural network; These are the optimal weights for the LSTM neural network; This represents the optimal bias for the LSTM neural network.

[0105] Step 5: Concatenate the basic risk probability with the standardized time series data and input it into the trained trend risk prediction model to output the trend risk probability. The basic risk probability and the standardized time series data are concatenated by aligning them along time steps.

[0106] The specific steps to obtain the predicted trend risk value are as follows:

[0107] Step 1: The initial hidden layer state and initial cell state of the trained trend risk prediction model are set up so that... and This provides initial conditions for time series prediction; among which, This is the initial hidden layer state; This represents the initial cell state.

[0108] Step 2, for the first At that moment, the input features of the trained trend risk prediction model are:

[0109] X t = [ P ( R static , t ) , F 1 , t , ⋯ F n , t ] T ,

[0110] In the formula, For the first The input features at any given time include basic risk information at the current time, as well as multi-dimensional dynamic feature information.

[0111] Step 3: Substitute the optimal parameters obtained in Step 4 into the trained trend risk prediction model to obtain:

[0112] f t = σ [ W f * ⋅ ( h t − 1 , X t ) + b f * ] ,

[0113] i t = σ [ W i * ⋅ ( h t − 1 , X t ) + b i * ] ,

[0114] C ˜ t = tanh [ W c * ⋅ ( h t − 1 , X t ) + b c * ] ,

[0115] ,

[0116] o t = σ [ W o * ⋅ ( h t − 1 , X t ) + b o * ] ,

[0117] ,

[0118] In the formula, Output for the forget gate; For input gate output; Candidate cell state; In cellular state; Output gate output; Output for hidden layer; The Sigmoid activation function maps the output to the [0,1] interval, adapting to gated switch logic; The hyperbolic tangent activation function maps the output to the [-1,1] interval, enriching the information expression of cell state; It is an element-level product, which enables gating to perform weighted filtering of information.

[0119] Step 4: Output based on hidden layer ,calculate Probability of trend risk at any given moment:

[0120] P ( R dynamic , t + 1 ) = σ [ W h * ⋅ h t + b h * ] ,

[0121] In the formula, For the predicted first The probability of trend risk at any given moment.

[0122] Step 5: Verify each time step. The range of values ​​is all within [ 0 , 1 ] Within the interval, after correcting outliers, the hidden layer state and cell state of the trained trend risk prediction model are updated, resulting in the following trend risk probability sequence:

[0123] ,

[0124] In the formula, In order to be in The probability of trend risk at any given moment; In order to be in The probability of trend risk at any given moment; In order to be in The probability of trend risk at any given moment.

[0125] Step Six: Weight and fuse the basic risk probability and the trend risk probability to obtain the final fire prediction probability. The specific steps are as follows:

[0126] Step 1: Set the basic risk weights Trend risk weight is Prioritize dynamic temporal risks to meet the real-time early warning requirements for electrical fires.

[0127] Step 2: When conducting an overall risk assessment for only a single electrical system, calculate the... Final fire prediction probability at time:

[0128] ,

[0129] In the formula, For the first The final fire prediction probability at any given time; This describes the risk status in the event of an electrical fire.

[0130] When a combined assessment is required for multiple specific risks such as over-temperature, arc faults, and residual current, the extended formula is used:

[0131] P z ( R fire , t + 1 ) = ∑ j = 1 n [ β P ( R static , j , t ) + ( 1 − β ) P ( R dynamic , j , t + 1 ) ⋅ W ′ j ] ,

[0132] In the formula, For the first The final fire prediction probability for various specific risks at any given time; For the first Class features in the first The basic risk probability at any given moment; For the first Class features in the first The probability of trend risk at any given moment; For the first Weights of class-specific risks.

[0133] Step 7: Set risk probability grading thresholds; determine the warning level based on the final fire prediction probability and generate corresponding warning information. Based on actual engineering needs, preset three warning thresholds to achieve graded warnings for electrical fires: when... When the condition is normal, there is no fire hazard; when When the situation is deemed a warning, indicating a potential electrical fire hazard, staff should be alerted to conduct an inspection; when When the alarm is triggered, it is determined to be an emergency alarm state, indicating a serious electrical fire hazard with an immediate risk of ignition. It is necessary to trigger an audible and visual alarm and cut off the relevant power supply.

[0134] like Figure 2 As shown, the raw data undergoes standardized preprocessing: based on the line's safe operating parameters—rated voltage 220V, rated current 10A, and maximum power 2200W—the electrical parameters are normalized to the [0, 1] interval; unknown feature parameters are represented by -1, and missing data is supplemented using the pre-value method. The electrical parameters at each moment are calculated using a Bayesian network to obtain a basic risk probability value. This basic risk probability value is combined with the original perceived features to form a new input feature vector, which serves as the input to a trend risk prediction model built on an LSTM neural network. The trend risk prediction model built on the LSTM neural network captures the dynamic characteristics of electrical parameters evolving over time through time-series modeling analysis, and iteratively updates the risk probability by combining the static probability value, ultimately outputting a real-time risk probability assessment result.

[0135] Table 1. Partial electrical fire verification data:

[0136] Record number Voltage / V Current / A Power / W Arc mark Temperature / °C Remark 1 221.3 0.28 61.964 0 25.1 Normal no-load 2 220.8 0.29 64.032 0 25.3 Normal no-load 3 219.7 0.31 68.107 0 25.5 Normal no-load 4 197.2 5.83 1149.670 1 25.7 Arc triggering 5 220.5 0.45 99.225 0 25.9 steady state operation 6 220.1 0.43 94.643 0 25.5 steady state operation 7 219.8 0.42 92.316 0 25.5 steady state operation

[0137] The preprocessed data was verified using partial electrical operation data, based on Table 1 and... Figure 2 It can be seen that when a real electric arc occurs and the arc indicator = 1, the prediction probability of the trend risk prediction model built based on the LSTM neural network is as high as 0.318, indicating that the Bayesian model has the ability to comprehensively analyze risks. When only a single indicator alarms, it does not mean that the overall risk level is very high, demonstrating its strong positive sample recognition ability. During the no-load or steady-state operation, the prediction probability is very low, showing that the trend risk prediction model built based on the LSTM neural network combined with the LSTM neural network can accurately exclude non-arc situations. When comparing the present invention with traditional threshold alarms, it was found that in the continuous monitoring of multiple characteristic parameters such as voltage, current, power, arc, and temperature, when only a single characteristic parameter shows an instantaneous abnormal state in a very short time, the traditional threshold alarm method will generate an electrical fire alarm signal, and the number of alarm signals is proportional to the instantaneous abnormal state. However, the results of the electrical fire early warning method for urban underground integrated pipe corridors provided by the present invention show that the probability of electrical fire risk is extremely low at this time, so no alarm event is triggered, effectively avoiding frequent alarms and improving the feasibility of the comprehensive electrical fire early warning platform.

[0138] like Figure 3 As shown, taking an underground utility tunnel in a certain city as a pilot project, the electrical fire early warning method for underground utility tunnels provided by this invention was compared with the traditional electrical fire early warning method. The pilot project recorded 87 instances of abnormal data for a single characteristic parameter, of which 12 were in a continuous abnormal state, and 75 of them returned to normal immediately after triggering an alarm due to momentary abnormalities. Using the traditional electrical fire early warning method, an alarm notification event was triggered every time a momentary abnormality exceeded a set threshold, resulting in 87 alarms. However, the electrical fire early warning method for underground utility tunnels provided by this invention did not generate alarm notification events for the 75 momentary abnormal states, significantly reducing the interference caused by false alarms. Therefore, the electrical fire early warning method for underground utility tunnels provided by this invention is significantly superior to the traditional electrical fire early warning method in terms of false alarm suppression.

[0139] This invention provides an electrical fire early warning method for urban underground integrated pipe corridors, which can achieve accurate early warning of electrical fire risks and reduce false alarm and missed alarm rates. It collects multi-dimensional data on electrical system operation and environment through various sensing devices, and combines a series of processing steps such as missing data completion, data cleaning, and normalization to ensure data quality and provide a reliable data foundation for accurate early warning. It integrates Bayesian network basic risk assessment with LSTM neural network trend risk prediction, considering both the basic risk characteristics accumulated from historical data and capturing the dynamic changes in real-time operating status, comprehensively characterizing the multi-factor coupling characteristics of electrical fires. It uses reinforcement learning algorithms to train and optimize the LSTM neural network, effectively improving the model's convergence speed and prediction accuracy through mechanisms such as experience replay and dual-network architecture, reducing prediction errors. It integrates basic risk probability and trend risk probability, overcoming the limitations of single-feature assessment and significantly reducing the false alarm and missed alarm rates of fire prediction. It sets three-level early warning thresholds to generate early warning information including risk level, location, and handling suggestions, avoiding false alarm and missed alarm problems caused by fixed thresholds and providing clear handling guidance for operation and maintenance personnel, improving the practicality and operability of the early warning.

[0140] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for early warning of electrical fires in urban underground utility tunnels, characterized in that, include: Step 1: Collect monitoring data for electrical fire early warning by installing sensing devices at key nodes of the underground utility tunnel; Step 2: Preprocess the monitoring data to obtain a standardized time-series dataset; Step 3: Construct a basic risk assessment model based on Bayesian networks and calculate the basic risk probability; Step 4: Construct a trend risk prediction model based on an LSTM neural network, and train and optimize the model. Step 5: Concatenate the basic risk probability with the standardized time series data, input it into the trained trend risk prediction model, and output the trend risk probability. Step 6: Weight and fuse the basic risk probability and the trend risk probability to obtain the final fire prediction probability; Step 7: Set risk probability classification thresholds; determine the warning level based on the final fire prediction probability, and generate corresponding warning information.

2. The method for early warning of electrical fires in urban underground utility tunnels according to claim 1, characterized in that, The sensing devices include: a current sensor, a voltage sensor, a temperature sensor, a fault arc detector, and a humidity sensor; the monitoring data includes: operating current, voltage, line temperature, fault arc signal, and ambient humidity data.

3. The method for early warning of electrical fires in urban underground utility tunnels according to claim 2, characterized in that, The preprocessing of the monitoring data includes: missing data completion, data cleaning, normalization, and feature extraction.

4. The method for early warning of electrical fires in urban underground utility tunnels according to claim 1, characterized in that, The specific process of constructing a basic risk assessment model based on Bayesian networks is as follows: Based on expert weighting and statistical analysis of historical electrical fire monitoring data, the prior probability, conditional probability, and feature weights of the basic risk are obtained; based on the full sample data of historical electrical fires, the marginal probabilities of the features are obtained; and through the prior probability, conditional probability, feature weights, and marginal probabilities of the features, the calculation formula for the basic risk probability is constructed.

5. The method for early warning of electrical fires in urban underground utility tunnels according to claim 4, characterized in that, The formula for calculating the basic risk probability is as follows: ， In the formula, Basic risk probability; The prior probability of basic risk; For conditional probability; For the first Marginal probabilities of class features; For feature weights; This represents the basic risk status in the event of an electrical fire. The product symbol is used. The feature sequence is the standardized time-series dataset. For the first Class features; For a specific moment; This represents the total number of feature categories.

6. The method for early warning of electrical fires in urban underground utility tunnels according to claim 1, characterized in that, The trend risk prediction model based on LSTM neural network includes: input layer, LSTM hidden layer and output layer.

7. The method for early warning of electrical fires in urban underground utility tunnels according to claim 6, characterized in that, In step four, the trend risk prediction model is trained and optimized using a reinforcement learning algorithm. The specific steps are as follows: Step 1: Using the trend risk prediction model as the agent, the electrical time series dataset as the interaction environment, the hidden layer state of the LSTM neural network at the previous moment, the cell state, the basic risk probability at the current moment, and the standardized time series data as the state, the weight and bias update of the LSTM neural network as the action, and the trend risk prediction error as the core to construct the reward function, and construct the current evaluation network and the target evaluation network. Step 2: Initialize the parameters of the LSTM neural network, the current evaluation network parameters of the deep Q network, and the target evaluation network parameters using a random normal distribution, and synchronize the initialized current evaluation network parameters to the target evaluation network parameters; Step 3: The intelligent agent interacts with the environment and stores the state, action, reward and the state of the next time step as samples in the experience replay pool. When the number of samples in the experience replay pool reaches the preset sample size, a batch of samples is randomly selected for training. Step 4: Calculate the target Q-value of the evaluation network based on the extracted batch samples and construct the loss function; update the LSTM neural network parameters using the gradient descent algorithm; Step 5: Every preset number of steps, update the target evaluation network parameters to the current evaluation network parameters, and repeat steps 3 to 4 until the cumulative reward is stable, the loss function value is lower than the preset threshold, or the number of iterations reaches the preset upper limit. At this point, the parameters are considered to have converged, the iteration is stopped, the optimal parameters are output, and the training and optimization of the trend risk prediction model is completed.

8. The method for early warning of electrical fires in urban underground utility tunnels according to claim 5, characterized in that, The formula for calculating the probability of trend risk is: ， In the formula, This represents the probability of trend risk. This represents the trend risk status in the event of an electrical fire. Use the Sigmoid activation function; These are the optimal weights for the LSTM neural network; This represents the optimal bias for the LSTM neural network. This is the output of the hidden layer.

9. The method for early warning of electrical fires in urban underground utility tunnels according to claim 8, characterized in that, The formula for calculating the final fire prediction probability is as follows: ， In the formula, The final fire prediction probability; Basic risk weights; This represents the trend risk weight.

10. The method for early warning of electrical fires in urban underground utility tunnels according to claim 9, characterized in that, The risk probability grading thresholds include: when At that time, it was determined to be in a normal state with no fire hazard; when When the situation is deemed to be in a warning state, indicating a potential electrical fire hazard, staff should be alerted to conduct an inspection. when When the alarm is triggered, it is determined to be an emergency alarm state, indicating a serious electrical fire hazard with an immediate risk of ignition. It is necessary to trigger an audible and visual alarm and cut off the relevant power supply.

Images