Electrical fire risk early warning method and system based on Internet of Things

Abstract

The invention provides an electrical fire risk early warning method and system based on the Internet of Things, and relates to the technical field of electrical fire risk early warning, and the method comprises the steps: collecting the real-time and historical state data, including operation parameters, line state parameters and environment data, of electrical equipment, constructing a training set containing a historical state structured data set and a risk level; preprocessing real-time data by using edge computing nodes, and eliminating abnormal values to generate a structured data set; and constructing and training a risk assessment model based on a deep learning algorithm, inputting a real-time data sample to obtain a risk level, feeding back the risk level to a user terminal, and sending an early warning signal to operation and maintenance personnel when the risk level does not conform to expectation. The device state can be monitored in real time, the risk level can be accurately evaluated, feedback can be timely sent to the user and the operation and maintenance personnel, the safety and reliability of the device can be improved, and the operation and maintenance cost can be reduced.

Description

Technical Field

[0001] This application relates to the technical field of electrical fire risk early warning, and in particular to an electrical fire risk early warning method and system based on the Internet of Things. Background Technology

[0002] With the increasing electrification of modern society and the widespread application of electrical equipment in various fields, while bringing convenience and other advantages, the risk of electrical fires is also becoming increasingly prominent. Electrical fires are characterized by their suddenness, rapid spread, and high destructiveness, causing significant casualties and property losses, and threatening public safety. Therefore, developing effective methods for early warning of electrical fire risks, and proactively identifying and warning of potential risks, is of great significance for timely prevention, reducing the probability of fires, and minimizing losses.

[0003] Currently, there are various technologies for early warning of electrical fire risks. Common methods are based on electrical parameter monitoring, which involves installing sensors in electrical systems to monitor parameters such as current, voltage, and temperature in real time. For example, current exceeding a threshold may indicate an overload, while abnormal temperature increases may suggest a short circuit or fault. Other methods are based on image recognition, using cameras and image processing algorithms to analyze the appearance and operating status of electrical equipment, detecting abnormal heating, electrical sparks, and other indicators to assess risk. Some systems combine data from multiple sensors, using algorithmic models for comprehensive analysis to improve accuracy.

[0004] However, electrical parameter monitoring methods can only reflect local electrical conditions and are insufficient to comprehensively capture complex fault characteristics. Some faults that initially show no obvious parameter changes may not receive timely warnings. Image recognition methods are easily affected by environmental factors such as light and obstructions, and their ability to identify subtle early fault features is limited. Most existing comprehensive early warning algorithm models are simple, and their accuracy and reliability in the face of complex and ever-changing electrical fire risk scenarios need improvement, failing to meet prevention and control requirements. Summary of the Invention

[0005] To improve the accuracy and reliability of early warning, this application provides an electrical fire risk early warning method and system based on the Internet of Things.

[0006] Firstly, this application provides an electrical fire risk early warning method based on the Internet of Things, employing the following technical solution: An IoT-based method for early warning of electrical fire risks includes the following steps: Collect real-time status data of electrical equipment and construct a training set. The training set includes a structured dataset corresponding to the historical status data of the electrical equipment and risk levels. The status data includes: operating parameters, line status parameters and environmental data. Edge computing nodes are used to preprocess real-time status data, and outliers are removed to generate a structured dataset corresponding to the real-time status data. A risk assessment model is built based on deep learning algorithms. The model is trained using a training set to obtain the trained risk assessment model. Samples from the structured dataset corresponding to the real-time status data are input into the trained risk assessment model to obtain the risk level. The risk level is then fed back to the user terminal, and an early warning signal is issued to the operation and maintenance personnel when the risk level does not meet expectations.

[0007] This application comprehensively collects real-time status data of electrical equipment and constructs a training set. The training set includes historical status data and corresponding risk levels. Edge computing nodes are used to preprocess the real-time status data, enabling timely removal of outliers and generating a structured dataset. Edge computing nodes, located close to the data source, can process data quickly, reducing the time and bandwidth consumption of data transmission to the cloud. The structured dataset facilitates subsequent processing by deep learning algorithms, improving the efficiency of model training and inference.

[0008] This application constructs a risk assessment model based on deep learning algorithms and trains it using a training set. This model can automatically learn the relationship between features and risk levels from structured data corresponding to historical status data, avoiding the subjectivity and limitations of manually set rules. By inputting samples from the structured dataset corresponding to real-time status data into the trained model, risk levels can be quickly obtained, enabling real-time risk assessment.

[0009] Timely feedback of risk levels to user terminals enables users to understand the risk status of electrical equipment in real time, improving maintenance personnel's control over equipment operation and facilitating timely measures to mitigate risks. When the risk level does not meet expectations, early warning signals are issued to maintenance personnel, allowing them to take timely action and improve equipment reliability and availability.

[0010] By adopting the above solution, this application can monitor the equipment status in real time, accurately assess the risk level, and provide timely feedback to users and maintenance personnel, which helps to improve the safety and reliability of the equipment and reduce maintenance costs.

[0011] Optionally, the method further includes: Historical early warning data and actual fire data of electrical equipment are acquired. The historical early warning data includes the early warning time, target equipment ID, and risk level. Based on the historical early warning data and actual fire data, the weight parameters of the risk assessment model are dynamically adjusted through a reinforcement learning algorithm.

[0012] Traditional risk assessment models often use fixed weight parameters, making it difficult to adapt to changes in the operating status of electrical equipment and the environment. This application, however, uses a reinforcement learning algorithm to dynamically adjust the weight parameters, enabling the model to adjust its focus on different features (such as operating parameters and line status parameters) in real time based on feedback from historical early warning data and actual fire data. Over time, more historical early warning data and actual fire data accumulate. The reinforcement learning algorithm can continuously learn and adjust using this new data, continuously improving the performance of the risk assessment model. This continuous optimization characteristic allows the model to keep pace with the development of electrical equipment technology and environmental changes, maintaining a consistently high level of risk assessment accuracy.

[0013] This application utilizes reinforcement learning algorithms to dynamically adjust the weight parameters of the risk assessment model, which has significant advantages such as improving the accuracy of risk assessment, enhancing the model's generalization ability, and enabling continuous optimization, thus effectively improving the performance of electrical equipment risk assessment.

[0014] Optionally, the reinforcement learning algorithm employs a deep Q-network, and the method includes: The state space of reinforcement learning includes historical early warning data and actual fire data, the action space is the weight adjustment range of the risk assessment model, and the reward function is the deviation between the risk level predicted by the model and the actual fire data. Initialize the experience replay pool of the deep Q network and store the samples corresponding to historical early warning data and actual fire data into the experience replay pool; Each iteration randomly selects a batch of samples from the experience replay pool and updates the parameters of the deep Q network using gradient descent to maximize the reward function value, thus completing a single weight parameter adjustment.

[0015] Structured data contains multi-dimensional information such as operating parameters of electrical equipment, line status parameters, and environmental data. This information comprehensively reflects the current operating status of the equipment, providing sufficient decision-making basis for reinforcement learning algorithms. It allows the algorithm to adjust the weights of the risk assessment model based on the actual state of the equipment. The adjustment range of the risk assessment model's weights is used as the action space. By adjusting the weights, the algorithm can dynamically change the model's focus on different features, thereby optimizing the model's predictive performance and enabling it to more accurately assess the risk level of electrical equipment. The deviation between the model's predicted risk level and the actual fire data is used as the reward function. The smaller the deviation, the more accurate the model's prediction, and the greater the reward. This prompts the algorithm to continuously adjust the weight parameters to minimize prediction deviation and improve the accuracy of risk assessment.

[0016] Deep Q-Networks (DQNs) possess powerful learning and adaptive capabilities, continuously learning from historical data in an experience replay pool. In practical applications, the operating environment and state of electrical equipment change over time. DQNs can adjust the weights of the risk assessment model in real time, enabling the model to quickly adapt to these changes and maintain high predictive accuracy. Risk assessment of electrical equipment involves complex relationships between multiple factors, and DQNs can automatically learn these relationships and find the optimal weight adjustment strategy. Compared to traditional methods, DQNs do not require manually pre-setting complex rules and can more effectively handle complex nonlinear relationships.

[0017] Optionally, the method further includes: when the risk level does not meet expectations, using a causal inference algorithm to reverse analyze the influencing factors in the structured data, and feeding the influencing factors back to the user terminal.

[0018] In electrical equipment risk assessment scenarios, a risk level that does not meet expectations indicates a significant potential hazard. This application utilizes a causal inference algorithm to reverse-analyze influencing factors within structured data, enabling the identification of the root causes of high risk at the data level. Causal inference algorithms aim to determine the causal relationships between variables, and structured data contains various operating parameters and environmental data related to electrical equipment. These data exhibit complex causal connections, which the causal inference algorithm can effectively analyze to identify the influencing factors leading to high risk, aligning with the logic of extracting causal relationships from data. By analyzing the influencing factors through causal inference, maintenance personnel can clearly understand which factors have caused the equipment's risk status. Without feedback on influencing factors, maintenance personnel may have to rely on guesswork and trial and error based on experience, easily leading to blind operations that may not only fail to solve the problem but also introduce new risks. Clear feedback on influencing factors can prevent this, improving the efficiency of equipment maintenance and management.

[0019] Optionally, the causal inference algorithm employs a Bayesian network, and the method includes: Each state parameter in the structured data is used as a node in a Bayesian network, and a conditional probability table between nodes is constructed based on the correlation between historical state data and corresponding risk levels. When the risk level output by the trained risk assessment model does not meet expectations, the risk level is used as the result node, and the posterior probability of each state parameter node is calculated through the backward inference of the Bayesian network. State parameters whose posterior probabilities exceed a preset threshold are identified as influencing factors and sorted by posterior probabilities from high to low before being fed back to the user terminal.

[0020] This application constructs a conditional probability table between nodes based on the correlation between historical state data and corresponding risk levels. Historical data contains the actual risk situation of equipment in different states. By analyzing this data, the inherent relationship between each state parameter and the risk level can be uncovered, thereby determining the conditional probabilities between nodes. This allows the Bayesian network to accurately reflect the causal relationship between equipment state and risk. When the risk assessment model outputs a risk level that does not meet expectations, the posterior probability of each state parameter node is calculated using this risk level as the result node, conforming to the logic of causal inference. Through reverse reasoning, the cause can be traced from the result, identifying which state parameters have a significant impact on the current risk level, providing a basis for determining influencing factors.

[0021] By calculating the posterior probability of each state parameter node, the impact of each parameter on the unexpected risk level can be quantified. A higher posterior probability indicates a greater contribution of the state parameter to the current risk level, making it a key factor leading to risk anomalies. This improves the objectivity and accuracy of the analysis results and reduces errors caused by human factors. This application sorts state parameters with posterior probabilities exceeding a preset threshold from high to low and feeds them back to the user terminal, providing users with a clear decision-making order. Users can prioritize and address the state parameters with the greatest impact on risk based on the sorting results, and take targeted measures to reduce risk.

[0022] Optionally, the method further includes: The conditional probability table of the Bayesian network is updated periodically based on newly added historical state data and risk level data. During the update process, if there are samples in the new data that conflict with the current conditional probability table, a weighted average method is used to merge the probability distributions of the conflicting samples and historical samples. The weights are positively correlated with the time decay coefficients corresponding to the samples, and the more recent the sample, the higher the weight.

[0023] This application periodically updates the conditional probability table of the Bayesian network based on newly added historical state data and risk level data. This allows the model to capture these changes in a timely manner, ensuring that the causal relationships reflected by the model are consistent with the actual situation. When there are samples in the new data that conflict with the current conditional probability table, it is reasonable to use a weighted average method to fuse the probability distributions of conflicting samples and historical samples. Because samples from different times have different importance to the current model, samples from more recent times tend to better reflect the current state and risk characteristics of the equipment, so they are given higher weights. The method of determining weights based on the time decay coefficient conforms to the law of data evolution over time, enabling the updated conditional probability table to more accurately reflect the current situation.

[0024] By regularly updating the conditional probability table, the model can incorporate the latest data, thereby more accurately predicting the future risk level of equipment. Over time, the original conditional probability table may develop errors due to data bias or changes in the environment. Regular updates can correct these errors in a timely manner, preventing the accumulation of errors and maintaining the model's high accuracy.

[0025] This application enables the model to better adapt to the dynamic changes in the operation of electrical equipment by regularly updating the conditional probability table of the Bayesian network and using a weighted average method to fuse conflict samples, thus providing users with more reliable risk assessment results.

[0026] Optionally, the risk assessment model includes an input layer, a hidden layer, and an output layer, with the functions of each layer defined as follows: Input layer: Receives structured datasets preprocessed by edge computing nodes, filters and extracts feature parameters related to electrical fire risk through feature engineering module, the feature parameters include current fluctuation coefficient in real-time data, insulation resistance change rate in line status parameters, and temperature and humidity coupling value in environmental data; Hidden layer: An improved long short-term memory network is used to perform temporal correlation analysis on feature parameters to capture the early evolution characteristics of electrical fire risk; Output layer: Based on early evolutionary features, a softmax classifier is used to output risk levels, which include low risk, medium risk and high risk.

[0027] The risk assessment model adopts a classic three-layer architecture: input layer, hidden layer, and output layer. The input layer is responsible for receiving and initially processing the data, the hidden layer performs in-depth feature analysis and correlation mining, and the output layer provides the final risk assessment result. The input layer receives the structured dataset preprocessed by edge computing nodes, fully leveraging the advantage of edge computing nodes being close to the data source. This allows for rapid initial processing of the raw data, reducing the pressure on data transmission to the central server and improving data availability and accuracy. The hidden layer uses an improved Long Short-Term Memory (LSTM) network to perform temporal correlation analysis on the feature parameters. In electrical fire risk assessment, the operating status of equipment and environmental factors change over time. LSTM can analyze the changing trends and interrelationships of these feature parameters at different time points, thereby capturing the early evolution characteristics of electrical fire risk. Based on the early evolution characteristics captured by the hidden layer, the output layer outputs the risk level through a softmax classifier. The softmax classifier maps the input features to different risk level categories and provides the probability of each category, allowing users to quickly understand the degree of electrical fire risk and take appropriate measures.

[0028] Optionally, the improved long short-term memory network includes an input gate, a forget gate, and an output gate, and a residual connection structure is introduced in the input gate, forget gate, and output gate.

[0029] Long Short-Term Memory (LSTM) networks control the inflow, forgetting, and outflow of information through input, forgetting, and output gates to address the vanishing and exploding gradient problems in traditional Recurrent Neural Networks (RNNs), enabling them to better handle long sequences of data. Introducing residual connections provides additional paths for information propagation within the network. This structure allows gradients to propagate more directly in the backpropagation process, avoiding the gradual decay of gradients to zero in deep networks due to excessive layers. This allows the network to learn deeper and more complex feature representations, thereby enhancing its ability to capture complex data patterns.

[0030] In practical applications, training time and inference speed are also important considerations. Because the residual connection structure improves gradient propagation and accelerates network training convergence, the model can complete training in a shorter time. Simultaneously, during the inference phase, the improved model structure is relatively stable, enabling rapid processing of input data and output results, thus improving overall application efficiency.

[0031] Optionally, the hidden layer further includes an attention layer, which is used to assign weights to the feature parameters and use the weighted feature parameters as new feature parameters.

[0032] In practical applications, the input model has numerous feature parameters, but not all features contribute equally to the final prediction or classification result. The attention layer can automatically learn and identify the feature parameters that are more important to the task and assign them higher weights.

[0033] By assigning weights to different feature parameters, the original features are essentially recombined and strengthened. The weights reflect the importance of a feature within the overall data. Using the weighted feature parameters as new feature parameters more accurately expresses the inherent structure and patterns of the data, helping subsequent model layers to better understand and process the data, and improving the model's adaptability to complex data.

[0034] Because the attention layer highlights key features, the model is able to capture information closely related to the task more accurately when making predictions or classifying.

[0035] Secondly, this application provides an electrical fire risk early warning system based on the Internet of Things, which adopts the following technical solution: An IoT-based electrical fire risk early warning system includes: The data acquisition module is used to collect real-time status data of electrical equipment and build a training set; The data processing module is used to preprocess real-time status data using edge computing nodes, remove outliers, and generate a structured dataset. The risk assessment module is used to build a risk assessment model based on deep learning algorithms. The risk assessment model is trained using a training set to obtain the trained risk assessment model. Samples from the structured dataset are input into the trained risk assessment model to obtain the risk level. The risk level is then fed back to the user terminal, and an early warning signal is issued to the operation and maintenance personnel when the risk level does not meet expectations.

[0036] In summary, this application includes at least one of the following beneficial technical effects: 1. This application can monitor the equipment status in real time, accurately assess the risk level, and provide timely feedback to users and maintenance personnel, which helps to improve the safety and reliability of the equipment and reduce maintenance costs.

[0037] 2. This application utilizes reinforcement learning algorithms to dynamically adjust the weight parameters of the risk assessment model, which has significant advantages such as improving the accuracy of risk assessment, enhancing the model's generalization ability, and enabling continuous optimization, and can effectively improve the effect of electrical equipment risk assessment.

[0038] 3. By analyzing influencing factors through causal inference algorithms, maintenance personnel can clearly understand which factors led to the risk status of the equipment. Without feedback on influencing factors, maintenance personnel may only be able to guess and try based on experience, which can easily lead to blind operations. This may not only fail to solve the problem but also cause new risks. Clear feedback on influencing factors can avoid this situation and improve the efficiency of equipment maintenance and management. Attached Figure Description

[0039] Figure 1 This is a flowchart of Embodiment 1 of this application; Figure 2 This is a flowchart of Embodiment 3 of this application; Figure 3 This is a structural schematic diagram of Embodiment 5 of this application. Detailed Implementation

[0040] The following combination Figures 1 to 3 This application will be described in further detail.

[0041] Example 1: This example discloses an electrical fire risk early warning method based on the Internet of Things, referring to... Figure 1The method includes: S11 data acquisition, S12 data processing, and S13 assessment and early warning. Real-time and historical status data of electrical equipment, including operating parameters, line status parameters, and environmental data, are collected to construct a training set containing a structured dataset of historical status and risk levels. Real-time data is preprocessed using edge computing nodes to remove outliers and generate a structured dataset. A risk assessment model is constructed and trained based on a deep learning algorithm. The risk level is obtained by inputting real-time data samples and fed back to the user terminal. If the risk level does not meet expectations, an early warning signal is sent to maintenance personnel. The execution process of each step in this embodiment is as follows: S11 data acquisition collects real-time status data of electrical equipment, including operating parameters, line status parameters, and environmental data.

[0042] Operating parameters include: electrical parameters such as voltage, current, power factor, frequency, and harmonic content; mechanical parameters such as equipment vibration frequency, temperature (e.g., winding temperature, bearing temperature), and speed (for motor-type equipment); and time parameters such as running time, number of start-stop cycles, and load rate change curves.

[0043] Line status parameters include: insulation performance such as insulation resistance, dielectric loss factor, and partial discharge signal; connection status such as contact resistance, joint temperature, and arc detection; and line load such as real-time current distribution, line voltage drop, and three-phase imbalance.

[0044] Environmental data includes: physical environment such as temperature and humidity, air pressure, dust concentration, and corrosive gas content; electromagnetic environment such as electric field strength, magnetic field strength, and radio frequency interference level; and geographical environment such as equipment altitude and seismic zone level (for outdoor equipment).

[0045] Unstructured data (such as log files and raw sensor waveforms) is converted into a time-series database format, and the risk level (low risk, medium risk, high risk) of data samples is reverse-labeled based on historical fault records (such as short circuits and insulation breakdowns).

[0046] S12 data processing utilizes edge computing nodes to preprocess real-time status data, employing statistical methods (such as the 3σ principle and box plot analysis) or domain rules (such as physical limits of equipment parameters) to remove outliers. After removing outliers from the real-time status data, a corresponding structured dataset is generated.

[0047] S13 Assessment and Early Warning: Risk assessment models are built based on deep learning algorithms. The risk assessment model can be any model, such as neural networks or LSTM.

[0048] The risk assessment model is trained using a training set to obtain the trained risk assessment model. Real-time structured data after edge preprocessing is input into the trained model, and the model outputs the risk level of the current device.

[0049] The risk level is fed back to the user terminal, and an early warning signal is issued to the operation and maintenance personnel when the risk level does not meet expectations.

[0050] Example 2: This example differs from Example 1 in that the method further includes: Historical early warning data and actual fire data of electrical equipment are acquired. The historical early warning data includes the early warning time, target equipment ID, and risk level. The actual fire data records the fire occurrence time, equipment ID, and degree of damage, which are used to label the actual risk.

[0051] Based on historical early warning data and actual fire data, the weight parameters of the risk assessment model are dynamically adjusted using a reinforcement learning algorithm. Historical early warning data and actual fire data are matched by time window and device ID to generate labeled training samples. Example: If device A001 issues a high-risk warning at 14:00 and a serious fire occurs at 14:30, the sample label is positive (early warning valid); if no warning is issued but a fire occurs, it is a negative example (false alarm invalid).

[0052] The matched training samples are used as the state input of the reinforcement learning environment. The reinforcement learning algorithm is guided to optimize the weight adjustment strategy through the reward function (such as the penalty value for the model not issuing a warning when a fire occurs, and the penalty value for false alarms). For example, the weight coefficient of the risk level of frequent false alarms is reduced first.

[0053] In this embodiment, the reinforcement learning algorithm employs a deep Q-network, dynamically adjusting the weight parameters of the risk assessment model through the reinforcement learning algorithm, including: The state space of reinforcement learning includes historical early warning data and actual fire data, and the input features of the state space include: Historical structured data of the equipment (such as the average temperature and current over the past 24 hours).

[0054] Early warning statistical characteristics (such as the frequency of high-risk early warnings in the past 7 days).

[0055] Fire history tags (such as whether a fire has occurred in the past 30 days).

[0056] The input features can be represented as: State = [average temperature, average current, number of high-risk warnings, fire history markers].

[0057] The action space represents the weight adjustment range of the risk assessment model, for example: Action = {fully connected layer weights +5%, convolutional layer weights -3%} The reward function is the deviation between the model's predicted risk level and the actual fire data, aiming to minimize this deviation. The specific calculation method is as follows: Fire warning accuracy: including successful fire warning: Reward = +10, and no fire warning: Reward = -20; False alarm control: High-risk warning when there is no fire: Reward = -5; Dynamic weighting: Adjust the reward based on the degree of loss in the actual fire data (e.g., the reward for a severe fire is doubled).

[0058] Initialize the experience replay pool of the deep Q network, and store the samples corresponding to historical early warning data and actual fire data into the experience replay pool. The storage format of the samples is (State, Action, Reward, Next State).

[0059] Each iteration randomly draws a batch of samples from the experience replay pool, updates the parameters of the deep Q-network using gradient descent to maximize the reward function value, and completes a single weight parameter adjustment. The process is as follows: A pre-trained risk assessment model is loaded, and the parameters of a Deep Q-Network (DQN) are initialized simultaneously. A batch of samples is randomly drawn from the experience replay pool, and the target Q-value (current reward + maximum Q-value for the next state) is calculated. The DQN parameters are updated using gradient descent to minimize the loss between the predicted Q-value and the target Q-value. The online DQN parameters are periodically soft-updated to the target network (e.g., 1% update every 100 iterations). The target network is an auxiliary network in the Deep Q-Network (DQN) used to generate the target Q-value. During training, the main network is responsible for selecting actions and calculating Q-values ​​based on the current state, while the target network is used to calculate the maximum Q-value for the next state, thus generating the target Q-value. The presence of the target network helps stabilize the training process and reduces the volatility of Q-value estimation.

[0060] The target network keeps synchronized with the main network through soft updates. That is, every certain number of iterations, the parameters of the main network are copied to the target network at a certain percentage (e.g., 1%). This helps maintain the stability of the target network while allowing the main network to gradually learn better strategies.

[0061] The Q-value is a value assessment of state-action combinations within a reinforcement learning framework. It represents the expected cumulative reward obtained after performing an action in a given state. It is calculated using the Bellman equation, where the Q-value of the current state equals the immediate reward plus a discount to the maximum Q-value of the next state. The calculation model for the Q-value is as follows: ; in, Let be the Q-value of action a given state s; and r be the value of the reward function. This is a discount factor used to balance the importance of immediate rewards and future rewards; The next state; This represents the possible actions in the next state.

[0062] During training, DQN updates the parameters of the risk assessment model by minimizing the loss between the predicted Q-value and the target Q-value.

[0063] In each iteration step, DQN selects the action with the largest Q value as the optimal action based on the current state. After each iteration, the optimal action (weight adjustment magnitude) output by DQN is applied to the risk assessment model. If DQN outputs "fully connected layer weight +5%", the weights of the corresponding layer in the model are updated.

[0064] This embodiment enables DQN to gradually learn the optimal weight adjustment strategy, thereby improving the accuracy of the risk assessment model.

[0065] Example 3: Reference Figure 2 The difference between this embodiment and Embodiment 1 is that the method further includes: S31 analyzes influencing factors. When the risk level does not meet expectations, it uses a causal inference algorithm to reverse analyze the influencing factors in the structured data and feeds the influencing factors back to the user terminal.

[0066] The causal inference algorithm uses a Bayesian network. This step includes: State parameters (such as temperature, current, voltage, and ambient humidity) in structured data are used as nodes in a Bayesian network, and risk levels (low / medium / high) are used as target nodes. For example, the node set = {temperature, current, voltage, humidity, risk level}.

[0067] Based on historical data, statistical analysis is used to determine the conditional dependencies between nodes, for example: If the temperature is greater than 80℃ and the current is greater than 1.2 times the rated value, the probability of a high risk level is 90%. If only the temperature is >80℃, then the probability of a high risk level is 30%.

[0068] CPT is learned from historical data through maximum likelihood estimation or Bayesian estimation.

[0069] When the risk level output by the trained risk assessment model does not meet expectations (i.e., the risk level output by the trained risk assessment model is medium or high risk), the posterior probability of each state parameter node is calculated through back-inference of the Bayesian network, using the risk level as the result node. The process is as follows: The actual risk level (e.g., high risk) is input into the Bayesian network as an observation, and the BeliefPropagation algorithm is used to calculate the posterior probability of each state parameter node. For example, if high risk is observed, the posterior probability P(temperature>80℃|high risk) is calculated for temperature>80℃.

[0070] State parameters whose posterior probabilities exceed a preset threshold are identified as influencing factors. For example, the threshold is set as: posterior probability > 70%. State parameters that exceed the threshold are marked as key influencing factors and sorted by posterior probability: temperature (95%) > current (85%) > humidity (40%). The posterior probabilities are sorted from high to low and then fed back to the user terminal.

[0071] S32 updates periodically based on newly added historical state data and risk level data to update the conditional probability table of the Bayesian network, including: Collect newly added historical status data and corresponding risk levels, and for each node combination (e.g., temperature > 80℃ ∩ current > 1.2 times), calculate its conditional probability in the newly added data.

[0072] For example, in the original CPT, P(high risk | temperature > 80℃, current > 1.2 times) = 90%, and in the new data, the probability of this condition becomes 85%, then the CPT is updated.

[0073] Use the chi-square test to verify whether the dependencies between nodes have changed, and adjust the structure of the Bayesian network as necessary (such as adding nodes or edges).

[0074] S33 handles conflicts. During the update process, if there are samples in the newly added data that conflict with the current conditional probability table, a weighted average method is used to fuse the probability distributions of the conflicting samples and historical samples. The weights are positively correlated with the time decay coefficients corresponding to the samples, and the more recent the sample, the higher the weight. The specific process is as follows: Weight = Time decay coefficient × k, where k is an adjustment coefficient with a value of 0.2-0.5.

[0075] Time decay coefficient The calculation model is as follows: ; in, This is the difference between the sample time and the current time; The attenuation rate is 0.1.

[0076] ; in, For the new conditional probability; Weights for newly added samples; This represents the posterior probability of the newly added sample. The original sample weights; This represents the posterior probability of the original sample.

[0077] Example 4: The difference between this example and Example 1 is that the risk assessment model includes an input layer, a hidden layer, and an output layer, and the functions of each layer are defined as follows: Input layer: Receives the structured dataset preprocessed by the edge computing nodes, filters and extracts feature parameters related to electrical fire risk through the feature engineering module, selects feature parameters based on domain knowledge and statistical correlation (such as Pearson coefficient > 0.3), and uses Min-Max normalization to scale the feature parameters to the [0,1] interval.

[0078] The characteristic parameters include the current fluctuation coefficient in real-time data, the insulation resistance change rate in line status parameters, and the temperature and humidity coupling value in environmental data.

[0079] Hidden layer: An improved long short-term memory network is used to perform temporal correlation analysis on feature parameters to capture the early evolution characteristics of electrical fire risk. The improved long short-term memory network includes an input gate, a forget gate, and an output gate, which are responsible for feature input screening, historical information forgetting and discarding, and final output control, respectively. Residual connection structures are introduced into the input gate, forget gate, and output gate.

[0080] The hidden layer also includes an attention layer, which is used to assign weights to the feature parameters and use the weighted feature parameters as new feature parameters.

[0081] The attention layer is used to assign weights to the feature parameters, emphasizing features more important for electrical fire risk prediction. The feature parameters are weighted according to the weights assigned by the attention layer to generate new feature parameters. These new feature parameters are then processed by an improved LSTM network for temporal correlation analysis to capture the early evolution characteristics of electrical fire risk. In each gated unit of the LSTM, residual connection structures help retain more low-level feature information.

[0082] Output layer: Based on early evolutionary features, a softmax classifier outputs risk levels, including low risk, medium risk, and high risk. The calculation process of the softmax classifier is as follows: ; in, Corresponding to low risk, medium risk, and high risk; and These are the classifier parameters; These are the feature vectors corresponding to early evolutionary characteristics; and The parameters are the classifier parameters, representing the i-th column of the weight matrix and the i-th element of the bias vector, respectively. They together determine the contribution of the input features to the i-th class.

[0083] Based on the above calculations, the probability distribution of the three risk levels of early evolutionary characteristics is output.

[0084] Example 5: Refer to Figure 3 This embodiment discloses an electrical fire risk early warning system based on the Internet of Things, the system comprising: The data acquisition module is responsible for collecting multi-dimensional real-time status data from electrical equipment and constructing a training set for model training. During the construction of the training set, this module combines equipment maintenance records and historical fault events to label the risk level of data samples. It addresses the class imbalance problem by oversampling minority class samples (such as high-risk events) or synthesizing anomalous data using a generative adversarial network (GAN).

[0085] The data processing module is used to preprocess real-time status data using edge computing nodes. It completes data cleaning, feature extraction, and structure transformation on edge computing nodes (such as industrial gateways) to generate structured datasets.

[0086] The risk assessment module is responsible for building risk assessment models based on deep learning algorithms. It trains these models using a training set, including: using the training set as input and training the initial risk assessment model with deep learning algorithms. The purpose of training is to enable the risk assessment model to learn the risk characteristics and patterns in the data, ultimately outputting the trained risk assessment model.

[0087] The risk assessment model is trained by inputting samples from the structured dataset to obtain the risk level, which is then fed back to the user terminal. When the risk level does not meet expectations, an early warning signal is issued to the operation and maintenance personnel.

[0088] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An Internet of Things-based electrical fire risk early warning method, characterized in that, The method comprises the following steps: Collecting real-time state data of the electrical equipment, constructing a training set, the training set comprising a structured data set corresponding to historical state data of the electrical equipment and a risk level, the state data comprising operating parameters, line state parameters and environmental data; Using an edge computing node to pre-process the real-time state data, and generating a structured data set corresponding to the real-time state data after eliminating abnormal values; Constructing a risk assessment model based on a deep learning algorithm, training the risk assessment model using the training set, obtaining a trained risk assessment model, inputting a sample in the structured data set corresponding to the real-time state data into the trained risk assessment model, obtaining a risk level, feeding back the risk level to a user terminal, and sending a warning signal to an operation and maintenance personnel when the risk level does not meet expectations.

2. The IoT-based electrical fire risk pre-warning method according to claim 1, characterized in that, The method further comprises: Obtaining historical warning data and actual fire data of the electrical equipment, the historical warning data comprising a warning time, a target device ID and a risk level, and dynamically adjusting weight parameters of the risk assessment model through a reinforcement learning algorithm according to the historical warning data and the actual fire data. 3.The IoT-based electrical fire risk pre-warning method according to claim 2, characterized in that, The reinforcement learning algorithm uses a deep Q network, and the method comprises: A state space of the reinforcement learning comprises the historical warning data and the actual fire data, an action space is a weight adjustment amplitude of the risk assessment model, and a reward function is a deviation value between a model predicted risk level and the actual fire data; An experience replay pool of the deep Q network is initialized, and samples corresponding to the historical warning data and the actual fire data are stored in the experience replay pool; A batch of samples are randomly extracted from the experience replay pool in each iteration, parameters of the deep Q network are updated through a gradient descent method to maximize the reward function value, and a single weight parameter adjustment is completed.

4. The IoT-based electrical fire risk pre-warning method according to claim 1, wherein, The method further comprises: when the risk level does not meet expectations, reverse analyzing influence factors in the structured data through a causal inference algorithm, and feeding back the influence factors to the user terminal. 5.The IoT-based electrical fire risk pre-warning method according to claim 4, characterized in that, The causal inference algorithm uses a Bayesian network, and the method comprises: Each state parameter in the structured data is taken as a node of the Bayesian network, a conditional probability table between nodes is constructed based on an association between historical state data and a corresponding risk level; When the risk level output by the trained risk assessment model does not meet expectations, the risk level is taken as a result node, and a posterior probability of each state parameter node is calculated through reverse reasoning of the Bayesian network; State parameters with a posterior probability exceeding a preset threshold are determined as influence factors, and are fed back to the user terminal in descending order of the posterior probability. 6.The IoT-based electrical fire risk pre-warning method according to claim 5, wherein, The method further comprises: Periodically updating the conditional probability table of the Bayesian network based on newly added historical state data and risk level data; During the updating process, if there is a sample conflicting with the current conditional probability table in the newly added data, a weighted average method is used to fuse probability distributions of the conflicting sample and historical samples, the weight being positively correlated with a time decay coefficient corresponding to the sample, and the weight of a sample being higher if the sample is closer in time. 7.The IoT-based electrical fire risk pre-warning method according to claim 1, wherein, ​ The input layer receives the structured data set pre-processed by the edge computing node, filters and extracts feature parameters related to the electrical fire risk through a feature engineering module, the feature parameters including a current fluctuation coefficient in real-time data, an insulation resistance change rate in line state parameters, and a temperature and humidity coupling value in environmental data; The hidden layer uses an improved long short-term memory network to analyze the time sequence correlation of the feature parameters and capture early evolution characteristics of the electrical fire risk; The output layer outputs a risk level based on the early evolution characteristics through a softmax classifier, the risk level including low risk, medium risk, and high risk. 8.The IoT-based electrical fire risk early warning method according to claim 7, characterized in that, The improved long short-term memory network includes an input gate, a forget gate, and an output gate, and a residual connection structure is introduced in the input gate, the forget gate, and the output gate. 9.The IoT-based electrical fire risk early warning method according to claim 7, wherein, The hidden layer further includes an attention layer for assigning weights to the feature parameters and taking the feature parameters after the weights are assigned as new feature parameters.

10. An Internet of Things based electrical fire risk early warning system, the system being adapted for use in the method of any one of claims 1-9, characterized in that, The data acquisition module is configured to acquire real-time state data of the electrical equipment and construct a training set. The data processing module is configured to pre-process the real-time state data using the edge computing node, generate a structured data set after removing outliers, and construct a risk assessment model based on a deep learning algorithm. The risk assessment module is configured to train the risk assessment model using the training set, input samples in the structured data set into the trained risk assessment model, obtain a risk level, feed back the risk level to a user terminal, and send a warning signal to an operation and maintenance personnel when the risk level does not meet expectations. ​

Images