Intelligent Early Warning Method and System for Oil Depot Safety Risks Based on Data Fusion

Abstract

The present invention provides an intelligent early warning method and system for oil depot safety risks based on data fusion. First, the physical parameter sequences of equipment sensors in the oil depot area, the dynamic environmental index sets of environmental monitoring devices, and the abnormal event records in the historical accident database are collected. Then, they are input into a pre-trained feature fusion model, and joint feature vectors are extracted through cross-modal correlation analysis. Next, based on this, the risk diffusion weights between the target monitoring node and the adjacent areas are calculated, and the real-time risk propagation intensity value is generated by sliding and weighting with a dynamic time window. Then, the pre-configured risk assessment model is called to non-linearly map the real-time risk propagation intensity value and the safety threshold, and the comprehensive risk probability of the target monitoring node is output. Finally, according to the interval where the comprehensive risk probability is located, corresponding safety control instructions are triggered, including equipment start / stop, environmental regulation, and emergency response strategies, to realize the intelligent early warning of oil depot safety risks.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to an intelligent early warning method and system for oil depot safety risks based on data fusion. Background Art

[0002] In the field of oil depot safety management, existing safety risk early warning methods mainly rely on single data sources or simple data fusion technologies, facing many challenges such as low early warning accuracy, slow response speed, and difficulty in comprehensively capturing the dynamic of risk propagation. Currently, most of the commonly used methods in the industry only conduct isolated analysis on the physical parameters of equipment sensors or the indicators of environmental monitoring devices, ignoring the internal connections and mutual influences between different data sources, resulting in an incomplete and in-depth assessment of oil depot safety risks. At the same time, when dealing with historical accident data, existing technologies often only use it as a reference case and fail to fully explore the risk triggering conditions and propagation laws contained therein, making the early warning results lack foresight and guidance. Summary of the Invention

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide an intelligent early warning method for oil depot safety risks based on data fusion. The method includes:

[0004] Collect the real-time monitored physical parameter sequences of equipment sensors in the oil depot area, the dynamic environmental index set generated by environmental monitoring devices, and the abnormal event records in the historical accident database;

[0005] Input the physical parameter sequences, dynamic environmental index set, and abnormal event records into a pre-trained feature fusion model, and extract the joint feature vectors of equipment state degradation features, environmental abnormal propagation paths, and historical accident triggering conditions through cross-modal correlation analysis;

[0006] Calculate the risk diffusion weight between the target monitoring node and adjacent monitoring areas according to the joint feature vectors, and perform sliding weighting on the risk diffusion weight based on a dynamic time window to generate a real-time risk propagation intensity value;

[0007] Call a pre-configured risk assessment model to perform non-linear mapping on the real-time risk propagation intensity value and the current oil depot safety threshold, and output the comprehensive risk probability of the target monitoring node;

[0008] Trigger corresponding safety control instructions according to the predefined risk interval where the comprehensive risk probability is located. The safety control instructions include equipment start-stop control signals, environmental adjustment instructions, and emergency response strategies.

[0009] In another aspect, an embodiment of the present invention further provides an intelligent early warning system for oil depot safety risks based on data fusion, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present invention integrates the physical parameter sequence monitored by the device sensor in real time, the set of dynamic environment indicators generated by the environmental monitoring device, and the abnormal event records in the historical accident database. Through cross-modal correlation analysis, the joint feature vector of the device state degradation characteristics, the environmental anomaly propagation path, and the historical accident trigger conditions is accurately extracted. Further, by calculating the risk diffusion weight between the target monitoring node and the adjacent monitoring area and performing sliding weighting based on the dynamic time window, the dynamic changes of risk propagation can be captured in real time, and a real-time risk propagation intensity value with high timeliness and accuracy can be generated. The pre-configured risk assessment model performs a non-linear mapping of the real-time risk propagation intensity value and the current oil depot safety threshold, and outputs the comprehensive risk probability of the target monitoring node, which not only considers the absolute size of the risk, but also incorporates the dynamic change trend of the risk, making the risk assessment result closer to the actual safety condition. More importantly, according to the predefined risk interval where the comprehensive risk probability is located, the corresponding safety control instructions are triggered, realizing the closed-loop management from risk assessment to safety control, effectively improving the response speed and disposal efficiency of oil depot safety management. Thus, not only the accuracy and timeliness of oil depot safety risk early warning are improved, but also the ability of the oil depot to respond to sudden safety events is significantly enhanced through the generation and execution of intelligent safety control instructions. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic flowchart of the execution process of the intelligent early warning method for oil depot safety risks based on data fusion provided by an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of an exemplary hardware and software component of the intelligent early warning system for oil depot safety risks based on data fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of the intelligent early warning method for oil depot safety risks based on data fusion provided by an embodiment of the present invention. The intelligent early warning method for oil depot safety risks based on data fusion will be introduced in detail below.

[0014] Step S110, collect the sequence of physical parameters monitored in real time by the equipment sensors in the oil depot area, the set of dynamic environmental indicators generated by the environmental monitoring devices, and the abnormal event records in the historical accident database.

[0015] In this embodiment, in a certain oil depot, there are numerous devices and monitoring devices distributed. Specifically, the equipment sensors monitor various physical parameters in real time. For example, the pressure sensor installed on the storage oil tank records the pressure value inside the oil tank every 10 seconds, forming a pressure parameter sequence; the temperature sensor collects the oil temperature inside the oil tank every 15 seconds, generating a temperature parameter sequence; the flow sensor monitors the oil product flow in and out of the oil tank and records the flow data every 20 seconds, forming a flow parameter sequence. The above physical parameter sequences reflect the real-time operating status of the oil depot equipment.

[0016] For the environmental monitoring devices, gas concentration monitors are distributed in various corners of the oil depot, measuring the concentration of various harmful gases in the surrounding air every 30 seconds, thereby generating a dynamic set of gas concentration indicators; the temperature and humidity sensors record the temperature and humidity data of the environment every 45 seconds, forming a temperature and humidity indicator set; the anemometer measures the wind speed and direction every minute, constituting a wind speed and direction indicator set. The above indicator sets together form a set of dynamic environmental indicators, reflecting the environmental conditions around the oil depot.

[0017] In addition, the historical accident database stores the records of abnormal events that occurred in the past. For example, in a certain oil tank leakage accident, it details the time of the accident, the equipment numbers involved in the accident, the temperature and humidity, gas concentration, etc. of the environment at that time, as well as the triggering conditions such as operation mistakes and equipment failures that led to the accident.

[0018] Step S120, input the physical parameter sequence, the set of dynamic environmental indicators, and the abnormal event records into a pre-trained feature fusion model, and extract the joint feature vector of the equipment state degradation features, the environmental abnormal propagation path, and the historical accident triggering conditions through cross-modal correlation analysis.

[0019] In this embodiment, the collected physical parameter sequence can be input into the time-frequency domain decomposition module of the pre-trained feature fusion model. Taking the pressure fluctuation spectrum as an example, the pressure parameter sequence is decomposed in the time-frequency domain through a set algorithm, and the low-frequency energy ratio is calculated. Suppose the pressure parameter sequence shows complex fluctuations within a certain period of time. After time-frequency domain decomposition, it is found that the low-frequency energy ratio has increased from the original 30% to 40%, which may indicate a certain degree of degradation in the equipment state. For the temperature change curve, the extreme value of its second derivative is calculated. If the original temperature change is relatively stable and the extreme value of the second derivative is small, but now the extreme value of the second derivative suddenly increases, it indicates that the trend of temperature change has changed significantly, which may also be a manifestation of abnormal equipment state. The amplitude-frequency characteristics of the flow pulse signal are also extracted. For example, the amplitude-frequency characteristics of the original flow pulse signal are relatively stable within a certain frequency range, but now there are frequency offsets or abnormal increases in amplitude. The above features are all used as equipment state degradation features.

[0020] Process the dynamic environment index set, and generate a gas concentration gradient distribution map, a temperature and humidity field strength matrix, and a wind speed vector field through spatial interpolation calculation. Taking gas concentration as an example, there are multiple gas concentration monitoring points at different positions in the oil depot. Through the spatial interpolation algorithm, the above discrete monitoring data are used to generate a continuous gas concentration gradient distribution map. Suppose in a certain area, the gas concentration gradient change rate exceeds the preset threshold. The original concentration change rate per meter is within 0.5 ppm / m, and now it reaches 1 ppm / m in a certain area. This area is extracted as the environmental anomaly propagation path. The temperature and humidity field strength matrix and the wind speed vector field are also analyzed in a similar way to find the areas with abnormal changes.

[0021] Analyze the causal chain of the abnormal event record. For example, in a certain accident record, it is found that due to the seal failure of a certain valve (equipment failure mode), oil leakage occurred. At the same time, the environmental temperature suddenly increased (environmental mutation critical value), and the operator did not respond until 10 minutes after the accident (operation response delay time). The above factors are extracted as the triggering conditions of historical accidents.

[0022] Input the device status degradation characteristics, the coordinates of the environmental anomaly propagation path, and the historical accident trigger conditions into the multi-head attention module in the feature fusion model. In this multi-head attention module, calculate the spatial correlation weight between the device status degradation characteristics and the environmental anomaly propagation path. For example, if the device status degradation characteristics show abnormal pressure in a certain oil tank, and abnormal gas concentration gradient changes are detected in the area near the oil tank, then the spatial correlation weight between them will be relatively high. Suppose this spatial correlation weight is 0.8. At the same time, calculate the temporal dependence weight between the device status degradation characteristics and the historical accident trigger conditions. For example, there is a certain temporal correlation between equipment failure in historical accidents and the current device status degradation characteristics. If the current device status degradation characteristics are relatively close to the equipment failure mode in historical accidents, the temporal dependence weight may be 0.7. The causal association weight between the environmental anomaly propagation path and the historical accident trigger conditions will also be calculated. If some factors in the environmental factors in historical accidents are similar to those in the current environmental anomaly propagation path, the causal association weight may be 0.6.

[0023] Finally, perform feature concatenation and dimensionality reduction processing on the device status degradation characteristics, the environmental anomaly propagation path, and the historical accident trigger conditions according to the above weights. For example, concatenate multiple parameters of the device status degradation characteristics, the coordinate information of the environmental anomaly propagation path, and various factors of the historical accident trigger conditions according to the corresponding weights to form a high-dimensional feature vector. Then, through a dimensionality reduction algorithm, transform this high-dimensional vector into a joint feature vector containing spatio-temporal causal relationships.

[0024] Step S130, calculate the risk diffusion weight between the target monitoring node and the adjacent monitoring area according to the joint feature vector, and perform sliding weighting on the risk diffusion weight based on a dynamic time window to generate a real-time risk propagation intensity value.

[0025] In this embodiment, the time-frequency domain energy distribution matrix corresponding to the device status degradation characteristics and the gradient change rate tensor corresponding to the environmental anomaly propagation path can be extracted from the joint feature vector. Suppose the time-frequency domain energy distribution matrix corresponding to the device status degradation characteristics shows that the low-frequency energy ratio has increased from 35% to 45% in the past hour. According to this change, the cumulative degradation amount of the device status degradation characteristics within a preset time window (such as one hour) can be determined. In the gradient change rate tensor corresponding to the environmental anomaly propagation path, the gradient change rate in a certain area exceeds the preset threshold. Calculate the diffusion rate of the environmental anomaly propagation path in the adjacent monitoring area according to the coordinates of this area. Suppose the diffusion rate is 10 meters per hour.

[0026] Obtain the pipeline connection density parameter between the target monitoring node and each adjacent monitoring area, the mapping table of valve control logic relationships, and the turbulence coefficient in the gas diffusion physical model. Assume that the pipeline connection density parameter between the target monitoring node and adjacent monitoring area A is 0.6, indicating that the pipeline connections in this area are relatively dense; the mapping table of valve control logic relationships shows that the priority of the valves in this area is relatively high; the turbulence coefficient in the gas diffusion physical model is 0.8. Superimpose the pipeline connection density parameter and the turbulence coefficient according to a preset ratio (e.g., 0.4 and 0.6) to generate the initial diffusion weight base value, i.e., 0.6×0.4 + 0.8×0.6 = 0.72.

[0027] Input the initial diffusion weight base value into the dynamic weight allocation network. The first hidden layer of the dynamic weight allocation network receives the sequential concatenation vector of the cumulative degradation amount and the diffusion rate. Assume that the cumulative degradation amount is 0.1 (indicating the degree of degradation of the device state within a preset time window), and the diffusion rate is 10 (the diffusion distance per hour). Concatenate them in chronological order into a vector [0.1, 10]. The second hidden layer receives the priority tag sequence in the mapping table of valve control logic relationships. For example, [high, medium, low] forms a vector [0.8, 0.5, 0.2] after encoding.

[0028] In the dynamic weight allocation network, use a temporal convolutional kernel to extract local periodic patterns from the sequential concatenation vector to generate a time-sensitive factor. For example, through analysis by the temporal convolutional kernel, it is found that the cumulative degradation amount and the diffusion rate have shown a gradually increasing trend in the past few time windows, and the generated time-sensitive factor is 0.9. At the same time, use the graph attention mechanism to calculate the adjacency node correlation degree of the priority tag sequence to generate a spatial dependence factor. For example, according to the connection relationship and priority between the valves, the calculated spatial dependence factor is 0.7.

[0029] Multiply the time-sensitive factor and the spatial dependence factor element by element to obtain the dynamic diffusion correction coefficient, i.e., 0.9×0.7 = 0.63. And perform a weighted sum of the dynamic diffusion correction coefficient and the initial diffusion weight base value to generate the real-time risk propagation sub-intensity from each adjacent monitoring area to the target monitoring node. For example, 0.63×0.72 + (1 - 0.63)×0.72 = 0.72 (here assume the weights of the weighted sum are 0.63 and 0.37).

[0030] According to the sliding step of the dynamic time window, the real-time risk propagation sub-intensity within the current time window is superimposed with the decay residual intensity of the previous window, and the superimposed result is normalized to generate the real-time risk propagation intensity value of the target monitoring node. Assume that the dynamic time window length is one hour, the sliding step is half an hour, the decay residual intensity of the previous window is 0.5, and the real-time risk propagation sub-intensity of the current time window is 0.72. After superimposition, it is 0.5 + 0.72 = 1.22. After normalization, the real-time risk propagation intensity value is 1.22 / (1.22 + the superimposed value of other adjacent regions). Assume that the finally obtained value is 0.6.

[0031] Among them, the length of the dynamic time window is dynamically adjusted according to the proportion of low-frequency components in the time-frequency domain energy distribution matrix of the device state degradation characteristics. When the proportion of low-frequency components exceeds the preset critical value (such as 40%), the window length is shortened to enhance risk sensitivity. If the current proportion of low-frequency components reaches 45%, the dynamic time window length is shortened from one hour to half an hour to capture risk changes more timely.

[0032] Step S140, call the pre-configured risk assessment model to perform non-linear mapping on the real-time risk propagation intensity value and the current oil depot safety threshold, and output the comprehensive risk probability of the target monitoring node.

[0033] In this embodiment, the real-time risk propagation intensity value can be normalized in the time dimension to generate a standardized risk intensity sequence. Assume that the real-time risk propagation intensity values in the past few time windows are 0.4, 0.5, and 0.6 respectively. These values are mapped to the interval [0, 1] through a normalization algorithm. For example, using the formula (value - minimum value) / (maximum value - minimum value). Here, the minimum value is 0.4 and the maximum value is 0.6. After processing, a standardized risk intensity sequence [0, 0.5, 1] is generated.

[0034] The current oil depot safety threshold is decomposed into a dynamic threshold vector according to the preset risk level interval. Assume that the preset risk levels are divided into three levels: low, medium, and high, and the corresponding threshold upper and lower limits are: low risk [0, 0.3], medium risk [0.3, 0.7], high risk [0.7, 1], forming a dynamic threshold vector [0, 0.3, 0.7, 1].

[0035] The standardized risk intensity sequence and the dynamic threshold vector are input into the multi-layer perceptron of the risk assessment model. The multi-layer perceptron calculates the non-linear weighted sum between the standardized risk intensity sequence and the dynamic threshold vector through an activation function. For example, assume that the activation function is ReLU. After a series of weight calculations and activation function processing, the non-linear weighted sum is 0.5.

[0036] Perform probability conversion on the non-linear weighted sum to generate a preliminary risk probability distribution. Assume that the logistic regression function is used to convert the non-linear weighted sum into a probability value, and after calculation, the preliminary risk probability distribution is 0.6.

[0037] Adjust the preliminary risk probability distribution according to the historical scenario probability correction coefficient that matches the current joint feature vector in the historical accident database. For example, in a historical scenario similar to the current joint feature vector in the historical accident database, the probability correction coefficient for the occurrence of risk is 1.2. Then, the comprehensive risk probability of the target monitoring node after adjustment is 0.6 × 1.2 = 0.72.

[0038] Step S150, trigger the corresponding safety control instruction according to the predefined risk interval where the comprehensive risk probability is located. The safety control instruction includes equipment start-stop control signals, environmental adjustment instructions, and emergency response strategies.

[0039] In this embodiment, the comprehensive risk probability 0.72 can be matched with the threshold values of each level in the predefined risk interval. Since 0.72 is greater than 0.7, it is determined that the current risk level identifier of the target monitoring node is high risk.

[0040] Retrieve the corresponding equipment start-stop control signal set, environmental adjustment instruction sequence, and emergency response strategy priority queue from the preset instruction mapping table according to the current risk level identifier. For the high-risk level, the equipment start-stop control signal set may include immediately stopping the oil inlet and outlet operations of relevant oil tanks and closing relevant valves; the environmental adjustment instruction sequence may require increasing the power of the ventilation system and starting the gas extraction device; the emergency response strategy priority queue may prioritize evacuating the surrounding personnel and preparing fire-fighting equipment at the same time.

[0041] Based on the pipeline topological connection relationship of the target monitoring node, sort the execution order of the equipment start-stop control signal set to generate an equipment control instruction chain with time sequence constraints. For example, first close the valve closest to the target monitoring node, and then stop the operation of the oil pump, and form an equipment control instruction chain in this order.

[0042] According to the spatial distribution characteristics of the environmental anomaly propagation path, optimize the allocation of the start coordinates of the gas extraction device and the ventilation system power adjustment parameters in the environmental adjustment instruction sequence. Assume that the environmental anomaly propagation path shows that the gas mainly accumulates in a certain area. Then, start the gas extraction device near this area and reasonably adjust the power of the ventilation system according to the size of this area and the gas concentration.

[0043] Input the emergency response strategy priority queue into the dynamic decision-making engine, and combine the spatio-temporal change gradient of the real-time risk propagation intensity value to adjust the trigger timing and coverage range of the emergency response strategy. For example, if the real-time risk propagation intensity value shows that the risk is spreading in a certain direction, then the coverage range of the emergency response strategy should be expanded to that direction accordingly, and at the same time, evacuate the people in that direction in advance and adjust the trigger timing.

[0044] According to the execution feedback signal of the device control instruction chain and the response delay time of the environmental adjustment instruction sequence, dynamically update the instruction execution interval in the emergency response strategy priority queue. For example, if the feedback signal of a valve closing in the device control instruction chain is delayed, then the execution interval of the subsequent instructions should be appropriately extended to ensure the effective execution of the entire safety control instruction, and finally generate a set of safety control instructions and send them to the target execution terminal.

[0045] Based on the above steps, the embodiment of the present invention integrates the physical parameter sequence monitored in real time by the device sensor, the set of dynamic environmental indicators generated by the environmental monitoring device, and the abnormal event records in the historical accident database. Through cross-modal correlation analysis, the joint feature vector of the device state degradation characteristics, the environmental abnormal propagation path, and the historical accident trigger conditions is accurately extracted. Further, by calculating the risk diffusion weight between the target monitoring node and the adjacent monitoring area and performing sliding weighting based on the dynamic time window, the dynamic changes of risk propagation can be captured in real time, and a real-time risk propagation intensity value with high timeliness and accuracy can be generated. The pre-configured risk assessment model performs a non-linear mapping of the real-time risk propagation intensity value and the current oil depot safety threshold, and outputs the comprehensive risk probability of the target monitoring node, which not only considers the absolute size of the risk, but also incorporates the dynamic change trend of the risk, making the risk assessment result closer to the actual safety situation. More critically, according to the predefined risk interval where the comprehensive risk probability is located, the corresponding safety control instructions are triggered, realizing the closed-loop management from risk assessment to safety control, effectively improving the response speed and disposal efficiency of the oil depot safety management. Thus, not only the accuracy and timeliness of the oil depot safety risk warning are improved, but also the ability of the oil depot to respond to sudden safety events is significantly enhanced through the intelligent generation and execution of safety control instructions.

[0046] In a possible implementation manner, step S120 includes:

[0047] Step S121, perform time-frequency domain decomposition on the physical parameter sequence, and extract the low-frequency energy ratio in the pressure fluctuation spectrum, the extreme value of the second derivative of the temperature change curve, and the amplitude-frequency characteristics of the flow pulse signal as the device state degradation characteristics.

[0048] For example, for a sequence of physical parameters, a pressure sensor records the pressure value inside an oil tank every 10 seconds to form a pressure parameter sequence. Perform time-frequency domain decomposition on it to determine the proportion of low-frequency energy in the pressure fluctuation spectrum. For example, within a specific time period, 100 pressure values are collected. These pressure values are processed through a specialized time-frequency domain decomposition algorithm. First, arrange these 100 pressure values in chronological order, and then use methods such as Fourier transform to convert the pressure signal in the time domain to the frequency domain to obtain the pressure fluctuation spectrum. In the spectrum, divide the frequency range into different frequency bands, and calculate the proportion of the energy in the low-frequency band (set as 0-10 Hz) to the total energy. Through detailed calculation, add up the energy values corresponding to each frequency within the low-frequency band, and then divide by the total energy value of the entire spectrum to obtain the proportion of low-frequency energy. Assume that the total energy value is 1000 joules and the total energy of the low-frequency band is 300 joules, then the proportion of low-frequency energy is 300÷1000 = 30%.

[0049] A temperature sensor collects the oil temperature inside the oil tank every 15 seconds to form a temperature parameter sequence. Calculate the extreme values of the second derivative of the temperature change curve. First, draw the temperature change curve based on the collected temperature data. For example, within 2 hours, 480 temperature data points are collected, and these temperature data points are connected into a curve. Then, calculate the first derivative through numerical differentiation, that is, calculate the slope change between adjacent two data points. Next, perform the same calculation on the first derivative to obtain the second derivative. During the calculation process, precise calculations are performed for each data point. Find the maximum and minimum values in the second derivative, and these extreme values reflect the change situation of the temperature change rate. Assume that among the extreme values of the second derivative obtained through calculation, the maximum value is 0.5 °C / (minute)², and this value can be used as one of the extreme values of the second derivative of the temperature change curve.

[0050] A flow sensor records flow data every 20 seconds to form a flow pulse signal. Analyze its amplitude-frequency characteristics. First, perform Fourier transform on the flow pulse signal to convert it from the time domain to the frequency domain. In the frequency domain, analyze the amplitude sizes corresponding to different frequency components. For example, within 1 hour, 180 flow data are collected. After Fourier transform, the amplitude distribution at different frequencies is obtained. It is found that the amplitude within a certain frequency band (such as 5-10 Hz) has changed significantly compared with the past. Originally, the amplitude in this frequency band was stable at about 5 cubic meters per hour, and now it has risen to 8 cubic meters per hour. These amplitude change situations are the amplitude-frequency characteristics of the flow pulse signal, and these characteristics are used as equipment state degradation characteristics.

[0051] Step S122, perform spatial interpolation calculation on the dynamic environment index set to generate a gas concentration gradient distribution map, a temperature-humidity field strength matrix, and a wind speed vector field, and extract the regions where the gradient change rate exceeds the preset threshold as the environmental anomaly propagation paths.

[0052] For the set of dynamic environmental indicators, the gas concentration monitor measures the concentrations of various harmful gases in the surrounding air every 30 seconds, forming discrete monitoring point data. Kriging spatial interpolation is performed on this data to generate a gas concentration gradient distribution map. For example, 10 gas concentration monitoring points are set in an area of an oil depot, and the concentration values of a certain harmful gas are measured respectively. Using the Kriging interpolation method, considering factors such as the distance between monitoring points, the spatial position relationship, and the change trend of known monitoring point data, the concentration values are estimated for the areas between the monitoring points. Through a series of complex calculations, a reasonable gas concentration value is calculated for each interpolation point, thus generating a gas concentration gradient distribution map with continuous spatial coverage. In this gas concentration gradient distribution map, the concentration gradient change rate at different positions is calculated, and the preset threshold is set to 0.1 ppm / m. By calculating the ratio of the concentration change amount between adjacent positions to the distance, the areas where the gradient change rate exceeds the preset threshold are found. For example, if the calculated concentration gradient change rate in a certain area reaches 0.15 ppm / m, this area is extracted as part of the environmental anomaly propagation path.

[0053] The temperature and humidity sensor records the temperature and humidity data of the environment every 45 seconds, forming discrete data. Similarly, spatial interpolation calculation is used to generate a temperature and humidity field strength matrix. For example, there are 20 temperature and humidity monitoring points in a large area of an oil depot. Through the interpolation algorithm, the temperature and humidity values are estimated for each position in the whole area, forming a matrix representing the temperature and humidity distribution. In this matrix, the change situations of temperature and humidity at different positions are analyzed, and the gradient change rate is calculated. When the temperature gradient change rate in a certain area exceeds the preset threshold (assumed to be 0.2 °C / m), this area is marked as a possible environmental anomaly propagation path area.

[0054] The anemometer measures the wind speed and wind direction every minute, and generates a wind speed vector field through spatial interpolation. For example, 15 anemometer monitoring points are set around an oil depot. The wind speed and wind direction data of these points are interpolated to generate the wind speed vector field of the whole area. The change situations of wind speed at different positions in the vector field are analyzed, and the gradient change rate is calculated. The areas where the gradient change rate exceeds the preset threshold (assumed to be 0.5 m / (s·m)) are extracted as the environmental anomaly propagation path.

[0055] Step S123, perform a causal chain analysis on the abnormal event record, and extract the equipment failure mode, environmental mutation critical value, and operation response delay time that led to the accident trigger as the historical accident trigger conditions.

[0056] For abnormal event records, such as the record of a certain oil tank leakage accident. Through the analysis of the causal chain, it is found that the valve seal of a certain oil tank has aged, resulting in the failure of the valve seal (equipment failure mode). At that time, the ambient temperature suddenly increased from the normal 25°C to 35°C (ambient mutation critical value), and the operator received the alarm and started the response operation 15 minutes after the accident occurred (operation response delay time). These factors are extracted as the triggering conditions for historical accidents.

[0057] Step S124, input the equipment state degradation characteristics, environmental abnormal propagation path coordinates, and historical accident triggering conditions into the multi-head attention module in the feature fusion model, and calculate the spatial correlation weight between the equipment state degradation characteristics and the environmental abnormal propagation path, the time dependence weight between the equipment state degradation characteristics and the historical accident triggering conditions, and the causal association weight between the environmental abnormal propagation path and the historical accident triggering conditions.

[0058] In this embodiment, when calculating the spatial correlation weight between the equipment state degradation characteristics and the environmental abnormal propagation path, for example, the equipment state degradation characteristics show that the pressure of a certain oil tank is abnormal, and abnormal changes in the gas concentration gradient are detected in the surrounding area of the oil tank. Considering their spatial position relationship, the oil tank position coordinates are (10, 20), and the center coordinates of the gas concentration abnormal area are (12, 22). By calculating the distance between the two and considering surrounding environmental factors, etc., the spatial correlation weight between them is obtained using existing algorithms. Suppose after a series of complex calculations, the weight value is 0.7, indicating a strong correlation between the two in space.

[0059] When calculating the time dependence weight between the equipment state degradation characteristics and the historical accident triggering conditions, for example, the temperature change situation in the current equipment state degradation characteristics is relatively close to the temperature mutation time in the historical accident triggering conditions. The historical accident occurred at 10 am, and the time when the temperature began to change abnormally in the current equipment state degradation characteristics is 10:15 am. By analyzing factors such as the time interval and the temperature change trend, using a special time analysis algorithm, the time dependence weight is calculated to be 0.6, indicating a certain dependence relationship between the two in time.

[0060] When calculating the causal association weight between the environmental abnormal propagation path and the historical accident triggering conditions, if the environmental factors in the historical accident are similar to some factors in the current environmental abnormal propagation path. For example, in the historical accident, the abnormal increase in gas concentration led to the accident, and a similar trend of gas concentration increase also appears in the current environmental abnormal propagation path. By comparing factors such as the gas type and the concentration change range of the two, after detailed causal analysis and calculation, the causal association weight is obtained as 0.5.

[0061] Step S125: Perform feature splicing and dimensionality reduction processing on the device state degradation features, environmental anomaly propagation paths, and historical accident triggering conditions according to the spatial correlation weight, time dependence weight, and causal association weight, and generate a joint feature vector containing spatio-temporal causal relationships.

[0062] In this embodiment, each parameter of the device state degradation features (such as the proportion of low-frequency energy of pressure, the extreme value of the second derivative of temperature, the amplitude-frequency characteristics of flow, etc.), the coordinate information of the environmental anomaly propagation path, and each factor of the historical accident triggering conditions (device failure mode, environmental mutation critical value, operation response delay time, etc.) can be spliced according to the corresponding weights. For example, the parameters of the device state degradation features are spliced with a weight of 0.4, the coordinates of the environmental anomaly propagation path are spliced with a weight of 0.3, and the historical accident triggering conditions are spliced with a weight of 0.3 to form a high-dimensional feature vector. Then, a dimensionality reduction algorithm such as principal component analysis is used to process this high-dimensional vector, remove some dimensions with low correlation, and retain the main information dimensions, finally generating a joint feature vector containing spatio-temporal causal relationships.

[0063] In a possible implementation manner, the pre-trained feature fusion model is trained through the following steps:

[0064] Step S210: Extract the physical parameter sample sequence, environmental index sample set, and historical accident sample set with labeled accident triggering conditions under normal working conditions from the sample database as training input data.

[0065] In a possible implementation manner, before step S210, it further includes:

[0066] Step S201: Perform moving average filtering on the physical parameter sample sequence, and generate a stable physical parameter sequence after removing impulse interference points.

[0067] Step S202: Perform Kriging spatial interpolation on the discrete monitoring point data in the environmental index sample set to generate an environmental index surface covering the continuous space.

[0068] Step S203: Convert the text records in the historical accident sample set into binary event trigger vectors with time stamps according to a preset event coding table.

[0069] Step S204: Align the stable physical parameter sequence with the environmental index surface in a grid manner according to the second-level time stamp to generate a spatio-temporally matched physical-environment joint input matrix.

[0070] Step S205: Perform time series slicing splicing on the binary event trigger vector and the physical-environment joint input matrix for a set time period before and after the event to generate a labeled cross-modal training sample set.

[0071] In this embodiment, for the physical parameter sample sequence, taking the data collected by the oil tank pressure sensor as an example, a set of pressure sample sequences are collected within a certain period of time. The data shows certain fluctuations, and there may be pulse interference points. The moving average filtering method is adopted, and the size of the moving window is set to 5 data points. Starting from the first data point, the first 5 data points are added together, that is, the value of the first point plus the value of the second point plus the value of the third point plus the value of the fourth point plus the value of the fifth point. Assuming these 5 values are 10, 12, 15, 8, and 11 respectively, the sum is 10 + 12 + 15 + 8 + 11 = 56. Then the sum is divided by the window size 5 to obtain the average value 56÷5 = 11.2, which is the value after the first moving average filtering. Next, the window moves backward by one data point to calculate the second moving average filtering value, that is, the average value from the second point to the sixth point. And so on, the entire physical parameter sample sequence is processed to generate a stable physical parameter sequence after removing the pulse interference points.

[0072] For the environmental index sample set, taking gas concentration monitoring as an example, there are multiple discrete gas concentration monitoring points distributed at different positions in the oil depot. The data collected at these points constitutes discrete monitoring point data. The Kriging spatial interpolation method is used to generate an environmental index surface covering the continuous space. Suppose there are 9 monitoring points in a rectangular area, and the gas concentration values are recorded respectively. First, according to the spatial coordinates and gas concentration values of these monitoring points, the distances and variograms between each point are calculated. The variogram reflects the relationship between the spatial distance and the difference in variable values. By analyzing the data of the known monitoring points, the model parameters of the variogram are determined. Then, for other positions in the area that need to be interpolated, using the Kriging interpolation formula, considering factors such as the distance weights of the surrounding monitoring points, the estimated gas concentration value at that position is calculated. For example, for an interpolation point to be determined, by calculating its distances from the 3 nearest surrounding monitoring points and the gas concentration values of these 3 monitoring points, combined with the weights determined by the variogram, the estimated gas concentration value at this point is calculated. After calculating the estimated values of all interpolation points to be determined, a gas concentration surface covering the continuous space is generated, that is, an environmental index surface covering the continuous space.

[0073] For the historical accident sample set, which contains text records. The text records are converted into timestamped binary event trigger vectors according to a preset event coding table. For example, in the text record of an oil tank leakage accident, information such as the time of the accident, the equipment involved, and environmental factors is recorded. According to the preset event coding table, the equipment information, environmental factors, etc. are encoded respectively. For example, equipment A is encoded as 001, and too high environmental temperature is encoded as 010. These encodings are arranged in a certain order and combined with the timestamp of the accident to generate a timestamped binary event trigger vector. Suppose the accident occurs at 10:30:00, and the encoded binary event trigger vector is [103000, 001, 010, …].

[0074] Each data point in the stationary physical parameter sequence has a timestamp accurate to the second, and the environmental index surface is also grid-divided according to the second-level time. The parameter value corresponding to each timestamp in the physical parameter sequence is combined with the value at the corresponding grid position in the environmental index surface at the same timestamp to form an initial feature matrix. Suppose at a certain second, the pressure value in the physical parameter sequence is 12, and the gas concentration value at the corresponding position in the environmental index surface is 5 ppm. They are combined at the corresponding position in the matrix, and finally a spatio-temporal matching physical-environment joint input matrix is generated.

[0075] Then, the binary event trigger vector and the physical-environment joint input matrix are spliced by time series slices in the set time period before and after the event to generate a labeled cross-modal training sample set. The set time period is set to 5 minutes before and after the event. For each binary event trigger vector corresponding to an accident record, it is spliced with the data in the physical-environment joint input matrix within 5 minutes before and after the accident time. For example, if the accident occurs at 10:30:00, then the data in the physical-environment joint input matrix within 10 minutes from 10:25:00 to 10:35:00 is spliced with the binary event trigger vector of this accident to form a labeled cross-modal training sample. Among them, the binary event trigger vector is used as a label to indicate whether the sample corresponds to an accident condition or a normal condition, and so on, to generate the entire labeled cross-modal training sample set.

[0076] Step S220, input the physical parameter sample sequence into the time-frequency decomposition module of the feature fusion model to generate an initial feature matrix of the equipment state degradation feature, input the environmental index sample set into the spatial interpolation module to generate an initial gradient tensor of the environmental anomaly propagation path, and input the historical accident sample set into the causal analysis module to extract the initial causal vector of the historical accident trigger condition.

[0077] Taking the flow parameter sample sequence as an example, this flow parameter sample sequence records the flow changes of oil products over a period of time. The time-frequency decomposition module will convert the flow signal in the time domain to the frequency domain. First, arrange the flow sample sequence in chronological order, and then use methods such as Fourier transform for time-frequency decomposition. Divide the entire time series into multiple small segments, perform Fourier transform on each small segment, and calculate the amplitudes and phases of different frequency components. For example, in a 10-minute flow sample sequence, it is divided into 60 small segments, each 10 seconds long. After performing Fourier transform on each small segment, the amplitude distributions at different frequencies are obtained, and these amplitude distributions constitute the initial feature matrix of the equipment state degradation characteristics.

[0078] Continuing with the example of the temperature and humidity environment index sample set, the spatial interpolation module will generate the initial gradient tensor of the environmental anomaly propagation path based on the discrete temperature and humidity monitoring point data. First, calculate the distances and temperature and humidity change rates between different monitoring points. For example, there are three adjacent monitoring points. The temperature at point A is 25°C, the temperature at point B is 28°C, and the temperature at point C is 30°C. The distances between them are 5 meters and 3 meters respectively. By calculating the ratio of the temperature change amount to the distance, the temperature gradient from point A to point B is (28 - 25) ÷ 5 = 0.6°C / m, and the temperature gradient from point B to point C is (30 - 28) ÷ 3 ≈ 0.67°C / m. Organize the gradient information in different directions and positions into a tensor, that is, generate the initial gradient tensor of the environmental anomaly propagation path.

[0079] Furthermore, for the sample record of an oil tank explosion accident, the causal analysis module will analyze the cause of the accident. After detailed analysis, it is determined that it is due to the aging of the oil tank valve resulting in poor sealing (equipment failure mode), the oxygen concentration in the environment being too high (environmental mutation critical value), and the operator delaying 10 minutes to take measures after discovering the abnormality (operation response delay time). Arrange these factors in a certain order to form the initial causal vector of the historical accident trigger conditions, such as [valve aging and poor sealing, high oxygen concentration, 10 minutes].

[0080] Step S230, input the initial feature matrix, the initial gradient tensor, and the initial causal vector into the multi-head attention module to calculate the spatio-temporal correlation weight between the equipment state degradation characteristics and the environmental anomaly propagation path, and the causal mapping weight between the environmental anomaly propagation path and the historical accident trigger conditions.

[0081] For the degradation characteristics of equipment status and the propagation path of environmental anomalies, taking the initial feature matrix of pressure and the initial gradient tensor of gas concentration as examples, consider the relationship between pressure changes and gas concentration gradient changes in space and time. Suppose that within a certain time period, the area with increasing pressure and the area with increasing gas concentration gradient overlap spatially and are relatively close in time. By calculating factors such as the distance between their spatial positions and the degree of temporal synchronization, the spatio-temporal correlation weight is obtained using existing algorithms. For example, after complex calculations, considering that the central distance between the pressure change area and the gas concentration gradient change area is 10 meters and the temporal synchronization deviation is within 1 minute, the spatio-temporal correlation weight is calculated to be 0.7. For the propagation path of environmental anomalies and the triggering conditions of historical accidents, taking the gas concentration gradient and the critical value of environmental mutation in a certain accident as an example, analyze the causal relationship between the gas concentration gradient change and the environmental factors at the time of the accident. If the current trend of gas concentration gradient change is similar to the trend of environmental factors that led to the accident in the historical accident, by comparing factors such as the change amplitude and change time, the causal mapping weight is calculated. Suppose that after detailed comparison and calculation, the causal mapping weight is 0.6.

[0082] Step S240, weight and splice the initial feature matrix according to the spatio-temporal correlation weight to generate an enhanced representation vector of the equipment status degradation characteristics, dynamically screen the initial causal vector according to the causal mapping weight to generate a conditional screening vector of the historical accident triggering conditions, and perform channel fusion on the enhanced representation vector and the initial gradient tensor to generate a path fusion matrix of the environmental anomaly propagation path.

[0083] For example, for the initial feature matrix of equipment status degradation characteristics such as pressure and flow rate, according to the spatio-temporal correlation weight of 0.7, weighted calculations are performed on the elements in different feature matrices. Suppose there is an element value of 10 in the pressure feature matrix and the corresponding element value at the same position in the flow rate feature matrix is 8. According to the weight calculation, the value at the corresponding position in the new enhanced representation vector is 10×0.7 + 8×(1 - 0.7) = 7 + 2.4 = 9.4. And so on, the entire initial feature matrix is weighted and spliced to generate an enhanced representation vector of the equipment status degradation characteristics.

[0084] For example, there are multiple factors in the initial causal vector. According to the causal mapping weight of 0.6, each factor is evaluated. If the causal relationship between a certain factor and the environmental anomaly propagation path is weak and its evaluation value is lower than a certain threshold (suppose the threshold is 0.5), then this factor is screened out from the vector. After screening, a conditional screening vector of the historical accident triggering conditions is generated.

[0085] For example, each element in the enhanced representation vector of the device status degradation feature is fused and calculated with the element in the corresponding channel of the initial gradient tensor of the environmental anomaly propagation path. Suppose an element value in the enhanced representation vector is 9 and the element value in the corresponding channel of the initial gradient tensor is 5. Through a certain fusion algorithm (such as addition or weighted addition), assuming simple addition is used, the value at the corresponding position in the path fusion matrix is 9 + 5 = 14. Calculate all corresponding elements in this way to generate the path fusion matrix of the environmental anomaly propagation path.

[0086] Step S250: Concatenate the path fusion matrix and the conditional screening vector across modalities to generate a joint feature training vector. Compare the similarity between the joint feature training vector and a preset verification label vector, calculate the multi-dimensional feature loss value between the joint feature training vector and the verification label vector, and adjust the filter parameters of the time-frequency decomposition module, the interpolation coefficient of the spatial interpolation module, and the causal chain extraction threshold of the causal analysis module according to the multi-dimensional feature loss value.

[0087] In this embodiment, the path fusion matrix and the conditional screening vector can be concatenated in a certain order. For example, the path fusion matrix is placed in the front and the conditional screening vector is placed in the back to form a new vector, that is, the joint feature training vector.

[0088] On this basis, assume that the preset verification label vector represents normal working conditions, and the joint feature training vector and the verification label vector are compared in multiple dimensions. For example, in the first dimension, the value of the verification label vector is 1 and the value of the joint feature training vector is 0.8. Calculate the square of the difference as (1 - 0.8)² = 0.04. In the second dimension, the value of the verification label vector is 0.5 and the value of the joint feature training vector is 0.6. Calculate the square of the difference as (0.6 - 0.5)² = 0.01. And so on, calculate for all dimensions, and then add up the sum of the squares of these differences to obtain the multi-dimensional feature loss value. Suppose there are a total of 5 dimensions, and the sum of the squares of the differences after calculation in other dimensions is 0.05, then the total multi-dimensional feature loss value is 0.04 + 0.01 + 0.05 = 0.1.

[0089] If the multi-dimensional feature loss value is large, it means that the output of the current model has a large difference from the preset label. For the time-frequency decomposition module, it may be necessary to adjust parameters such as the cut-off frequency of the filter. Suppose the current cut-off frequency of the filter is 10 Hz. According to the multi-dimensional feature loss value and the adjustment strategy, adjust the cut-off frequency to 12 Hz. For the spatial interpolation module, adjust the interpolation coefficient. For example, the original interpolation coefficient is 0.6 and it becomes 0.7 after adjustment. For the causal analysis module, adjust the causal chain extraction threshold. Suppose the original threshold is 0.4 and it becomes 0.5 after adjustment, so that the model can extract features more accurately.

[0090] Step S260: Input the new joint feature vector output by the adjusted feature fusion model into the discriminator network to determine the class probabilities of the new joint feature vector belonging to normal working conditions or historical accidents.

[0091] In this embodiment, the discriminator network will, based on the features of the new joint feature vector, give the probabilities of this vector belonging to normal working conditions or historical accidents through a series of calculations and judgments. For example, after the calculation of the discriminator network, it is obtained that the probability of the new joint feature vector belonging to normal working conditions is 0.4, and the probability of belonging to historical accidents is 0.6.

[0092] Step S270: Optimize the number of attention heads of the multi-head attention module and the scaling factor of channel fusion according to the difference value between the class probability and the true working condition label until the discriminator network can no longer distinguish the joint feature distributions of normal working conditions and historical accidents.

[0093] In this embodiment, if there is a large difference between the class probability and the true working condition label, for example, the true working condition label is normal working conditions, but the discriminator network gives a relatively high probability of belonging to historical accidents. At this time, it is necessary to adjust the number of attention heads of the multi-head attention module and the scaling factor of channel fusion. Suppose the original number of attention heads is 4 and the channel fusion scaling factor is 0.5. According to the size of the difference value and the optimization strategy, the number of attention heads is increased to 5, and the channel fusion scaling factor is adjusted to 0.6. Continuously repeat the above process until the discriminator network can no longer distinguish the joint feature distributions of normal working conditions and historical accidents, that is, the model achieves a better training effect.

[0094] In a possible implementation manner, step S130 includes:

[0095] Step S131: Extract the time-frequency domain energy distribution matrix corresponding to the equipment state degradation feature and the gradient change rate tensor corresponding to the environmental anomaly propagation path from the joint feature vector.

[0096] Taking a key oil tank in an oil depot as an example of the target monitoring node, the combined feature vector contains information in multiple aspects. For the characteristics of equipment state degradation, through the previous feature extraction and fusion, the time-frequency domain energy distribution matrix is obtained. This time-frequency domain energy distribution matrix records the energy distribution of equipment operation parameters (such as pressure, temperature, flow rate, etc.) in different frequency bands. For example, after performing time-frequency domain analysis on the pressure parameter of the oil tank, in the formed time-frequency domain energy distribution matrix, the frequency range is divided into multiple frequency bands, arranged in sequence from low frequency to high frequency. The energy value corresponding to each frequency band is obtained through a series of complex calculations. For example, the pressure signal is Fourier-transformed over a period of time to convert the time-domain signal to the frequency domain, and then the energy contribution of each frequency component is calculated. Finally, the energy distribution matrix is formed by summarizing. For the abnormal environmental propagation path, through spatial interpolation and analysis of the dynamic environmental index set, the corresponding gradient change rate tensor is obtained. Taking gas concentration monitoring as an example, multiple gas concentration monitoring points are distributed in each area of the oil depot. By generating a gas concentration gradient distribution map through spatial interpolation, the gradient change rate at different positions is calculated, and this gradient change rate information constitutes the gradient change rate tensor.

[0097] Step S132: Determine the cumulative degradation amount of the equipment state degradation characteristics within a preset time window according to the low-frequency energy ratio in the time-frequency domain energy distribution matrix, and calculate the diffusion rate of the abnormal environmental propagation path of adjacent monitoring areas based on the regional coordinates exceeding the preset threshold in the gradient change rate tensor.

[0098] Assume that the preset time window is one hour. In this time-frequency domain energy distribution matrix, the low-frequency energy ratio reflects the stability of the equipment operation state. For example, after calculation, within this one hour, the proportion of the energy in the low-frequency band (set as 0 - 10 Hz) in the total energy has increased from the original 30% to 35%. To determine the cumulative degradation amount, compare the current low-frequency energy ratio with the low-frequency energy ratio during historical normal operation. Assume that the low-frequency energy ratio during historical normal operation is stably at 30%. Then, within this one hour, the calculation method of the cumulative degradation amount is as follows: First, calculate the difference in the ratio, that is, 35% - 30% = 5%. This difference represents the change amplitude of the low-frequency energy ratio within the preset time window, and is used to measure the degradation degree of the equipment state. This 5% can be used as the cumulative degradation amount.

[0099] Further, the preset threshold is set to 0.1 ppm / m, indicating that the gradient change rate of the gas concentration within every meter of distance exceeding this value is regarded as abnormal. In the gradient change rate tensor, by checking the gradient change rates of each region one by one, it is found that the gradient change rate of the gas concentration in a certain adjacent monitoring region exceeds the preset threshold. Assuming that the coordinate information of this region is clear, the diffusion rate is calculated by analyzing the gradient change situation of this region over a period of time (such as half an hour). First, determine the gas concentration gradient values at the starting moment and the ending moment of this region. Assume that the gradient value at the starting moment is 0.12 ppm / m and the gradient value at the ending moment is 0.15 ppm / m, and the length of the region is 10 meters. Then, within this half an hour, the concentration change amount is 0.15 - 0.12 = 0.03 ppm, the diffusion distance is 10 meters, and the diffusion time is half an hour (i.e., 0.5 hours). According to the calculation method of the diffusion rate, the diffusion rate is equal to the concentration change amount divided by the diffusion time and then divided by the diffusion distance, that is, 0.03÷0.5÷10 = 0.006 ppm / (m·h), which is the diffusion rate of the environmental anomaly propagation path in this adjacent monitoring region.

[0100] Step S133, obtain the pipeline connection density parameter, the valve control logic relationship mapping table, and the turbulence coefficient in the gas diffusion physical model between the target monitoring node and each adjacent monitoring region, and superimpose the pipeline connection density parameter and the turbulence coefficient according to a preset ratio to generate an initial diffusion weight base value.

[0101] Taking the target oil tank and three adjacent monitoring areas (Area A, Area B, and Area C) as an example, the pipeline connection density parameter is determined by counting the number and distribution of pipelines between the target monitoring node and each adjacent monitoring area. Suppose there are 5 pipelines between the target oil tank and Area A, and the pipeline connection density parameter of this area is evaluated as 0.6; there are 3 pipelines between it and Area B, and the pipeline connection density parameter is 0.4; there are 4 pipelines between it and Area C, and the pipeline connection density parameter is 0.5. The valve control logic relationship mapping table records the opening and closing states of each valve and their control relationships. Different valve states will affect the flow of oil and gas, thus having an impact on risk diffusion. For example, valve V1 controls the connection between the target oil tank and Area A. When valve V1 is fully open, the flow of oil and gas is smooth, and the possibility of risk diffusion increases. The turbulence coefficient in the gas diffusion physical model takes into account the influence of the turbulent characteristics of air flow in the oil depot on gas diffusion. Suppose the turbulence coefficient of Area A is 0.7, that of Area B is 0.6, and that of Area C is 0.75. The preset ratio is set to 0.4 for the pipeline connection density parameter and 0.6 for the turbulence coefficient. For Area A, the calculation of the initial diffusion weight base value is: 0.6×0.4 + 0.7×0.6 = 0.24 + 0.42 = 0.66; for Area B, the calculation is: 0.4×0.4 + 0.6×0.6 = 0.16 + 0.36 = 0.52; for Area C, the calculation is: 0.5×0.4 + 0.75×0.6 = 0.2 + 0.45 = 0.65. These calculation results are the initial diffusion weight base values for each adjacent monitoring area.

[0102] Step S134, input the initial diffusion weight base value into the dynamic weight allocation network, where the first hidden layer of the dynamic weight allocation network receives the sequential splicing vector of the cumulative degradation amount and the diffusion rate, and the second hidden layer receives the priority marking sequence in the valve control logic relationship mapping table.

[0103] For example, for Area A, the cumulative degradation amount is 5% and the diffusion rate is 0.006 ppm / (m·h). They are spliced into a vector [5%, 0.006] in chronological order and input into the first hidden layer. In the valve control logic relationship mapping table, valve V1 of Area A has a higher priority. After encoding, it forms a priority marking sequence [0.8] and is input into the second hidden layer.

[0104] Step S135, in the dynamic weight allocation network, use a temporal convolutional kernel to extract local periodic patterns from the sequential splicing vector to generate a time-sensitive factor, and at the same time use a graph attention mechanism to calculate the adjacency node correlation degree of the priority marking sequence to generate a spatial dependence factor.

[0105] For example, the temporal convolution kernel analyzes the input time-series concatenated vectors to find periodic change patterns therein. For example, by analyzing the changes in the cumulative degradation amount and diffusion rate over multiple past time windows, it is found that they show a gradually increasing trend at a certain time interval. Suppose that after calculation by the temporal convolution kernel, the period of this increasing trend is determined to be 3 hours, and it is currently in the rising stage. Based on this information, a time-sensitive factor of 0.9 is generated, indicating a high sensitivity to risk changes at the current time point. At the same time, the graph attention mechanism is used to calculate the adjacency node correlation degree of the priority marking sequence to generate a spatial dependence factor. The graph attention mechanism considers the connection relationships between valves and priority information, and analyzes the correlation degree of adjacent nodes (such as other valves and monitoring areas) related to the target monitoring node. For example, through in-depth analysis of the valve control logic relationships, it is found that there is a close association between valve V1 and several other important valves, and it is in a key position in the entire oil depot system. After calculation, a spatial dependence factor of 0.7 is generated.

[0106] Step S136: Multiply the time-sensitive factor and the spatial dependence factor element by element to obtain a dynamic diffusion correction coefficient, and perform weighted summation of the dynamic diffusion correction coefficient and the initial diffusion weight base value to generate the real-time risk propagation sub-intensity from each adjacent monitoring area to the target monitoring node.

[0107] For example, for area A, the dynamic diffusion correction coefficient is 0.9×0.7 = 0.63. Suppose that when performing weighted summation, the weight of the dynamic diffusion correction coefficient is 0.7 and the weight of the initial diffusion weight base value is 0.3. Then the real-time risk propagation sub-intensity from area A to the target monitoring node is: 0.63×0.7 + 0.66×0.3 = 0.441 + 0.198 = 0.639. Similarly, the same calculation is performed for area B and area C to obtain their respective real-time risk propagation sub-intensities from them to the target monitoring node.

[0108] Step S137: According to the sliding step size of the dynamic time window, superimpose the real-time risk propagation sub-intensity within the current time window and the decay residual intensity of the previous window, and perform normalization processing on the superimposed result to generate the real-time risk propagation intensity value of the target monitoring node.

[0109] Assume that the length of the dynamic time window is one hour and the sliding step is half an hour. The decay residual intensity of the previous window is calculated by decaying the real-time risk propagation intensity value of the previous window. Assume that the real-time risk propagation intensity value of the previous window is 0.5. After decay calculation (such as according to the set decay formula, assume the decay coefficient is 0.8), the decay residual intensity of the previous window is 0.5×0.8 = 0.4. The real-time risk propagation sub-intensities from regions A, B, and C to the target monitoring node within the current time window are 0.639, 0.55, and 0.6 respectively (assuming the values obtained for regions B and C through the above calculation). Adding them to the decay residual intensity of the previous window, the sum is 0.4 + 0.639 + 0.55 + 0.6 = 2.189. After normalization, that is, dividing the sum by the sum of all intensity values participating in the superposition, the real-time risk propagation intensity value of the target monitoring node is: (0.639÷2.189) = 0.292 (here taking the proportion of the real-time risk propagation sub-intensity of region A in the normalization as an example, and the final real-time risk propagation intensity value is the comprehensive result after normalizing all regions).

[0110] Wherein, the length of the dynamic time window is dynamically adjusted according to the proportion of low-frequency components in the time-frequency domain energy distribution matrix of the device state degradation characteristics. When the proportion of low-frequency components exceeds a preset critical value, the window length is shortened to enhance risk sensitivity.

[0111] For example, assume the preset critical value is 35%. When the proportion of low-frequency components in the time-frequency domain energy distribution matrix of the device state degradation characteristics reaches 36%, the length of the dynamic time window is shortened from one hour to half an hour. This can update the risk propagation intensity value more frequently, capture risk changes in a timely manner, enhance the sensitivity to risks, and thus take corresponding safety measures more effectively to ensure the safe operation of the oil depot.

[0112] In a possible implementation manner, step S140 includes:

[0113] Step S141, perform time-dimensional normalization on the real-time risk propagation intensity value to generate a standardized risk intensity sequence.

[0114] Taking the target monitoring node as an example, assume that in the past few hours, the real-time risk propagation intensity values are 0.3, 0.5, 0.7, 0.6, and 0.4 respectively. To perform normalization processing in the time dimension, it is necessary to first find the maximum and minimum values in this set of data. In this set of data, the maximum value is 0.7 and the minimum value is 0.3. For the first value 0.3, the standardization calculation process is as follows: subtract the minimum value 0.3 from this value, getting 0.3 - 0.3 = 0, and then divide by the difference between the maximum value and the minimum value, that is, 0.7 - 0.3 = 0.4, so 0 ÷ 0.4 = 0. For the second value 0.5, the calculation process is: (0.5 - 0.3) ÷ (0.7 - 0.3) = 0.2 ÷ 0.4 = 0.5. For the third value 0.7, the calculation is: (0.7 - 0.3) ÷ (0.7 - 0.3) = 0.4 ÷ 0.4 = 1. For the fourth value 0.6, the calculation is: (0.6 - 0.3) ÷ (0.7 - 0.3) = 0.3 ÷ 0.4 = 0.75. For the fifth value 0.4, the calculation is: (0.4 - 0.3) ÷ (0.7 - 0.3) = 0.1 ÷ 0.4 = 0.25. After such calculations, the generated standardized risk intensity sequence is 0, 0.5, 1, 0.75, 0.25.

[0115] Step S142: Decompose the current oil depot safety threshold into a dynamic threshold vector according to a preset risk level interval, where the dynamic threshold vector includes the upper and lower limits of the thresholds corresponding to each risk level.

[0116] The preset risk levels are divided into three levels: low, medium, and high. The threshold range for the low risk level is set to 0 to 0.3, the threshold range for the medium risk level is 0.3 to 0.7, and the threshold range for the high risk level is 0.7 to 1. Then the dynamic threshold vector is composed of the upper and lower limits of the thresholds for these three levels, that is, [0, 0.3, 0.7, 1].

[0117] Step S143: Input the standardized risk intensity sequence and the dynamic threshold vector into the multi-layer perceptron of the risk assessment model, and calculate the non-linear weighted sum between the standardized risk intensity sequence and the dynamic threshold vector through an activation function.

[0118] There are multiple neuron layers inside the multi-layer perceptron, and each neuron performs a weighted sum on the input data and processes it through an activation function. Assume that each value in the standardized risk intensity sequence has a specific weight relationship with the corresponding value in the dynamic threshold vector. For example, for the first value 0 in the standardized risk intensity sequence, it has weights 0.1, 0.2, 0.3, and 0.4 respectively with the four values [0, 0.3, 0.7, 1] in the dynamic threshold vector. Then the weighted calculation process is: 0×0.1 + 0×0.2 + 0×0.3 + 0×0.4 = 0. For the second value 0.5, the weights are assumed to be 0.2, 0.3, 0.3, and 0.2, and the weighted calculation is: 0.5×0.2 + 0.5×0.3 + 0.5×0.3 + 0.5×0.2 = 0.1 + 0.15 + 0.15 + 0.1 = 0.5. For the third value 1, the weights are assumed to be 0.3, 0.3, 0.2, and 0.2, and the weighted calculation is: 1×0.3 + 1×0.3 + 1×0.2 + 1×0.2 = 0.3 + 0.3 + 0.2 + 0.2 = 1. For the fourth value 0.75, the weights are assumed to be 0.2, 0.3, 0.3, and 0.2, and the weighted calculation is: 0.75×0.2 + 0.75×0.3 + 0.75×0.3 + 0.75×0.2 = 0.15 + 0.225 + 0.225 + 0.15 = 0.75. For the fifth value 0.25, the weights are assumed to be 0.1, 0.2, 0.3, and 0.4, and the weighted calculation is: 0.25×0.1 + 0.25×0.2 + 0.25×0.3 + 0.25×0.4 = 0.025 + 0.05 + 0.075 + 0.1 = 0.25. After processing these weighted sums through an activation function (assuming the activation function is ReLU, that is, values greater than 0 remain unchanged, and values less than 0 become 0), the resulting non-linear weighted sum results are 0, 0.5, 1, 0.75, 0.25 (here because these values are all greater than 0, so they remain unchanged).

[0119] Step S144, perform probability conversion on the non-linear weighted sum to generate a preliminary risk probability distribution.

[0120] Assume that a logistic regression function is used for probability conversion. The formula of the logistic regression function is to convert the input value into a probability value through existing algorithms. For the first value 0 in the non-linear weighted sum result, after calculation by the logistic regression function, assuming the conversion formula is 1 / (1 + e to the power of the negative input value), where e is the natural constant approximately equal to 2.718. Then the calculation process is 1 / (1 + e to the power of 0) = 1 / (1 + 1) = 0.5. For the second value 0.5, the calculation is 1 / (1 + e to the power of -0.5). e to the power of -0.5 is approximately 0.607, so 1 / (1 + 0.607) ≈ 0.62. For the third value 1, the calculation is 1 / (1 + e to the power of -1). e to the power of -1 is approximately 0.368, so 1 / (1 + 0.368) ≈ 0.73. For the fourth value 0.75, the calculation is 1 / (1 + e to the power of -0.75). e to the power of -0.75 is approximately 0.472, so 1 / (1 + 0.472) ≈ 0.68. For the fifth value 0.25, the calculation is 1 / (1 + e to the power of -0.25). e to the power of -0.25 is approximately 0.779, so 1 / (1 + 0.779) ≈ 0.56. The preliminary risk probability distribution generated in this way is 0.5, 0.62, 0.73, 0.68, 0.56.

[0121] Step S145, adjust the preliminary risk probability distribution according to the historical scenario probability correction coefficient that matches the current combined feature vector in the historical accident database, and output the comprehensive risk probability of the target monitoring node.

[0122] Assume that the historical scenario probability correction coefficient that matches the current combined feature vector in the historical accident database is 1.2. Multiply each value in the preliminary risk probability distribution by this historical scenario probability correction coefficient. The first value 0.5 is adjusted to 0.5 × 1.2 = 0.6. The second value 0.62 is adjusted to 0.62 × 1.2 = 0.744. The third value 0.73 is adjusted to 0.73 × 1.2 = 0.876. The fourth value 0.68 is adjusted to 0.68 × 1.2 = 0.816. The fifth value 0.56 is adjusted to 0.56 × 1.2 = 0.672. Combining these adjusted values, obtain the comprehensive risk probability of the target monitoring node through a certain comprehensive calculation method (such as taking the average). Assume the average value is calculated as (0.6 + 0.744 + 0.876 + 0.816 + 0.672) ÷ 5 = 3.708 ÷ 5 = 0.7416.

[0123] In a possible implementation manner, step S150 includes:

[0124] Step S151, perform interval matching between the comprehensive risk probability and each level threshold in the predefined risk interval to determine the current risk level identifier of the target monitoring node.

[0125] For example, the comprehensive risk probability 0.7416 is matched with each level threshold in the predefined risk interval to determine the current risk level identifier of the target monitoring node. Since 0.7416 is greater than 0.7 and is in the high-risk level interval, the current risk level identifier of the target monitoring node is determined to be high risk.

[0126] Step S152, retrieve the corresponding device start-stop control signal set, environment adjustment instruction sequence, and emergency response strategy priority queue from the preset instruction mapping table according to the current risk level identifier.

[0127] For example, for the high-risk level, the device start-stop control signal set may include immediately stopping the operation of the feed pump of the target oil tank and closing all feed valves connected to the oil tank. For example, there is a feed pump numbered P1, and the corresponding control signal is "stop the P1 feed pump"; there are valves V1, V2, and V3 connected to the oil tank, and the control signals are "close the V1 valve", "close the V2 valve", and "close the V3 valve" respectively. The environment adjustment instruction sequence may require increasing the power of the ventilation system and starting multiple gas extraction devices. For example, the current power of the ventilation system is 50%, and the adjustment instruction is "increase the power of the ventilation system to 100%"; there are gas extraction devices A, B, and C, and the instructions are "start gas extraction device A", "start gas extraction device B", and "start gas extraction device C". The emergency response strategy priority queue may give top priority to evacuating the surrounding personnel and preparing fire-fighting equipment at the same time. For example, the emergency response strategies include "immediately evacuate all personnel within a radius of 500 meters" and "start fire-fighting equipment and deploy fire trucks to the designated location", and the priority of evacuating personnel is higher than starting fire-fighting equipment.

[0128] Step S153, based on the pipeline topological connection relationship of the target monitoring node, sort the execution order of the device start-stop control signal set to generate a device control instruction chain with time sequence constraints.

[0129] For example, the target monitoring node is connected to multiple devices through pipelines. The pipeline topological connection relationship shows that the valve closest to the oil tank should be closed first, and then the feed pump should be stopped. So the execution order of the device control instruction chain is: first, "close the V1 valve", and the V1 valve is the feed valve closest to the oil tank; then, "close the V2 valve"; then, "close the V3 valve"; finally, "stop the P1 feed pump". By following the execution order determined by the pipeline topological connection relationship, it can ensure the safe and orderly shutdown of relevant equipment in high-risk situations and prevent further leakage of oil products or other hazards.

[0130] Step S154, according to the spatial distribution characteristics of the environmental anomaly propagation path, optimize the allocation of the start coordinates of the gas extraction devices and the ventilation system power adjustment parameters in the environment adjustment instruction sequence.

[0131] For example, through the previous analysis of the environmental anomaly propagation path, it is found that the gas mainly accumulates in a specific area around the oil tank. Then, for the gas exhaust device, start the gas exhaust device A at the central position coordinates of this accumulation area (assumed to be (x1, y1)), start the gas exhaust device B at the position coordinates near the personnel activity area on the edge of this area (assumed to be (x2, y2)), and start the gas exhaust device C at the position coordinates adjacent to other storage areas in this area (assumed to be (x3, y3)) to ensure the comprehensive and effective exhaust of the accumulated gas. For the ventilation system power adjustment parameter, according to the size of this area and the gas concentration situation, since this area is large and the gas concentration is high, increase the ventilation system power from the current 50% to 100%, and adjust the ventilation direction so that it ventilates towards the gas accumulation area to accelerate the diffusion and dilution of the gas.

[0132] Step S155: Input the emergency response strategy priority queue into the dynamic decision engine, and in combination with the spatio-temporal change gradient of the real-time risk propagation intensity value, adjust the triggering timing and coverage range of the emergency response strategy.

[0133] For example, the real-time risk propagation intensity value shows that the risk is spreading in a certain direction. For example, the risk is spreading towards the office area of the oil depot. Based on this information, the dynamic decision engine adjusts the triggering timing of the emergency response strategy, evacuates the personnel in the office area in advance, and focuses on adjusting the evacuation range in the strategy of "immediately evacuate all personnel within a radius of 500 meters" to the office area and the surrounding 500-meter range. At the same time, expand the coverage range of the emergency response strategy, and expand the deployment range of fire-fighting equipment to the roads near the office area to ensure that fire extinguishing and rescue work can be carried out in a timely manner in case of an emergency.

[0134] Step S156: Dynamically update the instruction execution interval in the emergency response strategy priority queue according to the execution feedback signal of the equipment control instruction chain and the response delay time of the environmental adjustment instruction sequence, generate the final safety control instruction set and send it to the target execution terminal.

[0135] Suppose the execution feedback signal of "closing valve V1" in the device control instruction chain shows that the valve closing time is 30 seconds later than expected, which may affect the execution effect of subsequent instructions. At the same time, the response delay time for the ventilation system power increase in the environmental adjustment instruction sequence is 60 seconds. Based on this feedback information, dynamically update the instruction execution intervals in the emergency response strategy priority queue. The original interval between the instruction to evacuate personnel and the instruction to activate fire-fighting equipment is 5 minutes. Considering the delays in the device and environmental adjustment, the interval is extended to 6 minutes to ensure that all measures can be executed orderly and effectively. After such adjustment, generate the final set of safety control instructions and send them to the target execution terminal to ensure that the oil depot can take corresponding measures quickly and accurately in the face of high-risk situations, guaranteeing the safety of personnel and the safety of oil depot facilities.

[0136] In a possible implementation manner, after step S156, it further includes:

[0137] Step S160, capture the valve opening feedback signal and the actual ventilation power value returned by the execution terminal in real time, and generate a device response status code.

[0138] Taking the important valve V1 for controlling the oil product inlet and outlet in the oil depot and the ventilation system as examples. After receiving the control instruction of "closing valve V1", the execution terminal will feedback the opening situation of the valve in real time. Suppose the value range of the valve opening is 0% to 100%. After a period of time after the instruction is issued, the valve opening feedback signal returned by the execution terminal shows 5%, which means the valve is not fully closed. For the ventilation system, the instruction requires the ventilation power to be increased to 100%, and the actual ventilation power value returned by the execution terminal is 80%. Based on this feedback information, a device response status code is generated. The device response status code is a coding method that comprehensively represents the operating state of the device. For valve V1, since it is not fully closed, a certain bit of the status code (assuming the first bit represents the closing state of valve V1, 0 means fully closed, 1 means not fully closed) will be set to 1; for the ventilation system, since the ventilation power does not reach the target value, another bit of the status code (assuming the second bit represents the ventilation system power state, 0 means reaching the target power, 1 means not reaching) will be set to 1. The finally generated device response status code may be a multi-bit binary number, such as "11", indicating that valve V1 is not fully closed and the ventilation system power does not reach the target value.

[0139] Step S170, perform dynamic time warping matching on the device response status code and the pre-stored ideal response curve, and calculate the execution delay time difference and the action completion degree deviation value.

[0140] The pre-stored ideal response curve is obtained based on a large amount of historical data and theoretical analysis, representing the response process of the device to control instructions under ideal conditions. For valve V1, the ideal response curve shows that it should be fully closed within 60 seconds after the control instruction is issued, that is, the opening degree linearly decreases from 100% to 0%. In actuality, when starting to time from the issuance of the instruction, 90 seconds have passed when the feedback signal with an opening degree of 5% is received. Calculating the execution delay time difference is the difference between the actual response time and the ideal response time, that is, 90 seconds - 60 seconds = 30 seconds. For the deviation value of the action completion degree, ideally the valve should be fully closed, that is, the action completion degree is 100%, while the actual action completion degree is (100% - 5%) = 95%, so the deviation value of the action completion degree is 100% - 95% = 5%. For the ventilation system, the ideal response curve requires the power to be smoothly increased from the current value to 100% within 90 seconds after the instruction is issued. In actuality, the power only reaches 80% at 90 seconds. The execution delay time difference is 0 seconds (because the comparison is made at the ideal 90-second time point), and the deviation value of the action completion degree is 100% - 80% = 20%.

[0141] Step S180, when the execution delay time difference exceeds the preset threshold, automatically extend the time trigger intervals of each signal in the subsequent device control instruction chain.

[0142] Suppose for valve operation, the preset threshold is 20 seconds, and the calculated execution delay time difference of valve V1 just now is 30 seconds, which exceeds the preset threshold. In this case, the time trigger intervals of other signals in the subsequent device control instruction chain need to be automatically extended. For example, originally in the device control instruction chain, the instruction to "stop the P1 feed pump" was to be executed 30 seconds after valve V1 was closed. Now, the time trigger interval needs to be extended. The extended time can be determined according to the ratio by which the execution delay time difference exceeds the preset threshold. Suppose according to the ratio calculation, the time trigger interval needs to be extended by 15 seconds. Then the trigger time of the instruction to "stop the P1 feed pump" changes from 30 seconds after valve V1 was closed originally to 45 seconds after valve V1 was closed.

[0143] Step S190, adjust the power adjustment parameter in the environmental adjustment instruction according to the deviation value of the action completion degree, generate a compensated adaptive environmental control instruction, and re-inject the adaptive environmental control instruction into the dynamic decision engine to replace the unexecuted instruction item in the original emergency response strategy priority queue.

[0144] For the ventilation system, the deviation value of the action completion degree is 20%, indicating that the ventilation power has not reached the expected target. To compensate for this deviation, it is necessary to adjust the power regulation parameters of the ventilation system. Assume that the power regulation of the ventilation system is achieved by changing the speed of the motor, and there is a linear relationship between the motor speed and the ventilation power. The originally set motor speed control parameter is the parameter value that enables the ventilation power to reach 100%. Now, it is adjusted according to the deviation value of the action completion degree. Since the deviation value is 20%, in order to make the ventilation power reach 100% as soon as possible, it is necessary to appropriately increase the motor speed control parameter. For example, the originally set motor speed control parameter is 5000 revolutions per minute. After calculation, in order to compensate for the 20% deviation, the motor speed control parameter is increased to 6000 revolutions per minute, thereby generating an adaptive environment control instruction after compensation, that is, "increase the motor speed of the ventilation system to 6000 revolutions per minute".

[0145] On this basis, the original instruction regarding the ventilation system in the emergency response strategy priority queue is "increase the ventilation system power to 100%". However, since the target was not achieved during actual execution, the generated adaptive environment control instruction "increase the motor speed of the ventilation system to 6000 revolutions per minute" is now injected into the dynamic decision-making engine to replace the original instruction. The dynamic decision-making engine will re-evaluate and adjust the execution order and method of the emergency response strategy according to the new instruction to ensure that the environmental adjustment measures can be executed more effectively and reduce the risks in the oil depot.

[0146] In a possible implementation manner, the method further includes:

[0147] Step S210, after the safety control instruction is executed, collect actual risk mitigation effect data, where the actual risk mitigation effect data includes a leakage rate decline curve, a gas concentration dilution efficiency, and an equipment shutdown response time.

[0148] Taking the oil tank leakage accident as an example, before the execution of the safety control instructions, the leakage rate measured by the leakage monitoring equipment is 10 liters per minute. After the execution of the safety control instructions such as closing the relevant valves, the leakage rate is measured at regular intervals to obtain the leakage rate decline curve. For example, at the 10th minute after the instruction execution, the leakage rate drops to 5 liters per minute; at the 20th minute, it drops to 2 liters per minute. For the gas concentration dilution efficiency, before the ventilation system is started, the concentration of harmful gases in a certain area of the oil depot is 500 ppm. After the ventilation system operates according to the control instructions for a period of time, the gas concentration is measured in the same area, and the measured value is assumed to be 200 ppm. Calculate the gas concentration dilution efficiency, that is, (initial concentration - final concentration) ÷ initial concentration × 100%, which is (500 - 200) ÷ 500 × 100% = 60%. The equipment shutdown response time refers to the time from issuing the equipment shutdown instruction to the actual stop of the equipment operation. For the P1 feed pump, after issuing the instruction of "stop the P1 feed pump", after measurement, the pump completely stops running after 30 seconds, which is the equipment shutdown response time.

[0149] Step S220, compare and analyze the actual risk mitigation effect data with the predicted comprehensive risk probability, and calculate the risk prediction deviation degree and the control instruction execution efficiency.

[0150] For example, the predicted comprehensive risk probability is obtained through the previous risk assessment model. It is assumed that after the execution of the current safety control instructions, the risk probability will be reduced to 0.3. From the perspective of the actual risk mitigation effect data, the decrease in the leakage rate, the dilution of the gas concentration, etc. indicate that the actual risk has been effectively controlled, but there may be a difference from the predicted risk reduction degree. Calculate the risk prediction deviation degree. Assume that through a certain comprehensive evaluation method, the actual risk situation is converted into an equivalent risk probability value, assumed to be 0.4. The risk prediction deviation degree is the absolute value of the difference between the actual risk probability and the predicted risk probability divided by the predicted risk probability, that is, |0.4 - 0.3| ÷ 0.3 ≈ 0.33. For the control instruction execution efficiency, it is measured by indicators such as the equipment shutdown response time and the gas concentration dilution efficiency. For example, the preset standard for the equipment shutdown response time is 20 seconds, while the actual equipment shutdown response time is 30 seconds; the preset standard for the gas concentration dilution efficiency is 70%, and the actual is 60%. Considering these indicators, through the set calculation method (assuming the same weight for each indicator and simple average calculation), calculate the control instruction execution efficiency. The execution efficiency of the equipment shutdown response time is 20 ÷ 30 ≈ 0.67, and the execution efficiency of the gas concentration dilution efficiency is 60% ÷ 70% ≈ 0.86. Then the comprehensive control instruction execution efficiency is (0.67 + 0.86) ÷ 2 = 0.765.

[0151] Step S230, when the risk prediction deviation degree exceeds the allowable range, trigger the parameter update mechanism of the feature fusion model, and use the incremental learning algorithm to fuse the latest collected physical parameter sequence and the environmental index set.

[0152] Suppose the allowable range of the risk prediction deviation degree is 0.2, and the just calculated risk prediction deviation degree is 0.33, which exceeds the allowable range. At this time, trigger the parameter update mechanism of the feature fusion model. The latest collected physical parameter sequence includes parameters such as the pressure, temperature, and flow rate of the oil tank, and the environmental index set includes indexes such as gas concentration, temperature and humidity, and wind speed. For example, the pressure of the oil tank has increased after the execution of the safety control instruction, rising from the original 5 atmospheres to 5.5 atmospheres; there is a new change trend in the gas concentration in some areas. Using the incremental learning algorithm, gradually integrate these new data into the feature fusion model. The incremental learning algorithm will compare and analyze the new data with the existing knowledge structure of the model and adjust the parameters of the model. For example, for the filter parameters used to analyze the pressure parameters in the time-frequency decomposition module, according to the new pressure change situation, the cut-off frequency of the filter may be adjusted from the original 10 Hz to 12 Hz to better capture the features in the pressure signal. For the interpolation coefficient used to generate the gas concentration gradient distribution map in the spatial interpolation module, according to the new gas concentration data distribution, it may be adjusted from 0.6 to 0.7 to make the generated distribution map more accurately reflect the actual situation.

[0153] Step S240, when the execution efficiency of the control instruction is lower than the preset standard, optimize the weight allocation strategy in the risk assessment model to enhance the response sensitivity to rapidly changing risks.

[0154] Suppose the preset standard for the execution efficiency of the control instruction is 0.8, and the actual calculated execution efficiency of the control instruction is 0.765, which is lower than the preset standard. At this time, it is necessary to optimize the weight allocation strategy in the risk assessment model. When calculating the comprehensive risk probability, the risk assessment model assigns different weights to different factors, such as the weights of factors such as the degradation characteristics of equipment status and the propagation path of environmental anomalies. Through analysis, it is found that in the current situation, the degradation characteristics of equipment status have a greater impact on the risk, but the weight assigned to it in the model is relatively low. Therefore, appropriately increase the weight of the factors related to the degradation characteristics of equipment status. For example, the original weight of the degradation characteristics of equipment status is 0.3, and it is increased to 0.4, and at the same time, the weights of other factors are adjusted accordingly to ensure that the total weight is 1. After such adjustment, the model can more sensitively capture the risk changes when facing the rapid change of equipment status and improve the accuracy of risk assessment.

[0155] Step S250: Synchronize the model parameters of the updated risk assessment model to all monitoring nodes, and evaluate the warning accuracy rate of the updated risk assessment model using a progressive verification method in the next monitoring cycle.

[0156] In this embodiment, the model parameters of the updated risk assessment model include various weight values, threshold values, and other information. These parameters are synchronized to all monitoring nodes through the monitoring network of the oil depot to ensure that each monitoring node can use the latest model for risk assessment. In the next monitoring cycle, a progressive verification method is used to evaluate the warning accuracy rate. For example, at the beginning of the monitoring cycle, risk assessment and warning tests are first performed on some representative monitoring nodes. Suppose 5 monitoring nodes are selected, and the risk situations obtained by model evaluation are compared with the actual risk situations. If among these 5 monitoring nodes, the model accurately warns of the risk situations of 4 nodes, then the initial warning accuracy rate is 4÷5 = 0.8. Then gradually expand the verification scope, increase the number of monitoring nodes, and continuously evaluate the warning accuracy rate until all monitoring nodes have been verified, and finally determine the warning accuracy rate of the updated risk assessment model to ensure that the model can reliably support the safety management of the oil depot.

[0157] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of an intelligent warning system 100 for oil depot safety risks based on data fusion that can implement the idea of the present invention provided by some embodiments of the present invention. For example, the processor 120 can be used on the intelligent warning system 100 for oil depot safety risks based on data fusion and is used to execute the functions in the present invention.

[0158] The intelligent warning system 100 for oil depot safety risks based on data fusion can be a general-purpose server or a special-purpose server, and both can be used to implement the intelligent warning method for oil depot safety risks based on data fusion of the present invention. Although only one server is shown in the present invention, for convenience, the functions described in the present invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0159] For example, the intelligent early warning system 100 for oil depot safety risks based on data fusion may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the intelligent early warning system 100 for oil depot safety risks based on data fusion may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention can be implemented according to these program instructions. The intelligent early warning system 100 for oil depot safety risks based on data fusion further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0160] For ease of explanation, only one processor is described in the intelligent early warning system 100 for oil depot safety risks based on data fusion. However, it should be noted that the intelligent early warning system 100 for oil depot safety risks in the present invention may also include multiple processors. Therefore, the steps performed by one processor described in the present invention can also be jointly performed or separately performed by multiple processors. For example, if the processor of the intelligent early warning system 100 for oil depot safety risks based on data fusion performs step A and step B, it should be understood that step A and step B can also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0161] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned intelligent early warning method for oil depot safety risks based on data fusion is implemented.

[0162] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. An intelligent early warning method for oil depot safety risks based on data fusion, characterized in that, The method includes: Collecting the physical parameter sequence monitored in real time by equipment sensors in the oil depot area, the set of dynamic environmental indicators generated by the environmental monitoring device, and the abnormal event records in the historical accident database; Inputting the physical parameter sequence, the set of dynamic environmental indicators, and the abnormal event records into a pre-trained feature fusion model, and extracting the joint feature vector of the equipment state degradation feature, the environmental anomaly propagation path, and the historical accident triggering condition through cross-modal correlation analysis; Calculating the risk diffusion weight between the target monitoring node and the adjacent monitoring area according to the joint feature vector, and performing sliding weighting on the risk diffusion weight based on a dynamic time window to generate a real-time risk propagation intensity value; Invoking a pre-configured risk assessment model to perform a non-linear mapping on the real-time risk propagation intensity value and the current oil depot safety threshold, and outputting the comprehensive risk probability of the target monitoring node; Triggering corresponding safety control instructions according to the predefined risk interval where the comprehensive risk probability is located, and the safety control instructions include equipment start-stop control signals, environmental adjustment instructions, and emergency response strategies; The step of inputting the physical parameter sequence, the set of dynamic environmental indicators, and the abnormal event records into a pre-trained feature fusion model, and extracting the joint feature vector of the equipment state degradation feature, the environmental anomaly propagation path, and the historical accident triggering condition through cross-modal correlation analysis includes: Performing time-frequency domain decomposition on the physical parameter sequence, and extracting the low-frequency energy ratio in the pressure fluctuation spectrum, the extreme value of the second derivative of the temperature change curve, and the amplitude-frequency characteristics of the flow pulse signal as the equipment state degradation feature; Performing spatial interpolation calculation on the set of dynamic environmental indicators, generating a gas concentration gradient distribution map, a temperature and humidity field strength matrix, and a wind speed vector field, and extracting the area where the gradient change rate exceeds a preset threshold as the environmental anomaly propagation path; Performing causal chain analysis on the abnormal event records, and extracting the equipment failure mode, the environmental mutation critical value, and the operation response delay time that cause the accident trigger as the historical accident triggering condition; Inputting the equipment state degradation feature, the environmental anomaly propagation path coordinates, and the historical accident triggering condition into the multi-head attention module in the feature fusion model, and calculating the spatial correlation weight between the equipment state degradation feature and the environmental anomaly propagation path, the time dependence weight between the equipment state degradation feature and the historical accident triggering condition, and the causal association weight between the environmental anomaly propagation path and the historical accident triggering condition; Performing feature splicing and dimensionality reduction processing on the equipment state degradation feature, the environmental anomaly propagation path, and the historical accident triggering condition according to the spatial correlation weight, the time dependence weight, and the causal association weight, and generating a joint feature vector containing spatio-temporal causal relationships.

2. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 1, characterized in that, The pre-trained feature fusion model is trained through the following steps: Extracting the physical parameter sample sequence, the set of environmental indicator samples, and the historical accident sample set labeled with accident triggering conditions under normal working conditions from the sample database as training input data; Input the physical parameter sample sequence into the time-frequency decomposition module of the feature fusion model to generate an initial feature matrix of the device state degradation features. Input the environmental index sample set into the spatial interpolation module to generate an initial gradient tensor of the environmental anomaly propagation path. Input the historical accident sample set into the causal analysis module to extract an initial causal vector of the historical accident triggering conditions; Input the initial feature matrix, the initial gradient tensor, and the initial causal vector into the multi-head attention module to calculate the spatio-temporal correlation weights between the device state degradation features and the environmental anomaly propagation path, and the causal mapping weights between the environmental anomaly propagation path and the historical accident triggering conditions; Perform weighted splicing on the initial feature matrix according to the spatio-temporal correlation weights to generate an enhanced representation vector of the device state degradation features. Dynamically screen the initial causal vector according to the causal mapping weights to generate a conditional screening vector of the historical accident triggering conditions. Perform channel fusion on the enhanced representation vector and the initial gradient tensor to generate a path fusion matrix of the environmental anomaly propagation path; Perform cross-modal splicing on the path fusion matrix and the conditional screening vector to generate a joint feature training vector. Compare the similarity between the joint feature training vector and a preset verification label vector, calculate the multi-dimensional feature loss value between the joint feature training vector and the verification label vector, and adjust the filter parameters of the time-frequency decomposition module, the interpolation coefficients of the spatial interpolation module, and the causal chain extraction threshold of the causal analysis module according to the multi-dimensional feature loss value; Input the new joint feature vector output by the adjusted feature fusion model into the discriminator network to discriminate the class probabilities of the new joint feature vector belonging to normal working conditions or historical accidents; Optimize the number of attention heads of the multi-head attention module and the scaling factor of channel fusion according to the difference value between the class probability and the true working condition label until the discriminator network cannot distinguish the joint feature distributions of normal working conditions and historical accidents.

3. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 2, wherein Before extracting the physical parameter sample sequence, environmental index sample set, and historical accident sample set with labeled accident triggering conditions under normal working conditions from the sample database as training input data, it also includes: Perform moving average filtering on the physical parameter sample sequence to remove impulse interference points and generate a stationary physical parameter sequence; Perform Kriging spatial interpolation on the discrete monitoring point data in the environmental index sample set to generate an environmental index surface with continuous spatial coverage; Convert the text records in the historical accident sample set into timestamped binary event trigger vectors according to a preset event coding table; Align the stationary physical parameter sequence with the environmental index surface at the second-level timestamp to generate a spatio-temporally matched physical-environment joint input matrix; Perform temporal slice splicing on the binary event trigger vector and the physical-environment joint input matrix within a set time period before and after the event to generate a labeled cross-modal training sample set.

4. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 1, wherein Calculating the risk diffusion weight between the target monitoring node and the adjacent monitoring area according to the combined feature vector, and performing sliding weighting on the risk diffusion weight based on a dynamic time window to generate a real-time risk propagation intensity value, including: Extracting the time-frequency domain energy distribution matrix corresponding to the device state degradation feature and the gradient change rate tensor corresponding to the environmental anomaly propagation path from the combined feature vector; Determining the cumulative degradation amount of the device state degradation feature within a preset time window according to the low-frequency energy proportion in the time-frequency domain energy distribution matrix, and calculating the diffusion rate of the environmental anomaly propagation path of the adjacent monitoring area based on the regional coordinates exceeding the preset threshold in the gradient change rate tensor; Obtaining the pipeline connection density parameter, the valve control logic relationship mapping table, and the turbulence coefficient in the gas diffusion physical model between the target monitoring node and each adjacent monitoring area, and superimposing the pipeline connection density parameter and the turbulence coefficient according to a preset ratio to generate an initial diffusion weight base value; Inputting the initial diffusion weight base value into a dynamic weight allocation network, where the first hidden layer of the dynamic weight allocation network receives the sequential splicing vector of the cumulative degradation amount and the diffusion rate, and the second hidden layer receives the priority marking sequence in the valve control logic relationship mapping table; In the dynamic weight allocation network, using a temporal convolutional kernel to extract local periodic patterns from the sequential splicing vector to generate a time-sensitive factor, and simultaneously using a graph attention mechanism to calculate the adjacency node correlation degree of the priority marking sequence to generate a spatial dependence factor; Performing element-wise multiplication on the time-sensitive factor and the spatial dependence factor to obtain a dynamic diffusion correction coefficient, and performing weighted summation of the dynamic diffusion correction coefficient and the initial diffusion weight base value to generate the real-time risk propagation sub-intensity from each adjacent monitoring area to the target monitoring node; According to the sliding step of the dynamic time window, superimposing the real-time risk propagation sub-intensity within the current time window and the attenuation residual intensity of the previous window, and performing normalization processing on the superimposed result to generate the real-time risk propagation intensity value of the target monitoring node; Among them, the length of the dynamic time window is dynamically adjusted according to the low-frequency component proportion in the time-frequency domain energy distribution matrix of the device state degradation feature. When the low-frequency component proportion exceeds the preset critical value, the window length is shortened to enhance risk sensitivity.

5. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 1, characterized in that, Invoking a pre-configured risk assessment model to perform non-linear mapping on the real-time risk propagation intensity value and the current oil depot safety threshold, and outputting the comprehensive risk probability of the target monitoring node, including: Performing normalization processing on the real-time risk propagation intensity value in the time dimension to generate a standardized risk intensity sequence; Decomposing the current oil depot safety threshold into a dynamic threshold vector according to a preset risk level interval, where the dynamic threshold vector includes the upper and lower limits of the thresholds corresponding to each risk level; Inputting the standardized risk intensity sequence and the dynamic threshold vector into the multi-layer perceptron of the risk assessment model, and calculating the non-linear weighted sum between the standardized risk intensity sequence and the dynamic threshold vector through an activation function; Perform probability transformation on the non - linear weighted sum to generate a preliminary risk probability distribution; Adjust the preliminary risk probability distribution according to the historical scenario probability correction coefficient matching the current joint feature vector in the historical accident database, and output the comprehensive risk probability of the target monitoring node.

6. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 1, characterized in that The triggering of the corresponding safety control instruction according to the predefined risk interval where the comprehensive risk probability is located includes: Perform interval matching between the comprehensive risk probability and each level threshold in the predefined risk interval to determine the current risk level identifier of the target monitoring node; Retrieve the corresponding set of equipment start - stop control signals, environmental adjustment instruction sequence, and emergency response strategy priority queue from the preset instruction mapping table according to the current risk level identifier; Based on the pipeline topological connection relationship of the target monitoring node, sort the execution order of the set of equipment start - stop control signals to generate an equipment control instruction chain with timing constraints; According to the spatial distribution characteristics of the environmental anomaly propagation path, perform path - optimized allocation on the start coordinates of the gas extraction device and the ventilation system power adjustment parameters in the environmental adjustment instruction sequence; Input the emergency response strategy priority queue into the dynamic decision - making engine, and combine the spatio - temporal change gradient of the real - time risk propagation intensity value to adjust the triggering timing and coverage range of the emergency response strategy; According to the execution feedback signal of the equipment control instruction chain and the response delay time of the environmental adjustment instruction sequence, dynamically update the instruction execution interval in the emergency response strategy priority queue, generate the final safety control instruction set, and send it to the target execution terminal.

7. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 6, characterized in that After generating the final safety control instruction set and sending it to the target execution terminal, it further includes: Real - time capture the valve opening feedback signal and the actual ventilation power value returned by the execution terminal to generate an equipment response status code; Perform dynamic time warping matching between the equipment response status code and the pre - stored ideal response curve, and calculate the execution delay time difference and the action completion degree deviation value; When the execution delay time difference exceeds the preset threshold, automatically extend the time trigger interval of each signal in the subsequent equipment control instruction chain; Adjust the power adjustment parameter in the environmental adjustment instruction according to the action completion degree deviation value to generate a compensated adaptive environmental control instruction; Re - inject the adaptive environmental control instruction into the dynamic decision - making engine to replace the unexecuted instruction items in the original emergency response strategy priority queue.

8. The intelligent early warning method for oil depot safety risks based on data fusion according to claim 1, characterized in that The method further includes: After the execution of the safety control instruction, collect the actual risk mitigation effect data, where the actual risk mitigation effect data includes the leakage rate decline curve, gas concentration dilution efficiency, and equipment shutdown response time; Compare and analyze the actual risk mitigation effect data with the predicted comprehensive risk probability, and calculate the risk prediction deviation degree and the control instruction execution efficiency; When the risk prediction deviation degree exceeds the allowable range, trigger the parameter update mechanism of the feature fusion model, and use the incremental learning algorithm to fuse the latest collected physical parameter sequence and environmental index set; When the execution efficiency of the control instruction is lower than the preset standard, optimize the weight allocation strategy in the risk assessment model to enhance the response sensitivity to rapidly changing risks; Synchronize the model parameters of the updated risk assessment model to all monitoring nodes, and use a progressive verification method to evaluate the early warning accuracy of the updated risk assessment model in the next monitoring cycle.

9. An intelligent early warning system for oil depot safety risks based on data fusion, characterized in that, The intelligent early warning system for oil depot safety risks based on data fusion includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the intelligent early warning method for oil depot safety risks based on data fusion according to any one of claims 1-8 above.

Images