Multi-sensor fusion automatic early warning system for marine engine room fire

Abstract

This invention relates to the field of fire early warning technology and discloses a multi-sensor fusion automatic early warning system for ship engine room fires. The system includes: a data analysis unit, a baseline adjustment unit, a data correction unit, an anomaly analysis unit, a collaborative computing unit, and a fire early warning unit. The data analysis unit generates a global operating condition coefficient that accurately reflects the thermal state of the engine room. The baseline adjustment unit generates a dynamic sensing baseline that adapts to changes in operating conditions in real time. The data correction unit eliminates interference and generates refined anomaly signals. The anomaly analysis unit generates a single-dimensional burst intensity index using parameters such as abnormal energy flux to achieve accurate scoring. The collaborative computing unit and the fire early warning unit cooperate to generate fire early warning commands, improving the accuracy of multi-sensor collaborative fire identification and enhancing the adaptability of this early warning scheme to dynamic operating conditions, thus ensuring the accuracy of engine room fire monitoring.

Description

Technical Field

[0001] This invention relates to the field of fire early warning technology, specifically to a multi-sensor fusion automatic early warning system for fires in ship engine rooms. Background Technology

[0002] Currently, ships typically use multiple types of sensors to collect different data during navigation, such as temperature and smoke. Then, based on the data collected under the ship's no-load conditions, and the normal values ​​of each sensor are used to set warning thresholds, when the corresponding data exceeds the threshold, it is determined that a fire has occurred and an automatic warning is issued.

[0003] However, the above-mentioned early warning methods still have the following drawbacks when applied to ship engine rooms: the ship's engine room is a dynamic environment, and the operating conditions of the engine room will drift with the sailing status during the ship's navigation. For example, when the main engine is running at full load, the overall temperature of the engine room is higher than when it is unloaded. At this time, using the early warning threshold set by the ship based on the unloaded operating conditions can easily misjudge the normal operating temperature as a fire signal, resulting in frequent false early warnings and affecting the effectiveness of the early warning. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multi-sensor fusion automatic early warning system for ship engine room fires, which solves the aforementioned problems.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] A multi-sensor fusion automatic early warning system for ship engine room fires includes:

[0007] The data analysis unit is used to acquire real-time sensor data from multiple sensors in the target warning area and the operating condition data of the host in the target warning area, process the real-time operating condition data, and generate global operating condition coefficients. The target warning area is the engine room of the ship.

[0008] The baseline adjustment unit is used to acquire the static thresholds of each sensor within the target warning area, adjust the static thresholds according to the global operating condition coefficient, and generate a dynamic sensing baseline for each sensor.

[0009] The data correction unit is used to compare the sensor data of each sensor with the corresponding dynamic sensing baseline, eliminate data drift, and generate a refined abnormal signal for each sensor.

[0010] The anomaly analysis unit is used to analyze the burst intensity of the refined anomaly signal of each sensor and generate a one-dimensional burst intensity index.

[0011] The collaborative computing unit is used to collaboratively calculate the single-dimensional burst intensity index corresponding to all sensors and the global operating condition coefficient to generate collaborative early warning threshold and collaborative mutation energy.

[0012] The fire early warning unit is used to compare the collaborative early warning threshold and the collaborative mutation energy to generate a fire early warning command.

[0013] Furthermore, the real-time operating data is processed to generate global operating coefficients, including:

[0014] The host speed and host output power in the real-time operating data are calculated, the heat work of the host on the nacelle is analyzed, and the instantaneous heat load value is generated.

[0015] The airflow in the cabin is acquired in real time. Based on the real-time airflow and instantaneous heat load value, the heat retention rate in the cabin is analyzed, and the cabin thermal inertia is generated.

[0016] Furthermore, the real-time operating data is processed to generate global operating coefficients, which also includes:

[0017] Analyze the instantaneous fluctuations of the main unit's exhaust temperature in real-time operating data to generate a combustion oscillation index;

[0018] The cabin thermal stability is obtained by fusing the cabin thermal inertia and the combustion oscillation index.

[0019] The instantaneous heat load value and the thermal stability of the engine room are calculated together to generate a global operating condition coefficient.

[0020] Furthermore, the static threshold is adjusted based on the global operating condition coefficient to generate a dynamic sensing baseline for each sensor, including:

[0021] Based on the global operating condition coefficient, the influence of heat flow and air turbulence in the cabin is analyzed to generate disturbance entropy value;

[0022] Based on the instantaneous heat load value and cabin thermal inertia, the impact load on sensor stability is analyzed, and its influence on the static threshold of each sensor is determined, generating the load transient gradient.

[0023] By integrating the combustion oscillation index and the load transient gradient, field-induced coupling weights for different sensors are generated.

[0024] Furthermore, the static threshold is adjusted based on the global operating condition coefficient to generate a dynamic sensing baseline for each sensor, which also includes:

[0025] The adaptive offset is obtained by fusing the perturbation entropy value, the load transient gradient, and the field-induced coupling weight.

[0026] The adaptive offset is dynamically calibrated based on the static threshold of each sensor to generate a dynamic sensing baseline.

[0027] Furthermore, for each sensor, its sensor data is compared with the corresponding dynamic sensing baseline, and data drift is eliminated to generate a refined anomaly signal for each sensor, including:

[0028] For each sensor, calculate the real-time difference sequence between its real-time data and the dynamic sensing baseline, analyze the statistical boundary of the real-time difference sequence in the time domain based on the field-induced coupling weight, and generate anomaly judgment confidence interval;

[0029] Analyze abrupt changes in the real-time difference sequence outside the anomaly detection confidence interval and generate event significance values;

[0030] The confidence interval for anomaly determination is calculated in conjunction with the event significance value to obtain a refined anomaly signal after removing drift and impulse interference.

[0031] Furthermore, the burst intensity of the refined abnormal signal from each sensor is analyzed to generate a one-dimensional burst intensity index, including:

[0032] Analyze the refining anomaly signals to generate anomaly energy flux;

[0033] Identify the critical crossing points of refined abnormal signals, analyze the critical crossing points, and generate morphological sharpness factors.

[0034] Track the evolution trend of refined abnormal signals over multiple consecutive time segments and generate trend coherence coefficients;

[0035] By fusing the abnormal energy flux, morphological sharpness factor, and trend coherence coefficient, a one-dimensional burst intensity index representing the burst intensity of anomaly signals from a single sensor is generated.

[0036] Furthermore, the single-dimensional burst intensity index corresponding to all sensors is collaboratively calculated with the global operating condition coefficient to generate a collaborative early warning threshold and a collaborative mutation energy, including:

[0037] The temporal relationship between the one-dimensional burst intensity indices of different sensors is analyzed, a heterogeneous signal network is constructed, and the network correlation entropy is generated by calculating the heterogeneous signal network.

[0038] Based on the thermal stability of the engine room, the weights of various connection edges in the heterogeneous signal network are dynamically adjusted to generate the operating condition modulation conductance.

[0039] Based on network correlation entropy and operating condition modulation transmissivity, anomalies of multiple sensors are analyzed to generate global instability.

[0040] By integrating global instability with cabin thermal inertia, the energy accumulation of fire in the cabin is analyzed, and an environmental gain factor that dynamically adjusts the early warning sensitivity is generated.

[0041] Furthermore, the single-dimensional burst intensity index corresponding to all sensors is collaboratively calculated with the global operating condition coefficient to generate a collaborative early warning threshold and a collaborative mutation energy, which also includes:

[0042] The network correlation entropy, operating condition modulation transmissivity, global instability and environmental gain factor are fused to generate a collaborative early warning threshold;

[0043] Based on the heterogeneous signal network, the one-dimensional burst intensity index and environmental gain factor of all sensors are analyzed to generate a cooperative mutation energy representing cabin anomalies.

[0044] Furthermore, by comparing the collaborative early warning threshold and the collaborative mutation energy, a fire early warning instruction is generated, including:

[0045] When the collaborative mutation energy is greater than or equal to the collaborative early warning threshold and lasts for more than 3 seconds, a fire early warning command is automatically generated.

[0046] Otherwise, generate a security instruction.

[0047] In summary, the present invention has the following main beneficial effects:

[0048] By deeply mining operating data such as main engine speed and output power, the generated global operating condition coefficient can provide a comprehensive understanding of the dynamic thermal state of the ship's engine room. By fusing parameters such as disturbance entropy, load transient gradient, and field-induced coupling weight, a dynamic sensing baseline is generated, enabling adaptive calibration of static threshold operating conditions. The data correction unit efficiently removes pulse interference and data drift based on the anomaly judgment confidence interval and event significance value, obtaining refined anomaly signals.

[0049] By fusing anomalous energy flux, morphological sharpness factor, and trend coherence coefficient from the anomaly analysis unit, a one-dimensional burst intensity index is generated, enabling accurate scoring of anomalous signals from a single sensor. The collaborative computing unit constructs a heterogeneous signal network and combines parameters such as network correlation entropy and operating condition modulation transmissibility to generate collaborative early warning thresholds and collaborative mutation energy, achieving collaborative identification of anomalies from multiple sensors. Finally, through the generation of fire early warning commands by the fire early warning unit, this solution improves the accuracy and anti-interference capability of fire early warning, reduces the false alarm rate under dynamic operating conditions, and adapts to the drift characteristics of cabin operating conditions. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the multi-sensor fusion automatic early warning system for ship engine room fires according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] refer to Figure 1 A multi-sensor fusion automatic early warning system for ship engine room fires, including:

[0053] The data analysis unit is used to acquire real-time sensor data from multiple sensors in the target warning area and the operating condition data of the main engine (i.e., the engine) in the target warning area. It processes the real-time operating condition data to generate global operating condition coefficients. The target warning area is the engine room of the ship.

[0054] Data from multiple sensors includes: temperature, smoke concentration, carbon monoxide concentration, etc.

[0055] Operating data includes: main unit speed, main unit output power, main unit exhaust temperature, etc.

[0056] The baseline adjustment unit is used to acquire the static thresholds of each sensor within the target warning area, adjust the static thresholds according to the global operating condition coefficient, and generate a dynamic sensing baseline for each sensor.

[0057] The data correction unit is used to compare the sensor data of each sensor with the corresponding dynamic sensing baseline, eliminate data drift, and generate a refined abnormal signal for each sensor.

[0058] The anomaly analysis unit is used to analyze the burst intensity of the refined anomaly signal of each sensor and generate a one-dimensional burst intensity index.

[0059] The collaborative computing unit is used to collaboratively calculate the single-dimensional burst intensity index corresponding to all sensors and the global operating condition coefficient to generate collaborative early warning threshold and collaborative mutation energy.

[0060] The fire early warning unit is used to compare the collaborative early warning threshold and the collaborative mutation energy to generate a fire early warning command.

[0061] In one embodiment, real-time operating data is processed to generate global operating coefficients, including:

[0062] The main engine speed and output power in the real-time operating data are calculated to analyze the heat work of the main engine on the engine room and generate an instantaneous heat load value. Specifically, this includes: obtaining the rated power and rated speed of the main engine; dividing the main engine speed by the rated speed to obtain the speed value; dividing the main engine output power by the rated power to obtain the power value; and then multiplying the speed value and the power value to obtain the original heat work index.

[0063] Then, the product of the square of the speed rating and the reciprocal of the power rating is calculated as the heat loss compensation factor. Finally, the original heat work index, the heat loss compensation factor, and the rated power are multiplied to obtain the heat work contribution of the main engine to the engine room per unit time, which is the instantaneous heat load value. The instantaneous heat load value is mainly used to reflect the basic heat load level of the engine room under dynamic operating conditions, because the switching of the main engine from no load to full load during ship navigation will cause drastic changes in the heat load of the engine room. This avoids misjudging the normal heat load when the main engine is running at full load as the heat anomaly in the early stage of a fire.

[0064] The air volume in the cabin is acquired in real time. Based on the real-time air volume and instantaneous heat load value, the heat retention rate in the cabin is analyzed and the cabin thermal inertia is generated. Specifically, the real-time air volume is divided by the instantaneous heat load value to obtain the air volume heat dissipation coefficient, and the reciprocal of the air volume heat dissipation coefficient is calculated to obtain the real-time heat retention rate.

[0065] By obtaining the net volume of the engine room and multiplying the real-time heat retention rate by the net volume, the thermal inertia of the engine room can be obtained. The thermal inertia of the engine room is mainly used to distinguish between the slow accumulation of normal heat load and the rapid heat accumulation of fire. Under dynamic operating conditions, the normal heat generation of the main engine will cause the heat to be slowly retained, but a fire will cause the heat to accumulate rapidly. This avoids the slow temperature rise caused by the normal heat retention under dynamic operating conditions being mistaken for a fire, and at the same time, it captures the rapid heat accumulation signal unique to fire.

[0066] In one embodiment, processing real-time operating data to generate global operating coefficients further includes:

[0067] The instantaneous fluctuation of the main engine exhaust temperature in the real-time operating data is analyzed to generate a combustion oscillation index. Specifically, this includes: continuously sampling the real-time main engine exhaust temperature, calculating the temperature difference between adjacent sampling points, and recording the sign of each difference, which can be positive, negative, or zero, thereby forming a symbol sequence. The number of segments in the symbol sequence that have the same non-zero sign consecutively is counted. A sign that appears twice consecutively is considered consecutive. The reciprocal of the number of segments is then used as the combustion stability factor.

[0068] The standard deviation of all sampled temperature values ​​within the same time period is calculated and used as the temperature oscillation intensity. The combustion stability factor is then multiplied by the temperature oscillation intensity to obtain the combustion oscillation index. The combustion oscillation index is used to identify abnormal fluctuations in the combustion status of the main engine in the ship's cabin. Under normal dynamic operating conditions, the combustion fluctuation of the main engine is within a stable range; however, a fire may cause abnormal combustion of the main engine, such as local high temperature affecting combustion. In this case, the combustion oscillation index will increase significantly.

[0069] The cabin thermal stability is obtained by fusing the cabin volume thermal inertia and the combustion oscillation index, specifically including:

[0070] The arithmetic square root of the cabin's thermal inertia is calculated to obtain the thermal inertia gain, which represents the cabin's ability to buffer thermal shocks. Simultaneously, the reciprocal of the combustion oscillation index is calculated to obtain the oscillation damping factor. The geometric mean of the thermal inertia gain and the oscillation damping factor is then calculated to obtain the cabin thermal stability. Cabin thermal stability is used to comprehensively assess whether the cabin's thermal environment is in a stable state. Under dynamic operating conditions, high cabin thermal stability indicates that the cabin's thermal environment can buffer normal thermal fluctuations; low cabin thermal stability indicates that the thermal environment is more sensitive, and minor thermal anomalies may be precursors to fire.

[0071] The instantaneous heat load value and the thermal stability of the engine room are calculated together to generate a global operating condition coefficient. Specifically, this includes: calculating the ratio of the instantaneous heat load value to the thermal stability of the engine room, and using the ratio as the heat load stability ratio, which mainly reflects the tension of the current heat load relative to the thermal buffer capacity; then calculating the arithmetic square root of the rated power of the main engine, which is used as the power reference factor; dividing the heat load stability ratio by the power reference factor to obtain the global operating condition coefficient, which is mainly used to reflect the overall thermal state and potential risks of the engine room.

[0072] By accurately processing real-time operating data such as engine speed and output power, instantaneous heat load values ​​are generated, which accurately reflect the dynamic changes in the basic heat load of the engine room, avoid misjudgment of heat load caused by engine load switching, and distinguish between normal heat accumulation and rapid heat accumulation in fire based on the thermal inertia of the engine room. Combustion oscillation index is used to identify abnormalities in engine combustion, while engine room thermal stability can assess the thermal environment buffering capacity. Finally, the global operating condition coefficient is used to achieve an accurate assessment of the overall thermal state and potential risks of the engine room, improving the accuracy and reliability of fire early warning.

[0073] In one embodiment, adjusting the static threshold based on the global operating condition coefficient to generate a dynamic sensing baseline for each sensor includes:

[0074] Based on the global operating condition coefficient, the influence of heat flow and air turbulence in the cabin is analyzed to generate a disturbance entropy value. Specifically, the global operating condition coefficient is used as the heat flow intensity value, and the reciprocal of the cabin thermal stability is used as the turbulence potential coefficient. The turbulence potential coefficient represents the increasing trend of air disturbance when the thermal environment is unstable. The natural logarithm of the product of the heat flow intensity value and the turbulence potential coefficient is calculated to obtain the disturbance entropy value. The disturbance entropy value mainly reflects the degree of disturbance of the cabin heat flow and air turbulence to the sensor.

[0075] Based on the instantaneous heat load value and cabin thermal inertia, the impact load on sensor stability is analyzed, and its influence on the static threshold of each sensor is determined. The load transient gradient is generated, which specifically includes: real-time acquisition of instantaneous heat load value and cabin thermal inertia for three consecutive sampling periods, with the sampling period set to 0.5 seconds; calculation of the absolute value of the difference between instantaneous heat load values ​​in adjacent periods to obtain the heat load transient rate; and calculation of the ratio of the difference between cabin thermal inertia in adjacent periods to the cabin thermal inertia in the previous period to obtain the cabin thermal inertia decay rate.

[0076] Multiply the thermal load transient rate by a weight of 0.6 and the cabin thermal inertia decay rate by 0.4, and normalize the calculation result to the 0-1 range to obtain the sensor impact load coefficient. The sensor impact load coefficient represents the degree of influence of the impact load on the stability of the sensor.

[0077] Next, multiply the sensor's impact load coefficient by the corresponding sensor's static threshold, and normalize the product to the 0-1 range. This is the load transient gradient. The load transient gradient mainly reflects the impact of impact load on the sensor under dynamic working conditions, avoiding misjudging sudden changes in thermal load under normal working conditions as fires, while ensuring that drastic changes caused by fires can be accurately captured.

[0078] Among them, the transient rate of heat load mainly represents the magnitude of sudden changes in cabin heat load. Since such sudden changes may originate from core risk sources such as abnormal heating of the main engine or potential fire, it is the main cause that directly impacts the stability of sensors and causes their data to deviate from the true value. Therefore, it is given a higher weight of 0.6. The cabin thermal inertia decay rate mainly reflects the cabin's ability to buffer thermal shock and represents the dynamic change of this buffering ability. However, its essence is a derivative effect of cabin space characteristics, not the risk source itself. Therefore, it is given a secondary weight of 0.4.

[0079] By fusing the combustion oscillation index and the load transient gradient, field-induced coupling weights for different sensors are generated. Specifically, this includes: calculating the logarithm of the combustion oscillation index with the natural constant e as the base to obtain the combustion pulsation intensity; simultaneously performing a hyperbolic tangent function transformation on the load transient gradient to obtain the gradient influence saturation; and calculating the product of the combustion pulsation intensity and the gradient influence saturation to obtain the original field coupling energy.

[0080] Calculate the absolute value of the difference between the current period and the previous period for each sensor, divide the absolute value by its corresponding static threshold, and obtain the signal autovariability of the sensor.

[0081] Using the signal self-variance rate as an exponent, the original field coupling energy is exponentially operated on, and the calculation result is normalized to the 0-1 interval, which is the field-induced coupling weight of each sensor. The field-induced coupling weight is used to reflect the degree of influence of the dynamic environment on different sensors.

[0082] In one embodiment, adjusting the static threshold based on the global operating condition coefficient to generate a dynamic sensing baseline for each sensor further includes:

[0083] The perturbation entropy, load transient gradient, and field-induced coupling weights are fused to obtain an adaptive offset. Specifically, this involves multiplying the perturbation entropy by the load transient gradient and using the fourth root of the product as the basic perturbation factor. Simultaneously, the arcsine of the field-induced coupling weights is calculated to obtain the weight modulation coefficient.

[0084] The original offset is obtained by adding the basic disturbance factor and the weighted modulation coefficient and then multiplying it by the reciprocal of the cabin net volume. The absolute value of the original offset is divided by the cube root of the main engine's rated power, and the calculation result is normalized to the 0-1 interval.

[0085] The adaptive offset is dynamically calibrated based on the static threshold of each sensor to generate a dynamic sensing baseline. Specifically, this includes: multiplying the adaptive offset by the field-induced coupling weight of the corresponding sensor and taking the cube root of the product as the individualized offset gain; and adding 1 to the absolute value of the current signal self-variation rate of the sensor to obtain the signal activity coefficient.

[0086] The geometric mean of the individualized offset gain and the signal activity coefficient is then calculated to obtain the dynamic modulation factor. The static threshold of each sensor is multiplied by (1 + dynamic modulation factor) to obtain the dynamic sensing baseline of the sensor. The dynamic sensing baseline is used to dynamically raise the alarm baseline of each sensor, thus solving the false alarm caused by the fixed threshold.

[0087] By accurately reflecting the degree of disturbance of heat flow and turbulence to the sensor through the perturbation entropy value, the load transient gradient can distinguish between normal heat load changes and violent fire changes, the field-induced coupling weight realizes individualized adaptation to the degree of influence of different sensor environments, and the dynamic sensing baseline can accurately follow the drift of the cabin operating conditions and adjust in real time, which not only avoids false alarms under normal operating conditions, but also ensures the accurate capture of fire signals, improves the pertinence and reliability of the warning, and strengthens the adaptability and safety of cabin fire monitoring.

[0088] In one embodiment, for each sensor, its sensor data is compared with the corresponding dynamic sensing baseline, and data drift is eliminated to generate a refined anomaly signal for each sensor, including:

[0089] For each sensor, the real-time difference sequence between its real-time data and the dynamic sensing baseline is calculated. Based on the field-induced coupling weight analysis, the statistical boundary of the real-time difference sequence in the time domain is analyzed to generate anomaly judgment confidence intervals. Specifically, this includes: for each sensor, calculating the difference between its real-time data and the dynamic sensing baseline, combining multiple differences to form a real-time difference sequence; calculating the mean and median of the real-time difference sequence, and using the natural logarithm of the absolute difference between the mean and the median as a discrete anchor point.

[0090] Perform a base-2 logarithmic operation on the field-induced coupling weights, and use the reciprocal of the result plus 1 as the weight influence attenuation coefficient; multiply the maximum value in the real-time difference sequence by the weight influence attenuation coefficient to obtain the positive fluctuation characteristic quantity;

[0091] Divide the minimum value in the real-time difference sequence by the weighted influence attenuation coefficient to obtain the negative fluctuation feature. Then multiply the discrete anchor point, the positive fluctuation feature, and the negative fluctuation feature, and use the cube root of the product as the boundary expansion base value.

[0092] Finally, by adding and subtracting the boundary expansion base value from the mean of the real-time difference sequence, two boundary values ​​are obtained. These two boundary values ​​constitute the anomaly detection confidence interval of the sensor. The anomaly detection confidence interval is used to provide boundaries for anomaly detection. Under dynamic operating conditions, the normal fluctuation range of sensor data will change. The anomaly detection confidence interval is adjusted in real time with the dynamic operating conditions to ensure that normal fluctuation data falls within the interval, while real abnormal data will exceed the anomaly detection confidence interval.

[0093] Analyze abrupt change points in the real-time difference sequence outside the anomaly detection confidence interval and generate event significance values. Specifically, this includes: for each data point in the real-time difference sequence outside the anomaly detection confidence interval, treat it as an abrupt change point, calculate the absolute value of the difference between the real-time difference of the abrupt change point and the anomaly detection confidence interval, where if the real-time difference is greater than the upper boundary value of the anomaly detection confidence interval, calculate the absolute value of the difference between the real-time difference and the upper boundary value; if the real-time difference is less than the lower boundary value of the anomaly detection confidence interval, calculate the absolute value of the difference between the real-time difference and the lower boundary value.

[0094] Divide the absolute value by the boundary expansion base value to obtain the single-point boundary crossing intensity; at the same time, count the total number of mutation points in each of the three sampling periods before and after the mutation point, and multiply this total number by the sampling period (0.5 seconds) to obtain the neighborhood event density.

[0095] Then, the sign of the difference between the mutation point and the previous sampling point is calculated and compared with the product of the signs of the two consecutive differences before that. If the signs are opposite, the trend reversal factor is 2; if the signs are the same, the trend reversal factor is 1. The sign is represented as positive, negative or zero.

[0096] The single-point boundary crossing intensity, neighborhood event density, and trend reversal factor are multiplied together, and the absolute value of the result is taken as the event significance value of the mutation point. The event significance value is used to reflect the degree of abnormal significance of the mutation point, thereby distinguishing between pulse interference, data drift and real abnormal mutations under dynamic conditions. For example, the event significance value is low for instantaneous fluctuations of sensors caused by airflow disturbance (pulse interference), while the event significance value is high for continuous mutation events caused by fire, ensuring that the subsequent analysis is an abnormal signal that truly reflects the fire risk.

[0097] The confidence interval for anomaly detection is calculated in conjunction with the event significance value to obtain a refined anomaly signal after removing drift and impulse interference. Specifically, this involves using 1.8 times the mean of the event significance value as a threshold. For each abrupt change, if the event significance value exceeds this threshold, the continuity of data before and after the abrupt change is analyzed.

[0098] If the data from the previous sampling period before the mutation point is outside the anomaly confidence interval, it is determined to be the starting point of a continuous anomaly.

[0099] If the data in the next sampling period after the mutation point is still outside the anomaly detection confidence interval, then further confirmation is needed. A mutation point that meets one of the following conditions is considered a valid anomaly:

[0100] First, if the event significance value of the mutation point exceeds twice the threshold, it is considered a strong abnormal signal;

[0101] Second, the event's significance value exceeds the threshold, and within three consecutive sampling periods, at least two sampling points have values ​​outside the anomaly judgment confidence interval;

[0102] Abrupt points that do not meet the above conditions are identified as pulse interference or data drift and are eliminated. Finally, the effective anomalies are used as signal values ​​and integrated according to the time series to obtain refined anomaly signals. The refined anomaly signals are the effective anomaly signals after eliminating pulse interference and data drift, retaining the real anomaly signals related to fire, eliminating invalid interference signals under dynamic operating conditions, and reducing false alarms caused by faults or transient interference.

[0103] By adjusting the confidence interval for anomaly detection in real time with the field-induced coupling weight, it accurately adapts to the normal fluctuation range of the sensor. Furthermore, the event significance value is precisely quantified by multiple dimensions such as single-point boundary crossing intensity and neighborhood event density to accurately distinguish the degree of anomaly at the mutation point, effectively differentiating pulse interference, data drift and real anomalies. The resulting refined anomaly signal can accurately retain real fire-related anomalies, reduce false alarms caused by interference under dynamic operating conditions, and improve the accuracy of fire early warning.

[0104] In one embodiment, the burst intensity of the refined abnormal signal of each sensor is analyzed to generate a one-dimensional burst intensity index, including:

[0105] The refining anomaly signal is analyzed to generate anomaly energy flux. Specifically, this includes: setting a sliding analysis window with a length of 5 seconds, calculating the sum of the absolute values ​​of all signal values ​​in the refining anomaly signal within the window as the original cumulative amount; identifying the inflection point where the slope sign changes within the window, taking the time span between adjacent inflection points as a half-wave period, calculating the standard deviation of all half-wave periods, and taking its reciprocal as the rhythm regularization factor.

[0106] Next, calculate the absolute value of the difference between each two adjacent signal values ​​in the refined abnormal signal within the window, and calculate the variance of the three largest values ​​to obtain the gradient eddy current. Multiply the original cumulative amount, the rhythm regularization factor, and the gradient eddy current together to obtain the abnormal energy flux of the window. The abnormal energy flux mainly reflects the energy accumulation intensity of the abnormal signal, distinguishing between slight abnormal energy under normal operating conditions and severe abnormal energy caused by fire, and avoiding misjudging slight abnormal operating conditions as fire.

[0107] Identify the critical crossing points of refining anomaly signals and analyze them to generate a morphological sharpness factor. Specifically, this includes using the mean of all signal values ​​in the refining anomaly signals within the current sliding analysis window as a baseline, identifying the points where the signal value crosses the baseline twice consecutively, and recording these as critical crossing points.

[0108] Calculate the absolute difference between the maximum and minimum values ​​of the signal peaks between any two adjacent critical crossing points to obtain multiple peak-valley retreat depths; then calculate the absolute value of the slope between all continuous signal values ​​within the window, and calculate the standard deviation of the five largest slope values ​​to obtain the steepness dispersion.

[0109] Multiplying the average of all peak and valley retreat depths by the steepness dispersion, and then multiplying by the reciprocal of the total number of critical crossing points, yields the morphological sharpness factor of the window. The morphological sharpness factor is used to capture the morphology of abnormal fire signals, because abnormal signals caused by fire are usually sharp with rapid rise and violent fluctuations, while slight abnormal signals under dynamic conditions have a smooth morphology. This allows for the differentiation between normal operating condition anomalies and fire anomalies based on the signal morphology, further reducing the probability of misjudgment.

[0110] The evolution trend of refined anomalous signals over multiple consecutive time segments is tracked to generate trend coherence coefficients. Specifically, this involves: dividing the current 5-second sliding analysis window into three non-overlapping continuous sub-windows; calculating the linear regression slope of the signal values ​​in the refined anomalous signals within each of the three sub-windows to obtain three sub-trend values; examining the sign of the product of the first and second sub-trend values, as well as the sign of the product of the second and third sub-trend values; if the sign is positive, it indicates that the trend direction is continuous, and is recorded as 1; otherwise, it is recorded as 0. The two sign results are added together to obtain the direction coherence index.

[0111] Calculate the standard deviation of all signal values ​​within each sub-window, then calculate the coefficient of variation of these three standard deviations (i.e., standard deviation divided by mean), and use the reciprocal of the coefficient of variation as the fluctuation coherence factor.

[0112] The trend coherence coefficient of a window is obtained by multiplying the direction coherence index, the fluctuation coherence factor, and the absolute value of the mean signal values ​​of the first and third sub-windows. The trend coherence coefficient reflects the trend continuity of abnormal signals, avoiding misjudging short-term fluctuations in operating conditions as fires.

[0113] The abnormal energy flux, morphological sharpness factor, and trend coherence coefficient are fused to generate a one-dimensional burst intensity index representing the burst intensity of anomaly signals from a single sensor. Specifically, this involves: calculating the negative exponential function value of the abnormal energy flux with the natural constant as the base to obtain the energy attenuation basis; multiplying the morphological sharpness factor by 2 and then calculating its logarithm base 2 to obtain the waveform modulation gain; and adding the energy attenuation basis and the waveform modulation gain to obtain the preliminary fusion intensity.

[0114] Multiplying the trend coherence coefficient by the initial fusion strength yields a one-dimensional burst intensity index representing the burst intensity of the sensor's abnormal signal. The one-dimensional burst intensity index represents a score for the abnormal signal of a single sensor, improving the accuracy of a single sensor in identifying real fires.

[0115] By accurately distinguishing the energy differences between minor operational anomalies and severe fire anomalies through abnormal energy flux, the morphological sharpness factor captures the unique sharpness characteristics of fire from the signal morphology, the trend coherence coefficient avoids misjudgment of short-term operational fluctuations, and the one-dimensional sudden intensity index achieves accurate scoring of the suddenness of abnormal signals from a single sensor, improving the accuracy of a single sensor in identifying real fires, reducing the probability of false alarms under dynamic operating conditions, and strengthening the accurate identification capability of cabin fire monitoring.

[0116] In one embodiment, the single-dimensional burst intensity index corresponding to all sensors is collaboratively calculated with the global operating condition coefficient to generate a collaborative early warning threshold and a collaborative mutation energy, including:

[0117] The temporal relationship between the single-dimensional burst intensity indices of different sensors is analyzed, a heterogeneous signal network is constructed, and the network correlation entropy is generated by calculating the network correlation entropy. Specifically, each sensor is treated as an independent node. For any two different nodes, the cross-correlation coefficient between the single-dimensional burst intensity index sequences of the past 20 sampling periods is calculated. One of the single-dimensional burst intensity index sequences is lagged by 1 period and 2 periods respectively, and the cross-correlation coefficient is recalculated. The maximum value of the absolute values ​​of these three correlation coefficients is taken as the potential correlation strength between the two nodes.

[0118] The median of all potential association strengths is used as a threshold. If the potential association strength between two nodes exceeds the threshold, an undirected connection edge is established between them, and the potential association strength is used as the weight of this edge. The network composed of all nodes and edges is called a heterogeneous signal network.

[0119] Calculate the local clustering coefficient of each node in the heterogeneous signal network, which is the ratio of the number of edges of that node to the total number of edges; then calculate the standard deviation of the local clustering coefficients of all nodes, and take the reciprocal of the standard deviation as the network cohesion dispersion; multiply the average weight of all edges in the network by the network cohesion dispersion, and then perform a logarithmic operation to the base 2 on the product to obtain the network correlation entropy within the time window. The network correlation entropy is used to reflect the spatiotemporal correlation between abnormal signals from different sensors. Since real fires usually cause abnormalities in different sensors simultaneously or in a specific order, a high network correlation entropy indicates that the abnormalities of multiple sensors are synchronized, which is a strong collaborative early warning signal.

[0120] Based on the thermal stability of the engine room, the weights of various connection edges in the heterogeneous signal network are dynamically adjusted to generate the operating condition modulation conductance. Specifically, this includes: using the reciprocal of the engine room thermal stability as the thermal stability attenuation coefficient; for edges of nodes in the heterogeneous signal network, if both ends of an edge are sensors of the same type, such as temperature sensors, the basic sensitivity coefficient of that edge is 0.3; if both ends of an edge are sensors of different types, the basic sensitivity coefficient of that edge is 0.7.

[0121] For each edge in the heterogeneous signal network, its weight is multiplied by the corresponding basic sensitivity coefficient, and then multiplied by the thermal stability attenuation coefficient to obtain the weight of the edge after modulation under operating conditions.

[0122] The sum of the absolute values ​​of the differences between the weights after modulated operation and the original weights of all edges is calculated. This sum is then divided by the product of the total number of edges in the heterogeneous signal network and (1 + thermal stability attenuation coefficient) to obtain the modulated operation conductance. The modulated operation conductance is used to quantify the dynamic impact of changes in environmental thermal stability on the transmission efficiency of abnormal signals in the network.

[0123] Based on network correlation entropy and operating condition modulation transmissivity, anomalies of multiple sensors are analyzed to generate global instability. Specifically, this includes: obtaining the network correlation entropy value and operating condition modulation transmissivity value within the current 5-second time window, calculating the harmonic mean of the network correlation entropy value and operating condition modulation transmissivity value, and obtaining the collaborative baseline value.

[0124] The network correlation entropy sequence and the operating condition modulation transmissivity sequence were obtained for five consecutive 5-second time windows. The coefficient of variation of the network correlation entropy sequence was calculated as the entropy coefficient of variation, and the coefficient of variation of the operating condition modulation transmissivity sequence was calculated as the transmissivity coefficient of variation.

[0125] The arithmetic mean of the entropy coefficient of variation and the conductivity coefficient of variation is calculated to obtain the average coefficient of variation. Then, the cooperative baseline value is multiplied by (1 + average coefficient of variation) to obtain the global instability. The global instability is mainly used to assess the degree of anomaly of multiple sensors in the cabin.

[0126] By integrating global instability with cabin thermal inertia, the energy accumulation of fire in the cabin is analyzed, and an environmental gain factor that dynamically adjusts the early warning sensitivity is generated. Specifically, this includes: calculating the arithmetic square root of the cabin thermal inertia value to obtain the root value of thermal inertia; multiplying the global instability value by the root value of thermal inertia; and then multiplying the root value by pi (π) to obtain an intermediate product value, which represents the potential energy accumulation of fire in the cabin.

[0127] The absolute value of the tangent function of the intermediate product is calculated to obtain the environmental gain factor, which is used to dynamically adjust the trigger sensitivity of the warning threshold.

[0128] In one embodiment, the method of collaboratively calculating the single-dimensional burst intensity index corresponding to all sensors with the global operating condition coefficient to generate a collaborative early warning threshold and a collaborative mutation energy also includes:

[0129] The network correlation entropy, operating condition modulation transmissivity, global instability and environmental gain factor are fused to generate a collaborative early warning threshold. Specifically, the global instability value is multiplied by pi (π), its tangent function value is taken and added by 1 to obtain the modulation coefficient. The modulation coefficient is multiplied by the collaborative reference value to obtain the preliminary threshold after modulation.

[0130] The collaborative early warning threshold is obtained by multiplying the reciprocal of the environmental gain factor by the modulated initial threshold.

[0131] Based on the heterogeneous signal network, the one-dimensional burst intensity index and environmental gain factor of all sensors are analyzed to generate the cooperative mutation energy representing cabin anomalies. Specifically, this includes: for each node in the heterogeneous signal network, calculating the sum of the weights of all edges directly connected to it to obtain the local edge weight sum of that node; then finding the maximum value of the local edge weight sum of all nodes in the heterogeneous signal network, and dividing the local edge weight sum of each node by this maximum value to obtain the local network weight ratio of that node.

[0132] The network modulation node energy of a node is obtained by multiplying its one-dimensional burst intensity index, its local network weight ratio, and the arithmetic square root of the number of connected edges of the node.

[0133] The network modulation node energies of all nodes in the network are summed to obtain the total network modulation energy. Finally, the total network modulation energy is multiplied by the environmental gain factor to obtain the cooperative mutation energy. The cooperative mutation energy is used to reflect the anomaly intensity and correlation degree of multiple sensors, thereby determining the overall anomaly of the cabin.

[0134] In one embodiment, the collaborative early warning threshold and the collaborative mutation energy are compared to generate a fire early warning command, including:

[0135] When the collaborative mutation energy is greater than or equal to the collaborative early warning threshold and lasts for more than 3 seconds, a fire early warning command is automatically generated.

[0136] Otherwise, generate a security instruction.

[0137] By co-calculating the single-dimensional burst intensity index and the global operating condition coefficient, and combining the construction of a heterogeneous signal network with the fusion of multiple parameters to generate a collaborative early warning threshold and a collaborative mutation energy, the problem of frequent false alarms caused by the drift of the cabin operating conditions due to traditional fixed thresholds is solved. In addition, the network correlation entropy accurately captures the spatiotemporal correlation of anomalies of multiple sensors, the operating condition modulation conductance can quantify the dynamic impact of thermal stability on signal transmission, and the global instability is used to comprehensively evaluate the degree of anomalies of multiple sensors. The final output fire early warning command can improve the accuracy and reliability of fire early warning, effectively distinguish between normal operating condition fluctuations and real fires, and enhance the stability and adaptability of cabin fire monitoring.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor fusion automatic early warning system for marine engine room fire, characterized in that, include: The data analysis unit is used to acquire real-time sensor data from multiple sensors in the target warning area and the operating condition data of the host in the target warning area. It processes the real-time operating condition data to generate global operating condition coefficients, including: calculating the host speed and host output power in the real-time operating condition data, analyzing the host's heat work on the cabin, and generating instantaneous heat load values. The air volume in the cabin is acquired in real time. Based on the real-time air volume and instantaneous heat load value, the heat retention rate in the cabin is analyzed, and the cabin thermal inertia is generated. The target warning area is the ship's engine room, where the global operating condition coefficient is mainly used to reflect the overall thermal state and potential risks of the engine room; The baseline adjustment unit is used to acquire the static thresholds of each sensor within the target warning area, adjust the static thresholds according to the global operating condition coefficient, and generate a dynamic sensing baseline for each sensor, including: Based on the global operating condition coefficient, the influence of heat flow and air turbulence in the cabin is analyzed to generate disturbance entropy value; Based on the instantaneous heat load value and cabin thermal inertia, the impact load on sensor stability is analyzed, and its influence on the static threshold of each sensor is determined, generating the load transient gradient. By integrating the combustion oscillation index and the load transient gradient, field-induced coupling weights for different sensors are generated. The adaptive offset is obtained by fusing the perturbation entropy value, the load transient gradient, and the field-induced coupling weight. The adaptive offset is dynamically calibrated based on the static threshold of each sensor to generate a dynamic sensing baseline. The data correction unit is used to compare the sensor data of each sensor with the corresponding dynamic sensing baseline, eliminate data drift, and generate a refined abnormal signal for each sensor. The anomaly analysis unit is used to analyze the burst intensity of refined anomaly signals from each sensor and generate a one-dimensional burst intensity index, including: Analyze the refining anomaly signals to generate anomaly energy flux; Identify the critical crossing points of refined abnormal signals, analyze the critical crossing points, and generate morphological sharpness factors. Track the evolution trend of refined abnormal signals over multiple consecutive time segments and generate trend coherence coefficients; By fusing the abnormal energy flux, morphological sharpness factor and trend coherence coefficient, a one-dimensional burst intensity index representing the burst intensity of anomalous signals from a single sensor is generated. The collaborative computing unit is used to collaboratively calculate the single-dimensional burst intensity index corresponding to all sensors and the global operating condition coefficient to generate collaborative early warning thresholds and collaborative mutation energy. The collaborative mutation energy is used to reflect the abnormal intensity and correlation of multiple sensors, thereby determining the overall abnormality of the cabin. The fire early warning unit is used to compare the collaborative early warning threshold and the collaborative mutation energy to generate a fire early warning command.

2. The marine engine room fire multi-sensor fusion automatic warning system according to claim 1, characterized in that, The process of processing real-time operating data to generate global operating coefficients also includes: Analyze the instantaneous fluctuations of the main unit's exhaust temperature in real-time operating data to generate a combustion oscillation index; The cabin thermal stability is obtained by fusing the cabin thermal inertia and the combustion oscillation index. The instantaneous heat load value and the thermal stability of the engine room are calculated together to generate a global operating condition coefficient.

3. The marine engine room fire multi-sensor fusion automatic warning system according to claim 1, characterized in that, For each sensor, its sensor data is compared with the corresponding dynamic sensing baseline, and data drift is eliminated to generate a refined anomaly signal for each sensor, including: For each sensor, calculate the real-time difference sequence between its real-time data and the dynamic sensing baseline, analyze the statistical boundary of the real-time difference sequence in the time domain based on the field-induced coupling weight, and generate anomaly judgment confidence interval; Analyze abrupt changes in the real-time difference sequence outside the anomaly detection confidence interval and generate event significance values; The confidence interval for anomaly determination is calculated in conjunction with the event significance value to obtain a refined anomaly signal after removing drift and impulse interference.

4. The marine engine room fire multi-sensor fusion automatic warning system according to claim 2, characterized in that, The single-dimensional burst intensity index corresponding to all sensors is collaboratively calculated with the global operating condition coefficient to generate a collaborative early warning threshold and a collaborative mutation energy, including: The temporal relationship between the one-dimensional burst intensity indices of different sensors is analyzed, a heterogeneous signal network is constructed, and the network correlation entropy is generated by calculating the heterogeneous signal network. Based on the thermal stability of the engine room, the weights of various connection edges in the heterogeneous signal network are dynamically adjusted to generate the operating condition modulation conductance. Based on network correlation entropy and operating condition modulation transmissivity, anomalies of multiple sensors are analyzed to generate global instability. By integrating global instability with cabin thermal inertia, the energy accumulation of fire in the cabin is analyzed, and an environmental gain factor that dynamically adjusts the early warning sensitivity is generated.

5. The marine engine room fire multi-sensor fusion automatic warning system according to claim 4, characterized in that, The method involves collaboratively calculating the single-dimensional burst intensity index corresponding to all sensors with the global operating condition coefficient to generate a collaborative early warning threshold and collaborative mutation energy. This also includes: The network correlation entropy, operating condition modulation transmissivity, global instability and environmental gain factor are fused to generate a collaborative early warning threshold; Based on the heterogeneous signal network, the one-dimensional burst intensity index and environmental gain factor of all sensors are analyzed to generate a cooperative mutation energy representing cabin anomalies.

6. The marine engine room fire multi-sensor fusion automatic warning system according to claim 5, characterized in that, By comparing the collaborative early warning threshold and the collaborative mutation energy, a fire early warning instruction is generated, including: When the collaborative mutation energy is greater than or equal to the collaborative early warning threshold and lasts for more than 3 seconds, a fire early warning command is automatically generated. Otherwise, generate a security instruction.

Images