Intelligent fire extinguishing system based on cloud collaboration

Abstract

The invention discloses an intelligent fire extinguishing system based on cloud collaboration, and particularly relates to the field of fire fighting, and the system comprises a prediction module, an optimization module, a difference module, a synthesis module, a verification module and an updating module. The prediction module is used for collecting fire data and inputting the fire data to the edge side deep learning model, and outputting a first prediction result and a corresponding first instant instruction; the optimization module is used for transmitting the fire data to a cloud side deep learning model, and generating a second prediction result and a corresponding second optimization instruction. A real-time instruction and an optimization instruction are generated on an edge side and a cloud side respectively, phase collision is executed on a time axis to form a difference track, a synthesis instruction is generated based on the difference track, and verification and correction are performed in combination with feedback data, so that the consistency and coordination of edge quick response and cloud global scheduling in multiple rounds of iteration are realized, and the efficiency of edge quick response and cloud global scheduling is improved. Therefore, the problem of disposition disorder caused by instruction inconsistency, execution conflict and communication delay in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of fire protection technology, and more specifically, to a cloud-based intelligent fire extinguishing system. Background Technology

[0002] In existing cloud-based collaborative intelligent fire suppression systems, there are significant hidden dangers in the command transmission and execution process. The system relies on the rapid response of edge devices on the one hand, and on the global scheduling of the cloud on the other hand, and the two often cannot maintain complete consistency in actual operation. When a fire breaks out, edge devices usually take immediate action based on real-time data from local sensors, while the cloud needs to integrate information from multiple sources and perform calculations before issuing a decision. As a result, the edge device may have already taken measures, but subsequent instructions from the cloud may require adjustments or even overturn the existing actions, causing instruction conflicts and confusion in handling. If the edge device chooses to wait for feedback from the cloud, it may miss the critical opportunity to extinguish the fire. Therefore, existing technologies lack a processing mechanism that can maintain stability and coordination even when there is communication delay, link interruption, or inconsistent instructions; secondly, this deficiency will also affect the reliability and effectiveness of fire extinguishing systems in emergency situations. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a cloud-based collaborative intelligent fire suppression system. This system generates real-time and optimized commands at the edge and cloud sides respectively, performs phase collision on the time axis to form differential trajectories, and then generates synthetic commands based on the differential trajectories and verifies and corrects them in conjunction with feedback data. This achieves consistency and coordination between rapid edge response and global cloud scheduling in multiple iterations, thereby solving the problems of inconsistent commands, execution conflicts, and communication delays that lead to chaotic handling in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a cloud-based collaborative intelligent fire extinguishing system, comprising a prediction module, an optimization module, a difference module, a synthesis module, a verification module, and an update module; The prediction module is used to collect fire data and input it into the edge-side deep learning model, and output the first prediction result and the corresponding first real-time command. The optimization module is used to transmit the fire data to the cloud-side deep learning model to generate a second prediction result and a corresponding second optimization instruction; The difference module is used to perform phase collision between the first immediate instruction and the second optimized instruction on the time axis after the instruction is generated, forming a difference trajectory. The difference trajectory represents the phase convergence state or phase deviation state of the edge side and the cloud side in the execution direction. The synthesis module generates a third synthesis instruction based on the difference trajectory. The third synthesis instruction is formed by superimposing the parameter weights of the first immediate instruction and the second optimized instruction. The verification module is used to execute the third synthesis instruction and collect fire extinguishing feedback data in real time during the execution process. The fire extinguishing feedback data is used to form a feedback curve, and the feedback curve is compared with the difference trajectory to identify the execution deviation of the third synthesis instruction. The update module is used to generate a fourth correction instruction when the comparison result of the verification module shows that the feedback curve and the difference trajectory have a continuous deviation. The fourth correction instruction and the difference trajectory are then uploaded to the cloud side to update the corresponding model parameters on the edge side and the cloud side for the next round of phase collision iteration.

[0005] In a preferred embodiment, the construction and training of the edge-side deep learning model includes: slicing the collected fire data on the edge side according to a fixed time window, and retaining only the portions of temperature, smoke, voltage and current changes exceeding a preset threshold within each time window to form a first sparse input sequence; When any type of data is missing in the first sparse input sequence, a second compensation input sequence is generated based on the gradient inference results of adjacent time windows and the comparison results of similar sensors, forming a continuous input to ensure that the inference link is not interrupted. The second compensation input sequence is input to the hierarchical memory computing unit, which forms a forgetting update vector in the first layer, a delayed cumulative memory vector in the second layer, and merges them to generate a third time-series representation sequence. The third temporal representation sequence is input into the sparse attention computing unit, which assigns computing weights only to the data types that trigger the anomaly, forming the fourth weighted representation sequence. The fourth weighted representation sequence is input to the edge-side classification calculation layer. The edge-side classification calculation layer receives the fourth weighted representation sequence in each time window and performs matrix multiplication and weighted summation operations to form a classification score vector. After forming the classification score vector, a normalization operation is performed on the classification score vector, the index value of each category is calculated, and the ratio of each index value to the sum of all index values ​​is calculated to output the fire occurrence probability distribution. After outputting the fire occurrence probability distribution, the upper limit of the probability values ​​in the fire occurrence probability distribution is identified, and a first prediction result is obtained. The first prediction result includes the category label and the corresponding probability value. The first prediction result and the fire occurrence probability distribution are used to generate a first instant command. When there is a difference between the first instant command and the optimization conditions returned from the cloud, the weight allocation rules of the hierarchical memory computing unit and the sparse attention computing unit are updated to achieve adaptive evolution of the edge-side model.

[0006] In a preferred embodiment, the construction and training of the cloud-side deep learning model includes: receiving fire data uploaded from the edge side and aligning it with historical fire data stored in the cloud according to the timestamp order to form a first full data sequence, wherein the first full data sequence includes temperature data, smoke data, voltage data and current data; In the first full data sequence, the mean, variance and trend slope of temperature data, smoke data, voltage data and current data are calculated within a preset number of historical time windows to generate a second statistical result. The second statistical results are compared one by one among multiple devices to calculate the temperature difference, smoke difference, voltage difference and current difference between different devices, forming the third comparison result; The third comparison results are aggregated according to a preset time window, and the rate of change of temperature difference, smoke difference, voltage difference and current difference are calculated between adjacent time windows. The rate of change is obtained by dividing the numerical difference between adjacent time windows by the time interval, thereby forming a continuous change gradient. After obtaining the continuous change gradient, the continuous change gradients of each data are combined in chronological order to form a fourth trend sequence. The fourth trend sequence is input into the classification calculation unit in the cloud. In the classification calculation unit, each type of data in the fourth trend sequence is multiplied by its corresponding weight value and summed to form a stage score. After the stage score is formed, the index value of each stage score is calculated in sequence, and each index value is divided by the sum of all index values ​​to output the probability distribution of the fire development stage. After obtaining the probability distribution of the fire development stage, the maximum value of the probability values ​​in the probability distribution is selected, and the fire stage corresponding to the maximum value is determined as the second prediction result. The second prediction result includes the category of the fire stage and the probability value corresponding to the maximum value. Based on the second prediction result and the first real-time command uploaded from the edge side, the probability value of fire occurrence is compared with the corresponding control parameters. When the comparison result shows that there is a difference in the probability value of fire occurrence or the control parameter categories are inconsistent, a second optimization command is generated. The second optimization command includes threshold parameters for correcting the fire triggering conditions determined by the edge side, and control parameters for adjusting the spray volume, spray time and spray direction of the fire extinguishing action on the edge side. The second optimization command is then sent back to the edge side.

[0007] In a preferred embodiment, the phase collision process of the difference module includes: after the first immediate instruction and the second optimized instruction are generated, the trigger time, trigger condition and control parameters of the first immediate instruction and the trigger time, trigger condition and control parameters of the second optimized instruction are extracted respectively to form a first instruction sequence and a second instruction sequence. Align the trigger times of the first instruction sequence with the trigger times of the second instruction sequence, and at each aligned time position, calculate the numerical difference between the control parameters in the first instruction sequence and the control parameters in the second instruction sequence to generate a control parameter difference sequence, and record the sign direction of each numerical difference in the control parameter difference sequence. In the control parameter difference sequence, the rate of change of control parameter difference at adjacent time positions is calculated sequentially, and the phase convergence state or phase deviation state of the first instantaneous instruction and the second optimization instruction at that time position is marked according to the positive and negative values ​​and magnitude of the rate of change of control parameter difference, thereby generating a convergence deviation marking sequence.

[0008] In a preferred embodiment, the phase collision result acquisition of the difference module includes: combining the control parameter difference sequence with the convergence deviation marker sequence to form a difference trajectory. The difference trajectory includes the control parameter difference, the rate of change of the control parameter difference, and the phase convergence state or phase deviation state at each time position. The difference trajectory characterizes the phase collision result of the first instantaneous instruction and the second optimized instruction in the execution direction.

[0009] In a preferred embodiment, the generation process of the third synthesis instruction in the synthesis module includes: Based on the control parameter differences and control parameter difference change rates recorded at each time position in the difference trajectory, the weight coefficients of the first instantaneous instruction and the second optimized instruction are calculated. The control parameters of the first instant instruction are multiplied by the corresponding weight coefficient, the control parameters of the second optimization instruction are multiplied by the corresponding weight coefficient, and the results at the same time position are summed to form a composite parameter sequence. The synthesis parameter sequence is combined in chronological order, and a third synthesis instruction is generated after the combination is completed. The third synthesis instruction includes control parameters after weight superposition.

[0010] In a preferred embodiment, the execution process of the verification module includes: The third synthesis command is executed, and fire extinguishing feedback data is collected in real time during the execution process. The fire extinguishing feedback data includes temperature change data, smoke concentration data, voltage change data, and current change data, and forms a feedback data sequence. The feedback data sequence is arranged in chronological order. At each adjacent time position, the temperature change difference, smoke concentration difference, voltage change difference, and current change difference are calculated, and the ratio of each difference to the time interval is calculated to form a feedback curve. The feedback curve and the difference trajectory are compared item by item at the same time position. During the comparison process, the difference and rate of change of the four types of data, namely temperature, smoke, voltage and current, are calculated respectively, and a comparison result sequence is formed. In the comparison result sequence, if the difference between any type of data such as temperature, smoke, voltage or current exceeds a preset threshold within a preset number of time windows, or if the difference in the rate of change is positive or negative within a preset number of time windows, then it is determined that there is an execution deviation in the third synthesis instruction.

[0011] In a preferred embodiment, the execution process of the update module includes: After the verification module completes the comparison, it statistically analyzes the difference and rate of change between the feedback curve and the difference trajectory in each time window. When the difference exceeds the preset threshold in a preset number of time windows, or the rate of change has the same sign in a preset number of time windows, it is determined that there is a continuous deviation between the feedback curve and the difference trajectory. After determining that there is a continuous deviation, a fourth correction instruction is generated based on the magnitude of the difference and the difference in the rate of change. The fourth correction instruction includes correction parameters for the forgetting factor and the accumulation factor on the edge side, as well as correction values ​​for the classification threshold and weight parameters on the cloud side. The fourth correction instruction and the difference trajectory are uploaded to the cloud side, and the corresponding model parameters are updated on the cloud side and the edge side respectively. The updated model parameters are stored as the initial input conditions for the next round of phase collision. After the parameter update is completed, the updated model parameters are used to re-drive the computation process of the edge-side deep learning model and the cloud-side deep learning model for the generation of the first immediate instruction and the second optimization instruction in the next round, and are iteratively executed in phase collision.

[0012] The technical effects and advantages of this invention are as follows: 1. In this solution, the problem of inconsistency between the edge side and the cloud side in the instruction execution stage is solved. By introducing a phase collision mechanism through the difference module, the system compares the first real-time instruction and the second optimized instruction on the time axis at the same time, generates the difference trajectory and quantifies the convergence or deviation state, so that the consistency and coordination of the instructions on both sides can be maintained even under the condition of communication delay or link instability, avoiding instruction conflict and execution chaos. 2. In addition, this scheme combines hierarchical memory computing units and sparse attention computing units at the edge, which can hierarchically represent short-term fire mutation information and long-term fire evolution trends; the fast forgetting update vector captures second-level changes, the delayed cumulative memory vector accumulates multi-window trends, and then the sparse attention computing focuses on abnormal data types, so that the edge prediction results can respond in a timely manner and avoid misjudgment caused by fluctuations in a single feature. 3. This solution utilizes the timestamp alignment and cross-device comparison of all data on the cloud side to form trend information and stage evolution characteristics of fire propagation; by calculating the statistical results and change rates of temperature, smoke, voltage and current, a fourth trend sequence is generated and the probability distribution of fire stages is output, so that cloud prediction can reflect the whole picture of fire development and provide global optimization conditions for edge side instructions. 4. In the synthesis stage, this solution dynamically allocates the weights of the first immediate command and the second optimized command through the difference trajectory, generates the third synthesized command and directly drives the fire extinguishing device. The weight allocation is based on the difference and the rate of change, ensuring that when the edge prediction is closer to the real situation, it is given a higher weight, and when the cloud prediction reflects the long-term trend, its role is enhanced, so that the final fire extinguishing action takes into account both timeliness and globality. 5. In the verification and update stage, this solution collects fire extinguishing feedback data and forms a feedback curve, which is compared with the difference trajectory item by item to identify execution deviations. When a deviation is found within a local window, it is determined to be an abnormal command execution. When it shows a continuous deviation at the statistical level, a fourth correction command is generated. This fourth correction command adjusts the forgetting factor and accumulation factor on the edge side, as well as the classification threshold and weight parameters on the cloud side, so as to achieve adaptive evolution of the model in the next iteration and ensure the convergence and stability of the system in multiple rounds of operation. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0014] 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.

[0015] Refer to the instruction manual appendix Figure 1 An embodiment of the present invention provides a cloud-based collaborative intelligent fire extinguishing system, comprising a prediction module, an optimization module, a difference module, a synthesis module, a verification module, and an update module. The prediction module is used to collect fire data and input it into the edge-side deep learning model, and output the first prediction result and the corresponding first real-time command. The optimization module is used to transmit the fire data to the cloud-side deep learning model to generate a second prediction result and a corresponding second optimization instruction; The difference module is used to perform phase collision between the first immediate instruction and the second optimized instruction on the time axis after the instruction is generated, forming a difference trajectory. The difference trajectory represents the phase convergence state or phase deviation state of the edge side and the cloud side in the execution direction. The synthesis module generates a third synthesis instruction based on the difference trajectory. The third synthesis instruction is formed by superimposing the parameter weights of the first immediate instruction and the second optimized instruction. The verification module is used to execute the third synthesis instruction and collect fire extinguishing feedback data in real time during the execution process. The fire extinguishing feedback data is used to form a feedback curve, and the feedback curve is compared with the difference trajectory to identify the execution deviation of the third synthesis instruction. The update module is used to generate a fourth correction instruction when the comparison result of the verification module shows that the feedback curve and the difference trajectory have a continuous deviation. The fourth correction instruction and the difference trajectory are then uploaded to the cloud side to update the corresponding model parameters on the edge side and the cloud side for the next round of phase collision iteration.

[0016] The construction and training of the edge-side deep learning model includes: slicing the collected fire data at the edge according to fixed time windows, and retaining only the portions of temperature, smoke, voltage, and current changes exceeding preset thresholds within each time window to form a first sparse input sequence; where edge-side refers to local computing and execution devices deployed at or near the fire source, including embedded processors, industrial controllers, or edge computing nodes, used to directly process sensor-collected data and generate real-time instructions locally, rather than relying on remote cloud servers; When any type of data is missing in the first sparse input sequence, a second compensation input sequence is generated based on the gradient inference results of adjacent time windows and the comparison results of similar sensors, forming a continuous input to ensure that the inference link is not interrupted. It should be noted that when temperature, smoke, voltage, or current data is missing within a certain time window on the edge side, the system retrieves the collected values ​​of the corresponding data in the previous and next time windows, calculates the numerical differences between adjacent time windows, and forms a continuously changing gradient. After completing the gradient calculation, the system extends the gradient to the time position of the missing point and solves for the estimated value of the missing point. The estimated value is the gradient extrapolation result. Specifically, after obtaining the continuously changing gradient, the system multiplies the gradient with the time position of the missing point and adds the end value of the previous time window to solve for the estimated value of the missing point. It should be noted that when temperature, smoke, voltage, or current data is missing within a certain time window on the edge side, the system retrieves the values ​​from sensors of the same type as the missing data within the same time window. This value is then compared to the historical values ​​of the missing data in the previous time window, and a stable interval for the difference is calculated across multiple time windows. After completing the difference calculation, the system compensates the values ​​within the stable interval to the time position of the missing point and solves for the compensation value at that point. This compensation value is the comparison result with similar sensors. Specifically, after obtaining the stable interval of the difference, the system calculates the weighted sum of the current time window value of similar sensors and the average difference within the stable interval, and combines this weighted sum with the historical baseline value of the missing feature to solve for the compensation value at that point. The second compensation input sequence is input to the hierarchical memory computing unit. The hierarchical memory computing unit forms a fast forgetting update vector in the first layer, a delayed cumulative memory vector in the second layer, and merges them to generate a third temporal representation sequence. It should be noted that the hierarchical memory computing unit is a multi-layer temporal processing structure deployed on the edge side. Its first layer is used to process the fast response of short-term fire features, the second layer is used to process the trend accumulation of long-term fire features, and continuous temporal representation results are formed through inter-layer merging. The fast forgetting update vector is a numerical sequence generated by the first layer of the hierarchical memory computing unit within each time window. This numerical sequence applies a high forgetting factor to the input features, retaining only short-term mutation information of the fire data to drive a second-level response. The high forgetting factor refers to assigning a near-zero decay weight to historical inputs in memory computing. The value is usually located in the lower limit of the weight interval, including close to 0.1, to quickly discard past data and highlight the data changes in the current time window. The delayed cumulative memory vector is a numerical sequence generated by the second layer of the hierarchical memory computing unit after calculation across time windows. This numerical sequence gradually accumulates the changing trends of multiple time windows by applying a low forgetting factor to the input features, and is used to identify the continuous evolution pattern of fire. The low forgetting factor refers to assigning a retention weight close to one to the historical input in the memory computing. The value is usually located in the upper limit of the weight interval, including close to 0.9, to retain data from multiple time windows and gradually accumulate to form a long-term changing trend. The third temporal representation sequence is input to the sparse attention computing unit, which assigns computational weights only to the data types that trigger anomalies, forming a fourth weighted representation sequence. The sparse attention computing unit is a computational structure based on attention weight allocation. Its input is the temporal representation sequence. During execution, it first calculates the correlation coefficient between each data dimension, then assigns non-zero computational weights to the data dimensions whose correlation coefficients exceed a preset threshold or trigger anomaly conditions, and assigns zero weights to other data dimensions that have not been triggered, thereby forming a sparse matrix containing only some effective weights. Finally, the sparse matrix is ​​used to perform a weighted summation on the input temporal representation sequence, and the weighted representation sequence is output. The fourth weighted representation sequence is input to the edge-side classification calculation layer. The edge-side classification calculation layer receives the fourth weighted representation sequence in each time window and performs matrix multiplication and weighted summation operations to form a classification score vector. After forming the classification score vector, a normalization operation is performed on the classification score vector, the index value of each category is calculated, and the ratio of each index value to the sum of all index values ​​is calculated to output the fire occurrence probability distribution. After outputting the fire occurrence probability distribution, the upper limit of the probability values ​​in the fire occurrence probability distribution is identified, and a first prediction result is obtained. The first prediction result includes a category label and the corresponding probability value, used to characterize the probability of a fire event occurring within the current time window. The system generates a first real-time command by combining the first prediction result with the fire occurrence probability distribution. When there is a difference between the first real-time command and the optimization conditions returned from the cloud, the system updates the weight allocation rules of the hierarchical memory computing unit and the sparse attention computing unit to achieve adaptive evolution of the edge-side model. The system compares the trigger parameters of the first real-time command with the control parameters of the optimization conditions returned from the cloud one by one at the edge. When the numerical difference exceeds a preset threshold or the types of actions executed are inconsistent, it can be determined that there is a difference. After the system completes the comparison, it marks the difference results as an update signal, which is used to trigger the adjustment of the weight allocation rules of the hierarchical memory computing unit and the sparse attention computing unit, thereby driving the edge-side model into the adaptive evolution process. Secondly, when the system detects a difference between the first real-time instruction and the optimization conditions returned from the cloud, it directly replaces the original values ​​of the forgetting factor and the accumulation factor in the hierarchical memory computing unit with the difference value, and writes the abnormal category corresponding to the difference into the weight matrix of the sparse attention computing unit. In the matrix, the weight value corresponding to the abnormal category is increased, and the weight value corresponding to the indifferent category is decreased, generating a new weight allocation rule and overwriting the original rule.

[0017] The construction and training of the cloud-side deep learning model includes: receiving fire data uploaded from the edge side and aligning it with historical fire data stored in the cloud according to the timestamp order to form a first full data sequence, the first full data sequence including temperature data, smoke data, voltage data and current data; In the first full data sequence, the mean, variance and trend slope of temperature data, smoke data, voltage data and current data are calculated within a preset number of historical time windows to generate a second statistical result, which can be understood as the second statistical result being used to characterize the long-term changes in the fire development process. The second statistical result is compared one by one among multiple devices to calculate the temperature difference, smoke difference, voltage difference, and current difference between different devices, forming a third comparison result. The third comparison result is used to identify the propagation relationship of the fire among different devices. Here, multiple devices refer to two or more non-overlapping monitoring units that independently collect data. These monitoring units generate temperature, smoke, voltage, and current data respectively. The cloud will compare their outputs one by one to determine whether the fire has spread or propagated between different physical units. The third comparison results are aggregated according to a preset time window, and the rate of change of temperature difference, smoke difference, voltage difference and current difference are calculated between adjacent time windows. The rate of change is obtained by dividing the numerical difference between adjacent time windows by the time interval, thereby forming a continuous change gradient. After obtaining the continuous change gradient, the continuous change gradients of each data are combined in chronological order to form a fourth trend sequence. The fourth trend sequence is used to identify the stage evolution of the fire. The fourth trend sequence is input into the classification calculation unit in the cloud. In the classification calculation unit, each type of data in the fourth trend sequence is multiplied by its corresponding weight value and summed to form a stage score. After the stage score is formed, the index value of each stage score is calculated in sequence, and each index value is divided by the sum of all index values ​​to output the probability distribution of the fire development stage. After obtaining the probability distribution of the fire development stage, the maximum value of the probability values ​​in the probability distribution is selected, and the fire stage corresponding to the maximum value is determined as the second prediction result. The second prediction result includes the category of the fire stage and the probability value corresponding to the maximum value. Based on the second prediction result and the first real-time command uploaded from the edge side, the probability value of fire occurrence is compared with the corresponding control parameters. When the comparison result shows that there is a difference in the probability value of fire occurrence or the control parameter categories are inconsistent, a second optimization command is generated. The second optimization command includes threshold parameters for correcting the fire triggering conditions determined by the edge side, and control parameters for adjusting the spray volume, spray time and spray direction of the fire extinguishing action on the edge side. The second optimization command is then sent back to the edge side to directly update the triggering conditions and execution content of the first real-time command. The corresponding control parameters refer to the numerical parameters in the first real-time command used to determine the specific execution method of the fire extinguishing action, including but not limited to the spray volume, spray time, spray direction and start / stop status of the spraying device.

[0018] The phase collision process of the difference module includes: after the first immediate instruction and the second optimized instruction are generated, the trigger time, trigger condition and control parameters of the first immediate instruction and the trigger time, trigger condition and control parameters of the second optimized instruction are extracted respectively to form a first instruction sequence and a second instruction sequence. Align the trigger times of the first instruction sequence with the trigger times of the second instruction sequence, and at each aligned time position, calculate the numerical difference between the control parameters in the first instruction sequence and the control parameters in the second instruction sequence to generate a control parameter difference sequence, and record the sign direction of each numerical difference in the control parameter difference sequence; wherein the sign direction of each numerical difference refers to the sign of the difference between the control parameters of the first immediate instruction and the control parameters of the second optimized instruction at the same time position, with a positive sign indicating that the control parameters of the first immediate instruction are greater than the control parameters of the second optimized instruction, and a negative sign indicating that the control parameters of the first immediate instruction are less than the control parameters of the second optimized instruction; In the control parameter difference sequence, the rate of change of control parameter difference at adjacent time positions is calculated sequentially, and the phase convergence state or phase deviation state of the first instantaneous instruction and the second optimized instruction at that time position is marked according to the positive and negative values ​​and magnitude of the rate of change of control parameter difference, thereby generating a convergence deviation marking sequence. Phase convergence state refers to the rate of change of the difference in control parameters gradually approaching zero at adjacent time positions, or the sign direction of the difference in control parameters gradually changing from inconsistent to consistent, indicating that the control parameters of the first immediate instruction and the second optimized instruction are gradually approaching each other on the time axis. Phase deviation state: refers to the situation where the rate of change of the difference in control parameters continues to increase at adjacent time positions, or the signs of the difference in control parameters remain inconsistent, indicating that the control parameters of the first immediate instruction and the second optimized instruction are gradually moving away from each other on the time axis.

[0019] The acquisition of the phase collision result of the difference module includes: combining the control parameter difference sequence with the convergence deviation marker sequence to form a difference trajectory. The difference trajectory includes the control parameter difference, the rate of change of the control parameter difference, and the phase convergence state or phase deviation state at each time position. The difference trajectory represents the phase collision result of the first instantaneous instruction and the second optimized instruction in the execution direction.

[0020] In the synthesis module, the generation process of the third synthesis instruction includes: Based on the control parameter differences and the rate of change of control parameter differences recorded at each time position in the difference trajectory, the weight coefficients of the first immediate instruction and the second optimized instruction are calculated. The weight coefficient calculation here includes: using the absolute value of the control parameter difference and the rate of change of the control parameter difference at each time position as the allocation basis; increasing the weight coefficient of the first immediate instruction when the control parameter difference is small and the rate of change approaches zero; and increasing the weight coefficient of the second optimized instruction when the control parameter difference is large and the rate of change continues to increase. Furthermore, the value of the weight coefficient can be determined by normalization, that is, limiting the weight values ​​of the first immediate instruction and the second optimized instruction to between 0 and 1, and ensuring that the sum of their weight values ​​at the same time position equals 1. The control parameters of the first instant instruction are multiplied by the corresponding weight coefficient, the control parameters of the second optimization instruction are multiplied by the corresponding weight coefficient, and the results at the same time position are summed to form a composite parameter sequence. The synthetic parameter sequence is combined in chronological order, and a third synthetic instruction is generated after the combination is completed. The third synthetic instruction includes control parameters after weight superposition, which are used for subsequent fire extinguishing actions. The execution of the fire extinguishing action refers to the process of directly driving the fire extinguishing device to complete the operation of the spray volume, spray time and spray direction of the fire extinguishing agent according to the control parameters after weight superposition in the third synthetic instruction.

[0021] The execution process of the verification module includes: The third synthesis command is executed, and fire extinguishing feedback data is collected in real time during the execution process. The fire extinguishing feedback data includes temperature change data, smoke concentration data, voltage change data, and current change data, and forms a feedback data sequence. Temperature change data is used to reflect the dynamic changes in fire source intensity and combustion rate during the fire extinguishing process. It corresponds to the temperature data obtained in the fire data acquisition stage. It is the feedback of this type of temperature data after the third synthetic command drives the fire extinguishing device to execute, and is used to verify whether the temperature suppression has reached the command expectation. Smoke concentration data is used to reflect the release of combustion products during the fire extinguishing process. It corresponds to the smoke data acquired during the fire data acquisition phase. It is the feedback formed by this type of smoke data during the execution of the third synthesis instruction and is used to determine whether the fire extinguishing action effectively reduces visible smoke and particulate matter. Voltage change data is used to reflect the stability of the electrical system during the fire extinguishing process. It corresponds to the voltage data obtained during the fire data acquisition phase. It is the feedback formed after the third synthesis command triggers the electrical-related control and is used to monitor whether the fire extinguishing action causes abnormal voltage fluctuations. Current change data is used to reflect the safety status of electrical loads during the fire extinguishing process. It corresponds to the current data obtained during the fire data acquisition phase. It is the feedback formed by this type of current data during the execution of the third synthesis command and is used to verify whether the fire extinguishing action avoids the risk of current overload or short circuit. The feedback data sequence is arranged in chronological order. At each adjacent time position, the temperature change difference, smoke concentration difference, voltage change difference, and current change difference are calculated, and the ratio of each difference to the time interval is calculated to form a feedback curve. The feedback curve and the difference trajectory are compared item by item at the same time position. During the comparison, the difference and rate of change of four types of data (temperature, smoke, voltage, and current) are calculated respectively, and a comparison result sequence is formed. The difference refers to the numerical result obtained by subtracting the difference of control parameters in the feedback curve from the difference of control parameters in the difference trajectory at the same time position, and is used to measure the degree of deviation between the two curves in terms of numerical magnitude. The rate of change difference refers to the numerical result obtained by subtracting the rate of change of the difference of control parameters in the feedback curve from the rate of change of the difference of control parameters in the difference trajectory at the same time position, and is used to measure the degree of deviation between the two curves in terms of trend of change. In the comparison result sequence, if the difference between any type of data such as temperature, smoke, voltage or current exceeds a preset threshold within a preset number of time windows, or if the signs of the rate of change differences are all positive or all negative within a preset number of time windows, then it is determined that there is an execution deviation in the third synthesis instruction; where the signs are all positive or all negative means that within a preset number of time windows, the values ​​of all rate of change differences are greater than zero or all rate of change differences are less than zero, which is used to indicate that the change trend is consistent in the same direction.

[0022] The execution process of the update module includes: After the verification module completes the comparison, it statistically analyzes the difference and rate of change between the feedback curve and the difference trajectory in each time window. When the difference exceeds a preset threshold within a preset number of time windows, or the rate of change is consistent in sign within a preset number of time windows, it is determined that there is a continuous deviation between the feedback curve and the difference trajectory. It should be noted that the judgment of continuous deviation is different from the execution deviation of the verification module. The continuous deviation is statistically judged on the overall trend of the entire feedback curve and the difference trajectory to identify long-term inconsistencies at the system level, while the execution deviation is only judged on the abnormal performance of a single type of data such as temperature, smoke, voltage or current within a local time window. After determining that there is a continuous deviation, a fourth correction instruction is generated based on the magnitude of the difference and the difference in the rate of change. The fourth correction instruction includes correction parameters for the forgetting factor and the accumulation factor on the edge side, as well as correction values ​​for the classification threshold and weight parameters on the cloud side. The fourth correction instruction and the difference trajectory are uploaded to the cloud side, and the corresponding model parameters are updated on the cloud side and the edge side respectively. The updated model parameters are stored as the initial input conditions for the next round of phase collision. After the parameter update is completed, the updated model parameters are used to re-drive the computation process of the edge-side deep learning model and the cloud-side deep learning model for the generation of the first immediate instruction and the second optimization instruction in the next round, and are iteratively executed in phase collision. It should be noted that the forgetting factor on the edge side refers to the parameter used to decay historical data in the hierarchical memory computation on the edge side, which determines the proportion of historical data retained in the current computation. Modifying this factor can change the balance between short-term and long-term data on the edge side. When a bias occurs, the dependence on historical input can be increased or decreased by adjusting the forgetting factor. The accumulation factor on the edge side refers to the parameter used to gradually accumulate data from multiple time windows in the edge-side hierarchical memory calculation. It is used to control the accumulation speed of long-term trends. Modifying this accumulation factor is used to adjust the sensitivity of the edge side to trend information and avoid distortion of real-time judgment due to accumulation that is too fast or too slow. Cloud-side classification threshold: refers to the boundary value used to determine the fire stage category in cloud-side classification calculation. When the probability distribution exceeds this value, it is determined to enter a certain stage. Adjusting the classification threshold can change the sensitivity of the cloud in determining the fire stage, making the cloud judgment result more in line with the actual feedback. The weighting parameter on the cloud side refers to the weight values ​​assigned to different features (temperature, smoke, voltage, current) when calculating the probability distribution of fire stages on the cloud side. Adjusting this parameter can change the influence ratio of various features in the classification results, and can be used to compensate for the excessive bias of a single feature.

[0023] The following is an overall explanation of this scheme: This scheme is based on a prediction module, an optimization module, a difference module, a synthesis module, a verification module, and an update module. It uses the core mechanism of phase collision to connect and correct the calculation results from the edge side and the cloud side. Specifically, the prediction module first collects fire data such as temperature, smoke, voltage and current at the edge. The data is sliced ​​and filtered within a fixed time window to remove invalid data below the threshold, resulting in the first sparse input sequence. When data is missing, the system uses the values ​​of adjacent time windows to calculate the gradient change and estimate the value of the missing point. At the same time, it combines the difference compensation results of similar sensors to form a continuous input, ensuring the integrity of the data link. Based on this, the data sequentially enters the hierarchical memory computing unit and the sparse attention computing unit. The former extracts short-term and long-term features through forgetting factors and accumulation factors, respectively, while the latter only assigns weights to the data types that trigger anomalies, ultimately obtaining the fourth weighted representation sequence. This fourth weighted representation sequence is sent to the edge-side classification computing layer, where a fire occurrence probability distribution is formed through matrix operations and normalization. From this distribution, the first prediction result is solved and the first immediate command is generated to drive the execution of the local fire extinguishing device. The optimization module receives data uploaded from the edge on the cloud side and aligns it with the stored historical data with timestamps to form a complete full data sequence. The system calculates the mean, variance and trend slope of various types of data within a specified time window to form statistical results, and compares them one by one between different devices to obtain the propagation relationship. The results are then aggregated and the rate of change between time windows is calculated to form a continuous gradient, which is finally combined into a fourth trend sequence. This fourth trend sequence is sent to the cloud classification and calculation unit, where the probability distribution of the fire stage is obtained through weighted and normalized processing. The stage with the highest probability value is selected as the second prediction result, and a second optimization instruction is generated. This second optimization instruction includes threshold parameters for adjusting the triggering conditions on the edge side and control parameters for correcting the injection volume, injection time, and injection direction. It is then sent back to the edge side to correct the real-time decision. After the first immediate instruction and the second optimized instruction are generated, the difference module extracts their respective trigger times and control parameters, aligns them on the time axis, calculates the parameter difference at each corresponding position, and records the sign and direction of the difference to obtain a difference sequence. The system further calculates the rate of change of the difference at adjacent time positions, and determines whether the instructions converge or deviate based on the magnitude and sign of the rate of change, thus forming a convergence or deviation marker sequence, which is combined with the difference sequence to form a difference trajectory. The synthesis module uses the difference and rate of change in the difference trajectory to dynamically allocate the weight coefficients of the two types of instructions, calculates the weighted results at the same time position, and performs superposition to generate a synthetic parameter sequence, finally obtaining the third synthetic instruction. This instruction contains the parameters after weight superposition and is used to directly drive the fire extinguishing device to perform fire extinguishing actions. The verification module collects feedback data in real time during the execution of the third synthesis instruction, including temperature changes, smoke concentration, voltage changes, and current changes, forming a feedback data sequence. The system calculates the difference in the sequence within a time window and its ratio to the time interval, forming a feedback curve. Subsequently, the feedback curve and the difference trajectory are compared one by one at the same time position, and the difference in value and the difference in rate of change are calculated respectively. When the difference in value exceeds the threshold in multiple time windows, or the difference in rate of change remains positive or negative in multiple time windows, it is determined that there is an execution deviation in the third synthesis instruction. Based on this, the update module further judges the overall trend. When the feedback curve and the difference trajectory show a continuous deviation at the statistical level, the system generates a fourth correction instruction. This correction instruction includes correction parameters for the forgetting factor and accumulation factor on the edge side, as well as correction values ​​for the classification threshold and weight parameters on the cloud side, and is uploaded to the cloud to update the models on both the edge and cloud sides simultaneously. After the update is completed, the new parameters are saved as the initial conditions for the next round of phase collision, ensuring that the system continuously corrects itself and tends to be consistent during the iteration process. This solution combines real-time edge-side response with global optimization on the cloud side. By using phase collision and differential trajectory analysis, it unifies and corrects command results from different sources, achieving prediction, optimization, synthesis, verification, and updating. This allows the solution to combine rapid response with long-term optimization, meeting the need for second-level response at fire scenes while avoiding misjudgments caused by data locality at the edge. It also achieves adaptive evolution and convergence in multiple iterations, ensuring the reliability of the fire extinguishing process.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent fire extinguishing system based on cloud-side cooperation, comprising a prediction module, an optimization module, a difference module, a synthesis module, a verification module and an update module, characterized in that: the prediction module is used to collect fire data and input the fire data into an edge-side deep learning model, and output a first prediction result and a corresponding first immediate instruction; the optimization module is used to transmit the fire data to a cloud-side deep learning model, and generate a second prediction result and a corresponding second optimization instruction; the difference module is used to perform phase collision on the first immediate instruction and the second optimization instruction on a time axis after the instructions are generated, form a difference trajectory, and represent a phase convergence state or a phase deviation state of the edge side and the cloud side in the execution direction through the difference trajectory; the synthesis module generates a third synthesis instruction based on the difference trajectory, and the third synthesis instruction is formed by superimposing parameter weights of the first immediate instruction and the second optimization instruction; the verification module is used to execute the third synthesis instruction, collect fire extinguishing feedback data in real time during the execution, form a feedback curve using the fire extinguishing feedback data, and compare the feedback curve with the difference trajectory to identify an execution deviation of the third synthesis instruction; and the update module is used to generate a fourth correction instruction when the comparison result of the verification module shows that the feedback curve and the difference trajectory are continuously deviated, and upload the fourth correction instruction and the difference trajectory to the cloud side to update corresponding model parameters of the edge side and the cloud side for iterative execution of the next round of phase collision.

2. The intelligent fire extinguishing system based on cloud-side cooperation according to claim 1, characterized in that: the construction and training of the edge-side deep learning model comprises: slicing the collected fire data on the edge side according to a fixed time window, and retaining only the part where the temperature, smoke, voltage and current change exceeds a preset threshold in each time window to form a first sparse input sequence; when any type of data in the first sparse input sequence is missing, generating a second compensation input sequence based on the gradient deduction result of the adjacent time window and the comparison result of the same type of sensor to form a continuous input to ensure that the inference link is not interrupted; inputting the second compensation input sequence into a hierarchical memory calculation unit, the hierarchical memory calculation unit forms a forgetting update vector in the first layer, forms a delay cumulative memory vector in the second layer, and combines to generate a third time sequence representation sequence; inputting the third time sequence representation sequence into a sparse attention calculation unit, the sparse attention calculation unit only allocates calculation weight to the data type triggering the anomaly to form a fourth weighted representation sequence; ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ inputting the fourth weighted feature sequence into an edge side classification calculation layer, the edge side classification calculation layer receiving the fourth weighted feature sequence in each time window and performing matrix multiplication and weighted summation operation to form a classification score vector; after forming the classification score vector, performing normalization operation on the classification score vector, calculating the exponential value of each category, and performing ratio operation of each exponential value with the sum of all exponential values to output a fire occurrence probability distribution; after outputting the fire occurrence probability distribution, identifying the upper limit of the probability value in the fire occurrence probability distribution and solving the first prediction result, the first prediction result including a category label and a corresponding probability value; the first prediction result and the fire occurrence probability distribution generate a first instant instruction, and when there is a difference between the first instant instruction and the optimization condition returned by the cloud, the weight distribution rules of the hierarchical memory calculation unit and the sparse attention calculation unit are updated to realize the adaptive evolution of the edge side model.

3. The intelligent fire extinguishing system based on cloud cooperation according to claim 2, wherein: the construction and training of the cloud side deep learning model includes: receiving the fire data uploaded by the edge side, and aligning with the historical fire data stored in the cloud according to the time stamp sequence to form a first full data sequence, the first full data sequence including temperature data, smoke data, voltage data and current data; in the first full data sequence, the mean, variance and trend slope of the temperature data, smoke data, voltage data and current data are calculated in a preset number of historical time windows to generate a second statistical result; the second statistical result is compared between a plurality of devices one by one to calculate the temperature difference, smoke difference, voltage difference and current difference between different devices to form a third comparison result; the third comparison result is aggregated according to a preset time window, and the change rate of temperature difference, smoke difference, voltage difference and current difference between adjacent time windows is calculated respectively, the change rate being solved by dividing the numerical difference of adjacent time windows by the time interval, thereby forming a continuous change gradient; after obtaining the continuous change gradient, the continuous change gradient of each data is combined in time sequence to form a fourth trend sequence; inputting the fourth trend sequence into the classification calculation unit of the cloud, multiplying each data in the fourth trend sequence by a corresponding weight value in the classification calculation unit and performing summation operation to form a stage score; after forming the stage score, the exponential value of each stage score is calculated in turn, and each exponential value is divided by the sum of all exponential values to output a probability distribution of fire development stage; after obtaining the probability distribution of fire development stage, selecting the maximum value of the probability value in the probability distribution, and determining the fire stage corresponding to the maximum value as the second prediction result, the second prediction result including the category of the fire stage and the probability value corresponding to the maximum value; The difference module includes a phase collision module and a difference result acquisition module.

4. The intelligent fire extinguishing system based on cloud collaboration according to claim 3, wherein: The phase collision process of the difference module includes: after the first instant instruction and the second optimized instruction are generated, the trigger time, trigger condition and control parameter of the first instant instruction and the trigger time, trigger condition and control parameter of the second optimized instruction are extracted respectively to form a first instruction sequence and a second instruction sequence; The trigger time of the first instruction sequence is aligned with the trigger time of the second instruction sequence, and at each aligned time position, the numerical difference between the control parameter in the first instruction sequence and the control parameter in the second instruction sequence is calculated to generate a control parameter difference value sequence, and the sign direction of each numerical difference is recorded in the control parameter difference value sequence; In the control parameter difference value sequence, the control parameter difference value change rate of adjacent time positions is calculated in sequence, and according to the positive and negative values and sizes of the control parameter difference value change rate, the phase convergence state or phase deviation state of the first instant instruction and the second optimized instruction at the time position is marked to generate a convergence deviation marker sequence.

5. The intelligent fire extinguishing system based on cloud collaboration according to claim 4, wherein: The phase collision result acquisition of the difference module includes: combining the control parameter difference value sequence and the convergence deviation marker sequence to form a difference trajectory, the difference trajectory including the control parameter difference value, the control parameter difference value change rate and the phase convergence state or phase deviation state of each time position, and the phase collision result of the first instant instruction and the second optimized instruction in the execution direction is represented by the difference trajectory.

6. The intelligent fire extinguishing system based on cloud collaboration according to claim 5, wherein: In the synthesis module, the generation process of the third synthetic instruction includes: Based on the control parameter difference value and the control parameter difference value change rate of each time position recorded in the difference trajectory, the weight coefficient of the first instant instruction and the weight coefficient of the second optimized instruction are calculated; The control parameter of the first instant instruction is multiplied by the corresponding weight coefficient, the control parameter of the second optimized instruction is multiplied by the corresponding weight coefficient, and the results at the same time position are added to form a synthetic parameter sequence; The synthetic parameter sequence is combined in time sequence, and the third synthetic instruction is generated after the combination is completed, the third synthetic instruction including the control parameter after weight superposition.

7. The intelligent fire extinguishing system based on cloud collaboration according to claim 6, wherein: The execution process of the verification module includes: The third synthesis instruction is executed, and fire extinguishing feedback data is collected in real time during execution, the fire extinguishing feedback data including temperature change data, smoke concentration data, voltage change data and current change data, and a feedback data sequence is formed; The feedback data sequence is arranged in time sequence, temperature change difference, smoke concentration difference, voltage change difference and current change difference are calculated at each adjacent time position, and the ratio of each difference to time interval is calculated to form a feedback curve; The feedback curve and the difference trajectory are compared item by item at the same time position, and the difference difference and the change rate difference of the four types of data of temperature, smoke, voltage and current are calculated respectively during the comparison, and a comparison result sequence is formed; In the comparison result sequence, when the difference difference of any one of temperature, smoke, voltage or current exceeds the preset threshold value within the preset number of time windows, or the sign of the change rate difference is positive or negative within the preset number of time windows, it is determined that the third synthesis instruction has execution deviation.

8. The intelligent fire extinguishing system based on cloud collaboration according to claim 7, characterized in that: The execution process of the update module includes: After the comparison module completes the comparison, the difference difference and the change rate difference of the feedback curve and the difference trajectory in each time window are counted, and when the difference difference exceeds the preset threshold value within the preset number of time windows, or the sign of the change rate difference is consistent within the preset number of time windows, it is determined that the feedback curve and the difference trajectory have sustained deviation; After it is determined that there is sustained deviation, a fourth correction instruction is generated according to the size of the difference difference and the change rate difference, the fourth correction instruction including correction parameters of the forgetting factor and the cumulative factor on the edge side and correction values of the classification threshold and the weight parameter on the cloud side; The fourth correction instruction and the difference trajectory are uploaded to the cloud side together, the corresponding model parameters are updated on the cloud side and the edge side respectively, and the updated model parameters are stored as the initial input condition for the next round of phase collision; After completing the parameter update, the calculation process of the edge side deep learning model and the cloud side deep learning model is driven again using the updated model parameters, which is used for generating the first immediate instruction and the second optimization instruction in the next round, and is iteratively executed in the phase collision.

Images