Fire-fighting equipment intelligent inspection system based on edge computing

Abstract

The application provides a fire-fighting equipment intelligent inspection system based on edge computing and relates to the field of fire-fighting equipment state monitoring. The system is composed of a data acquisition terminal, an edge computing gateway and a cloud server. The acquisition terminal is configured with a multi-source sensor and a differentiated sampling interpolation alignment mechanism, and outputs a uniform time sequence multi-source sensor data frame. The edge side extracts a short-time feature vector, calculates time consistency by starting typical correlation analysis of transient multi-source features, determines continuous abnormalities and classifies and caches abnormal data. The cloud side quantifies equipment degradation rate based on dynamic time warping and time acceleration coefficient, and generates a health degree adjustment signal. Finally, the D-S evidence theory is used to fuse two-way signals to output a comprehensive health degree score, which reduces the transmission bandwidth, improves the accuracy of abnormal identification, supports hierarchical operation and maintenance and equipment life cycle health management.

Description

Technical Field

[0001] This invention relates to the field of fire equipment condition monitoring, specifically to an intelligent fire equipment inspection system based on edge computing. Background Technology

[0002] Fire pump sets are the core power equipment of building fire protection water supply systems. According to fire protection codes, they must undergo regular start-stop self-inspections and statutory periodic inspections. Currently, the industry generally deploys various types of sensors to collect pump set operating parameters, triggering abnormal alarms with fixed thresholds. Some systems upload data to the cloud for statistical analysis, and routine start-stop self-inspections are used to verify the equipment's startability.

[0003] However, the characteristic of fire pump sets being idle for extended periods and only operating briefly during self-tests means that effective data is highly concentrated in the startup transient. During this period, the signal characteristics of normal operating condition shocks and early degradation are highly coupled. Existing judgment methods cannot distinguish between instantaneous disturbances, startup shocks, and continuous equipment degradation anomalies in extremely short transient processes, making it difficult to simultaneously address both missed and false diagnoses. Furthermore, the discretely distributed transient data from each self-test cannot directly form a continuous state evolution sequence. If existing technologies are to quantify equipment degradation trends within the legally mandated inspection cycle, they either need to upload all data, increasing bandwidth and deployment costs, or only transmit alarm points, which cannot support continuous degradation rate assessment. Ultimately, it remains impossible to achieve accurate quantification of the long-term health status of equipment with low transmission costs. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent inspection system for fire-fighting equipment based on edge computing, comprising:

[0005] A data acquisition terminal includes a multi-source sensor cluster and a data interface. The data interface is connected to each sensor and configured to encapsulate and send multi-source sensor data frames according to a unified timestamp.

[0006] The data processing module includes an edge computing gateway and a cloud server. The edge computing gateway is deployed at the fire equipment site and receives multi-source sensor data frames.

[0007] The edge computing gateway includes a first processing unit, a second processing unit, and a health fusion output unit;

[0008] The first processing unit is configured to extract the time-domain statistical features and frequency-domain features of multi-source sensor data frames in units of a first time window, and generate a short-time feature vector.

[0009] The second processing unit is connected to the first processing unit and is configured to accumulate short-time feature vectors of N consecutive first time windows, calculate the temporal consistency coefficient between the N short-time feature vectors, and determine that there is a continuous abnormality when the coefficient is lower than a preset threshold. The corresponding multi-source sensor data frame is cached in the local memory and a first health adjustment signal is generated.

[0010] The cloud server includes a third processing unit connected to an edge computing gateway. It is configured to receive multi-source sensor data frames corresponding to continuous anomalies cached in the local memory at a preset period, stitch multiple first health adjustment signals into a health time-series curve, align the health time-series curve with a pre-stored historical health benchmark curve, calculate the cumulative deviation, calculate the equipment degradation rate based on the cumulative deviation and the remaining days of the statutory inspection period, and generate and send a second health adjustment signal based on the equipment degradation rate.

[0011] The health fusion output unit receives the first health adjustment signal and the second health adjustment signal, weights and fuses them, outputs a comprehensive health score, and uses this score to inspect the equipment.

[0012] Preferably, the multi-source sensor cluster includes a vibration acceleration sensor installed on the bearing housing of a fire hydrant pump group or sprinkler pump group, a dynamic pressure sensor installed on the outlet pipe network of the pump group, and a temperature sensor installed on the motor winding of the pump group. Within the same sampling period, the data interface configures the sampling rate of the vibration acceleration sensor to an integer multiple of the sampling rate of the dynamic pressure sensor, and configures the sampling rate of the temperature sensor to 1 / M of the sampling rate of the dynamic pressure sensor, where M is an integer greater than or equal to 10. Using the sampling time of the dynamic pressure sensor as the reference timestamp, the sampling data of the vibration acceleration and temperature sensors are linearly interpolated and aligned, and encapsulated into a multi-source sensor data frame containing pump group start / stop status identifier bits and pipe network steady-state / transient flow status identifier bits. By configuring differentiated sampling rates, the efficiency of rapid fault feature extraction and slow parameter monitoring is balanced, and the interpolation alignment mechanism eliminates temporal deviations in the multi-source data, providing a unified temporal reference for multi-dimensional feature fusion.

[0013] Preferably, the first processing unit performs wavelet packet decomposition and sub-band energy calculation on the vibration acceleration signal, extracting the energy proportion of the pump group bearing fault characteristic band and the energy proportion of the impeller cavitation characteristic band. The energy proportion of the characteristic band satisfies the formula... In the formula The signal energy within the target's characteristic frequency band. The total energy of the vibration signal across the entire frequency band is used. Empirical mode decomposition is performed on the dynamic pressure signal to extract the dominant frequency of pressure pulsation and its harmonic energy distribution. Combined with the temperature rise rate within the window calculated from the temperature signal, a short-time characteristic vector is constructed. The impeller cavitation characteristic frequency band is determined by the offline calibrated cavitation initiation test spectrum. The calibration process involves gradually reducing the inlet pressure under the pump's rated speed condition, recording the corresponding vibration spectrum when initial cavitation bubbles are detected on the impeller surface, and defining the cavitation characteristic frequency band as the interval where the energy concentration in the spectrum is higher than a preset energy threshold and the frequency range is within ±5% of the impeller blade passing frequency. The impeller blade passing frequency satisfies the formula... In the formula, Z is the number of impeller blades and n is the rated speed of the pump set; the cavitation characteristic frequency band after calibration is stored in the non-volatile memory of the edge computing gateway, which can be directly called during online monitoring without repeating online calibration, thus reducing the computing load on the edge side while ensuring the accuracy of cavitation identification.

[0014] Preferably, the second processing unit calculates the temporal consistency coefficient through multi-source feature correlation analysis during the start-up transient period: the edge computing gateway identifies the pump group start-up and stop status flag bits in the multi-source sensor data frame, captures the pump group start-up process triggered by periodic self-tests, extracts the start-up transient period from the moment the motor is energized to the moment the pipeline pressure reaches a steady-state value, and constructs a transient feature matrix from the short-time feature vectors of N consecutive first time windows within this period; performs canonical correlation analysis on the feature submatrix corresponding to vibration acceleration and the feature submatrix corresponding to dynamic pressure, and extracts the maximum canonical correlation coefficient as the temporal consistency coefficient; the canonical correlation analysis constructs a generalized eigenvalue problem. Solve for the equation, where , These are the covariance matrices of the two sets of features. Let a be the cross-covariance matrix of two sets of features, and let a and b be the linear combination coefficient vectors of the two sets of features. Solve for the largest eigenvalue. The square root of the time series consistency coefficient is the time series consistency coefficient. The preset threshold for the time-series consistency coefficient is obtained by offline calibration of the 95th percentile of the normal start-up sample under rated operating conditions. When the time-series consistency coefficient is lower than the threshold, it is determined that there is a continuous anomaly, and a first health adjustment signal is generated that decreases linearly with the decrease of the time-series consistency coefficient. At the same time, three levels of storage priority are assigned to the cached data according to the anomaly level and the pump group operating condition type. When the local storage capacity is insufficient, historical data is overwritten in order of priority from low to high, and high-value anomaly data is retained first under limited storage resources.

[0015] Preferably, the third processing unit dynamically times-aligns the health time-series curve with the historical health baseline curve within the statutory inspection period, calculates the cumulative deviation between the two, and then calculates the equipment degradation rate by combining the remaining days of the statutory inspection period, thereby generating a second health adjustment signal; dynamic time-warping solves for the optimal alignment path of the two curves by constructing a cumulative distance matrix, and the cumulative deviation is the sum of the distances of all sampling points on the optimal alignment path, satisfying... The distance between individual points is calculated using Euclidean distance, i.e. These correspond to the sampling points of the health time series and the baseline series, respectively. The calculation of the equipment degradation rate incorporates a time acceleration coefficient to correct the nonlinear degradation process. A health sub-sequence of a single pump start-up process is extracted as the monitoring sub-cycle, and after alignment and mapping to the baseline curve, an equivalent historical time span is obtained. The time acceleration coefficient satisfies... In the formula, T eq For the equivalent historical time span, T act To monitor the actual duration of sub-cycles; the corrected equivalent cumulative deviation. Equipment degradation rate In the formula, T remain The remaining days of the statutory inspection cycle are used to compare the actual degradation rate with the normal degradation rate derived from the statutory inspection requirements. A second health adjustment signal is obtained through linear mapping, thereby achieving accurate quantification of the long-term degradation trend of the equipment.

[0016] Preferably, the health fusion output unit uses DS evidence theory to fuse and evaluate two health signals, constructs a recognition framework {normal, attentive, abnormal, severe}, treats the first and second health adjustment signals as independent evidence bodies, generates basic probability assignments for each state according to preset mapping rules, and calculates the conflict coefficient between the two evidence bodies. In the formula, A is the focal element of the evidence body corresponding to the first health adjustment signal, and B is the focal element of the evidence body corresponding to the second health adjustment signal. Both are subsets of the recognition frame, and the summation range covers all focal element pairs with empty intersections. When the conflict coefficient exceeds the preset threshold, it is determined that there is an evidence conflict between the short-term judgment result on the edge side and the long-term trend in the cloud. At this time, the weight of the second health adjustment signal is increased and the weight of the first health adjustment signal is decreased. At the same time, the third processing unit is triggered to extend the statistical duration of the health time sequence, and the device degradation rate and the second health adjustment signal are re-verified before being sent out. When the conflict coefficient does not exceed the preset threshold, the Dempster combination rule is adopted. Confidence scores are synthesized, and the final confidence scores of each state are used as weights to calculate a comprehensive health score with a clear value range, which supports subsequent hierarchical operation and maintenance decisions.

[0017] This invention provides an intelligent inspection system for fire-fighting equipment based on edge computing. It has the following advantages:

[0018] 1. This invention adopts a layered collaborative computing architecture of edge computing gateway and cloud server. Real-time data processing, feature extraction and initial anomaly judgment are completed at the edge. Only abnormal data is uploaded to the cloud, which greatly reduces the data transmission bandwidth consumption and cloud computing pressure, while ensuring the real-time response to on-site anomalies.

[0019] 2. This invention replaces the traditional single-parameter threshold alarm with a multi-source feature time sequence consistency judgment based on the transient process of pump group startup. It integrates multi-dimensional information such as vibration, pressure, and temperature, effectively filtering out false alarms caused by instantaneous sensor disturbances and operating condition fluctuations, and improving the accuracy and reliability of anomaly identification.

[0020] 3. This invention can more accurately quantify the rate of equipment degradation, while providing full-cycle operational data to support statutory inspections and providing complete technical support for the intelligent management of fire protection equipment throughout its entire life cycle. Attached Figure Description

[0021] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0022] Figure 1 This is a system framework diagram as described in this invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0024] This embodiment addresses the inspection scenario of fire hydrant pump sets in fire-fighting water supply systems of civil buildings, deploying an intelligent fire equipment inspection system based on edge computing, referencing... Figure 1 The system consists of a data acquisition terminal, an edge computing gateway, and a cloud server. The edge computing gateway is deployed at the fire pump room site, and the cloud server is deployed in the fire control center computer room. The data acquisition terminal and the edge computing gateway are connected via Ethernet wired connection and use the Modbus TCP protocol to transmit data. The data frame transmission period is synchronized with the first time window.

[0025] The data acquisition terminal includes a multi-source sensor cluster and a data interface. The multi-source sensor cluster includes a vibration acceleration sensor, a dynamic pressure sensor, and a temperature sensor. The vibration acceleration sensor is installed at a radial measuring point on the bearing housing of the fire hydrant pump unit's drive end. The dynamic pressure sensor is installed on a straight pipe section three times the pipe diameter after the pump unit's outlet flange. The temperature sensor is embedded at the end of the pump unit's motor stator winding. The data interface connects to each sensor. Within the same sampling period, the sampling rate of the vibration acceleration sensor is configured as an integer multiple of the sampling rate of the dynamic pressure sensor, and the sampling rate of the temperature sensor is configured as 1 / M of the sampling rate of the dynamic pressure sensor, where M is an integer greater than or equal to 10. In this embodiment, a single-stage centrifugal fire pump unit with a rated speed of 1450 r / min and 6 impeller blades is used as an example. The impeller blade passing frequency is calculated using the following formula: In the formula, Z represents the number of impeller blades, and n represents the rated speed of the pump unit. Substituting the values ​​into the formula yields... To fully capture the dominant frequency of pressure pulsations and its 6th harmonic component, the sampling rate of the dynamic pressure sensor is set to 1000Hz according to the Nyquist sampling theorem. The sampling rate of the vibration acceleration sensor is configured to be 8 times that of the dynamic pressure sensor, i.e., 8000Hz, which can cover the characteristic frequency of bearing faults and its higher harmonics, meeting the requirements for fault feature extraction. The sampling rate of the temperature sensor is configured to be 1 / 20 of that of the dynamic pressure sensor, i.e., M=20, corresponding to a sampling rate of 50Hz; the motor winding temperature is a slowly changing thermal process with a time constant greater than 10s, and this sampling rate can meet the accuracy requirements for monitoring the temperature rise rate.

[0026] The data interface uses the sampling time of the dynamic pressure sensor as the reference timestamp, and performs first-order linear interpolation to align the sampling data from the vibration acceleration sensor and temperature sensor, encapsulating multi-source data within the same sampling period into a multi-source sensor data frame. The multi-source sensor data frame includes a timestamp field, aligned sensor data fields, a pump start / stop status flag field, and a pipeline steady-state / transient flow status flag field, with a timestamp accuracy of 1ms. The pump start / stop status flag is determined by the motor circuit current signal: when the motor current is higher than 10% of the rated current, the flag indicates the start state; when the motor current is lower than 5% of the rated current and the duration exceeds 1 second, the flag indicates the stop state. The pipeline steady-state / transient flow status flag is determined by pressure signal fluctuations: if the pressure fluctuation amplitude is less than 2% of the rated working pressure within three consecutive sampling periods, it is considered steady state; otherwise, it is considered transient.

[0027] The edge computing gateway is deployed at the fire equipment site to receive multi-source sensor data frames sent by the data acquisition terminal. Internally, it includes a first processing unit, a second processing unit, and a health fusion output unit. The first processing unit extracts the time-domain statistical features and frequency-domain features of the multi-source sensor data frames in units of a first time window, generating a short-time feature vector. In this embodiment, the duration of the first time window is set to 0.5 seconds. For the vibration acceleration sensor signal, the first processing unit uses db4 wavelet packets for 3-level decomposition, dividing the signal frequency band into 8 sub-frequency bands, and calculating the signal energy of each sub-frequency band. The calculation formulas for the energy proportion of the bearing fault characteristic frequency band and the energy proportion of the impeller cavitation characteristic frequency band are as follows: In the formula The signal energy within the target's characteristic frequency band. This represents the total energy of the vibration signal across the entire frequency band.

[0028] The impeller cavitation characteristic frequency band is determined based on the pre-stored cavitation initiation test spectrum of the pump set at rated speed. The cavitation initiation test spectrum is obtained through offline calibration: under the rated speed condition of the pump set, the inlet pressure of the pump set is gradually reduced, and the impeller surface condition is monitored using a high-speed camera at a frame rate of 1000 frames per second. When the proportion of cavitation area in a single frame image of the impeller inlet area reaches 0.5%, it is determined that primary cavitation has occurred, and the spectral distribution of the vibration acceleration signal at this time is recorded. The energy concentration of each frequency interval in the spectrum is calculated. The energy concentration is the ratio of the energy in that frequency interval to the total energy of the entire frequency band. The frequency interval with an energy concentration higher than the preset energy threshold and the frequency range within ±5% of the impeller blade passage frequency is determined as the impeller cavitation characteristic frequency band. In this embodiment, the preset energy threshold is set to 15%, corresponding to a cavitation characteristic frequency band range of 137.75Hz to 152.25Hz. The calibrated impeller cavitation characteristic frequency band is stored in the non-volatile memory of the edge computing gateway, which can be directly accessed by the first processing unit during subsequent online monitoring.

[0029] For the signals from the dynamic pressure sensor, the first processing unit uses empirical mode decomposition (EMD) for processing. The decomposition terminates when the standard deviation of the obtained intrinsic mode functions is less than 0.2. After decomposition, Fourier transform is performed on each intrinsic mode function component to extract the energy distribution of the pressure pulsation dominant frequency and its first three harmonics. For the signals from the temperature sensor, the first processing unit calculates the temperature change within a single first time window to obtain the temperature rise rate, in °C / s. Finally, the short-time feature vector is composed of the energy proportion of the bearing fault characteristic frequency band, the energy proportion of the impeller cavitation characteristic frequency band, the pressure pulsation dominant frequency, the harmonic energy distribution, and the temperature rise rate.

[0030] The second processing unit is connected to the first processing unit. It accumulates short-time feature vectors from N consecutive first time windows and calculates the temporal consistency coefficient among these N short-time feature vectors. In this embodiment, N is set to 16, based on the following: the duration of the start-up transient period from the start-up of the fire pump unit to the steady-state pressure in the pipeline network is approximately 8 seconds, the duration of the first time window is 0.5 seconds, and N=16 corresponds to a total duration of 8 seconds, which can completely cover the entire start-up transient process and meet the sample size requirements for temporal consistency analysis.

[0031] The edge computing gateway identifies the pump start / stop status flag in multi-source sensor data frames. When it detects a pump start triggered by periodic self-testing, it extracts the transient start-up period from the moment the motor is powered on until the pipeline pressure reaches a steady-state value. The short-time feature vectors of 16 consecutive first time windows within this period are used to construct a transient feature matrix. The criteria for determining if the pipeline pressure reaches a steady-state value are: the pipeline pressure rises to more than 95% of the rated working pressure, and the pressure fluctuation amplitude is less than 2% of the rated working pressure within three consecutive first time windows. In the transient feature matrix, the feature submatrix X corresponding to the vibration acceleration sensor has a dimension of 16×2, containing two feature variables: the proportion of energy in the bearing fault frequency band and the proportion of energy in the impeller cavitation frequency band. The feature submatrix Y corresponding to the dynamic pressure sensor has a dimension of 16×2, containing the pressure pulsation main frequency, the energy of the first three harmonics, and two feature variables.

[0032] The second processing unit performs canonical correlation analysis on the two feature sub-matrices and extracts the maximum canonical correlation coefficient as the time series consistency coefficient. The solution process for canonical correlation analysis is as follows: calculate the covariance matrix of the two sets of features respectively. , With the cross-covariance matrix Constructing generalized eigenvalues: In the equation, a and b are the linear combination coefficient vectors of the two sets of features. Solving the above equation yields the largest eigenvalue. The square root of this value is the maximum canonical correlation coefficient, also known as the time-series consistency coefficient. .

[0033] The preset threshold for the timing consistency coefficient is determined through offline calibration: 20 sets of sample data from normal startup processes under rated operating conditions are collected, the timing consistency coefficient of each sample is calculated, and the 95th percentile of all sample coefficient values ​​is taken as the preset threshold. In this embodiment, the mean timing consistency coefficient of the 20 normal samples is 0.88, the standard deviation is 0.03, and the 95th percentile is approximately 0.83; therefore, the preset threshold is set to 0.83. When the timing consistency coefficient is lower than 0.83, it is determined that the pump group has a continuous abnormality. The second processing unit caches the multi-source sensor data frames of the corresponding time period to the local memory and generates a first health adjustment signal. The first health adjustment signal is a health correction coefficient ranging from 0 to 1, and its value decreases linearly with the decrease of the timing consistency coefficient. When the timing consistency coefficient is equal to the preset threshold of 0.83, the first health adjustment signal is 1.0; when the timing consistency coefficient drops to 0.5, the first health adjustment signal is 0.6.

[0034] The second processing unit also assigns storage priorities to multi-source sensor data frames cached in local memory based on the anomaly level of the persistent anomaly and the pump start / stop status flag in the multi-source sensor data frames. Anomaly levels are categorized by the deviation ratio of the timing consistency coefficient from a preset threshold: a deviation ratio greater than 30% is considered a severe anomaly, a deviation ratio between 10% and 30% is considered a general anomaly, and a deviation ratio less than 10% is considered a minor anomaly. Storage priorities are divided into three levels: the highest priority is assigned to abnormal data frames under emergency startup conditions, the medium priority to abnormal data frames under periodic self-test conditions, and the lowest priority to abnormal data frames under normal standby conditions. Local memory is divided into three cache queues corresponding to the three priority levels, with a total capacity of 16GB. When the local memory capacity reaches its limit, cached data in each queue is overwritten sequentially according to priority from low to high, with higher-priority abnormal data being retained first.

[0035] The cloud server includes a third processing unit, which is communicatively connected to the edge computing gateway. This third processing unit receives multi-source sensor data frames corresponding to persistent anomalies cached in the local memory at a preset period. In this embodiment, the preset period is set to 24 hours, with data synchronization completed once daily at midnight. The third processing unit concatenates multiple first health adjustment signals in chronological order into a health time-series curve. It then aligns the health time-series curve with a pre-stored historical health baseline curve using dynamic time warping, calculating the cumulative deviation. Based on the cumulative deviation and the remaining days of the statutory inspection period, it calculates the device degradation rate and generates a second health adjustment signal based on the device degradation rate, which is then sent to the edge computing gateway.

[0036] The third processing unit dynamically aligns the health status time-series curve with the pre-stored historical health status benchmark curve within the statutory inspection cycle of the fire protection equipment, and extracts the cumulative deviation from the historical health status benchmark curve. The historical health status benchmark curve is a health status decay curve obtained by fitting data collected under normal operating conditions for three consecutive months after the equipment passes its first statutory inspection. The statutory inspection cycle is set to 365 days according to relevant fire protection equipment specifications.

[0037] The calculation process for dynamic time warping is as follows: Let the health time series be... The historical health baseline sequence is The distance between the i-th and j-th sampling points in the two sequences is calculated using Euclidean distance: Construct the cumulative distance matrix D, whose elements satisfy the recurrence relation: The boundary condition is D(1,1)=d(1,1), and the first row and first column of the matrix can only be calculated recursively in one direction. Starting from the bottom right corner D(m,n) of the matrix and backtracking to the top left corner D(1,1), the optimal alignment path is obtained. The cumulative deviation is the sum of the distances between all points on the optimal alignment path, i.e.: .

[0038] The third processing unit calculates the equipment degradation rate based on the cumulative deviation and the remaining days of the statutory inspection cycle, and generates a second health adjustment signal based on the equipment degradation rate. The specific calculation process is as follows: based on the pump start / stop status flag in the multi-source sensor data frame corresponding to the health time-series curve, the health sub-sequence during the most recent pump start event is extracted as the monitoring sub-cycle; the monitoring sub-cycle is dynamically time-normalized and mapped to the time axis of the historical health baseline curve, and the equivalent historical time span T corresponding to the health change within the monitoring sub-cycle is extracted. eq This refers to the duration of a continuous interval on the historical health baseline curve where the difference in health status at the beginning and end of the monitoring sub-cycle is equal to the duration of the interval. The equivalent historical time span is calculated as T = (Equation T - 1) / (Equation T - 1). act The ratio of the two values ​​yields the time acceleration coefficient k. t ,in .

[0039] Multiply the cumulative deviation by the time acceleration factor to obtain the equivalent cumulative deviation. Then divide the equivalent cumulative deviation by the remaining number of days T in the statutory inspection cycle. remain The equipment degradation rate v is obtained, with units of health score / d, where .

[0040] The preset normal degradation rate v0 is objectively derived from the statutory inspection requirements: the initial value of the equipment health within the statutory inspection cycle is 100 points, and the requirement for passing the inspection is a health score of not less than 80 points. Therefore, the maximum allowable normal degradation within the statutory cycle is 20 points, and the normal degradation rate is... The second health adjustment signal is a correction coefficient ranging from 0 to 1, obtained by linearly mapping the ratio of the actual degradation rate to the normal degradation rate: when v≤v0, the second health adjustment signal is 1.0; when v>v0, the adjustment signal value decreases linearly as the degradation rate increases.

[0041] The health score fusion output unit is connected to the second and third processing units, respectively. It receives the first and second health score adjustment signals, weights and fuses them, and outputs a comprehensive health score. The fusion process is based on the DS evidence theory, constructing a recognition framework {normal, attentive, abnormal, severe}. The first and second health score adjustment signals are treated as independent evidence bodies, and the basic probability allocation corresponding to each evidence body is calculated according to a preset mapping rule.

[0042] The basic probability allocation mapping rule for the first health adjustment signal s1 is as follows: When s1 ≥ 0.8, m1 (normal) = 0.7, m1 (attention) = 0.2, m1 (abnormal) = 0.1, m1 (severe) = 0; when 0.6 ≤ s1 < 0.8, m1 (normal) = 0.2, m1 (attention) = 0.6, m1 (abnormal) = 0.2, m1 (severe) = 0; when 0.4 ≤ s1 < 0.6, m1 (normal) = 0, m1 (attention) = 0.2, m1 (abnormal) = 0.6, m1 (severe) = 0.2; when s1 < 0.4, m1 (normal) = 0, m1 (attention) = 0.1, m1 (abnormal) = 0.2, m1 (severe) = 0.7. The mapping rule for the second health adjustment signal s2 is consistent with the above rules.

[0043] The conflict coefficient K between two pieces of evidence is calculated using the following formula: The summation range covers all focal element pairs with empty intersections. The preset conflict threshold is set to 0.6, which is derived based on the following: According to the statistical analysis of historical operating data of similar fire pump sets, the 99th percentile of the conflict coefficient between the short-term judgment result on the edge side and the long-term trend analysis result on the cloud under normal operating conditions is 0.58. After taking the engineering margin, the threshold is determined to be 0.6.

[0044] When the conflict coefficient K exceeds 0.6, it is determined that there is an evidence conflict between the short-term anomaly on the edge side and the long-term trend in the cloud. At this time, the second health adjustment signal is used first, and the weight of the first health adjustment signal is reduced. The weight of the first health adjustment signal is set to 0.2, and the weight of the second health adjustment signal is set to 0.8. At the same time, the third processing unit is triggered to re-verify the device degradation rate. During the re-verification, the statistical duration of the health time series curve is extended to twice the original duration, the device degradation rate and the second health adjustment signal are recalculated, and then sent to the health fusion output unit.

[0045] When the conflict coefficient K does not exceed the preset threshold, the Dempster combination rule of the DS evidence theory is used to synthesize the confidence level. The combination rule formula is as follows: In the formula, A is any focal element in the recognition frame, and A is not equal to 0.

[0046] After synthesis, the final confidence scores for the four states are obtained. Values ​​are assigned to the four states within the recognition framework: 100 points for normal, 75 points for attentive, 50 points for abnormal, and 25 points for severe. A weighted sum is then calculated using the final confidence scores of each state as weights to obtain a comprehensive health score ranging from 0 to 100. The calculation formula is: Score = 100m (normal) + 75m (attentive) + 50m (abnormal) + 25m (severe).

[0047] The system automatically matches the maintenance level based on the comprehensive health score range: when the score is 85 or above, the equipment is determined to be in a healthy operating state, and the regular self-inspection cycle is maintained; when the score is between 60 and 85, the equipment is determined to have early signs of degradation, the self-inspection frequency is automatically increased and key inspection prompts are pushed to guide maintenance personnel to conduct targeted inspections of vulnerable parts such as bearings and impellers; when the score is below 60, the equipment is determined to have a clear abnormality, the early warning is automatically triggered and a maintenance work order is generated, so as to realize the early detection and early handling of abnormal conditions and avoid equipment failure in fire conditions due to operation with defects.

[0048] Because this comprehensive health score integrates dual evidence—short-term anomaly detection at the edge and long-term trend analysis in the cloud—and eliminates interference from single sensor disturbances through an evidence conflict verification mechanism, its evaluation results exhibit significantly higher stability and accuracy than traditional single-parameter threshold alarms, effectively reducing invalid inspection work orders. In equipment lifecycle management and statutory inspection scenarios, this score, combined with historical health time-series curves stored in the cloud, can continuously trace the equipment degradation process, providing full-cycle operational data to support statutory inspections and overcoming the limitation that statutory inspections only reflect the status at the moment of detection. Maintenance personnel can predict the remaining service life of equipment based on the long-term trend of the health score, proactively plan spare parts procurement and downtime maintenance, balance maintenance costs and equipment reliability, and improve the problem of both reactive maintenance and over-maintenance in traditional inspection models.

[0049] 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 fire equipment intelligent inspection system based on edge computing, characterized in that, include: A data acquisition terminal includes a multi-source sensor cluster and a data interface. The data interface is connected to each sensor and configured to encapsulate and send multi-source sensor data frames according to a unified timestamp. The data processing module includes an edge computing gateway and a cloud server. The edge computing gateway is deployed at the fire equipment site and receives multi-source sensor data frames. The edge computing gateway includes a first processing unit, a second processing unit, and a health fusion output unit; The first processing unit is configured to extract the time-domain statistical features and frequency-domain features of multi-source sensor data frames in units of a first time window, and generate a short-time feature vector. The second processing unit is connected to the first processing unit and is configured to accumulate short-time feature vectors of N consecutive first time windows, calculate the temporal consistency coefficient between the N short-time feature vectors, and determine that there is a continuous abnormality when the coefficient is lower than a preset threshold. The corresponding multi-source sensor data frame is cached in the local memory and a first health adjustment signal is generated. The cloud server includes a third processing unit connected to an edge computing gateway, configured to receive multi-source sensor data frames corresponding to continuous anomalies cached in the local memory at a preset period, splice multiple first health adjustment signals into a health time-series curve, align the health time-series curve with a pre-stored historical health benchmark curve, and calculate the cumulative deviation. The equipment degradation rate is calculated based on the cumulative deviation and the remaining days of the statutory inspection cycle, and a second health adjustment signal is generated and sent based on the equipment degradation rate. The health fusion output unit receives the first health adjustment signal and the second health adjustment signal, weights and fuses them, outputs a comprehensive health score, and uses this score to inspect the equipment.

2. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 1, characterized in that: The multi-source sensor cluster includes: a vibration acceleration sensor installed on the bearing housing of a fire hydrant pump set or a sprinkler pump set, a dynamic pressure sensor installed on the outlet pipe network of the pump set, and a temperature sensor installed on the motor windings of the pump set.

3. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 2, characterized in that: The data interface is further configured to: within the same sampling period, configure the sampling rate of the vibration acceleration sensor to an integer multiple of the sampling rate of the dynamic pressure sensor, configure the sampling rate of the temperature sensor to 1 / M of the sampling rate of the dynamic pressure sensor, where M is an integer greater than or equal to 10, and use the sampling time of the dynamic pressure sensor as a reference timestamp to perform linear interpolation and alignment on the sampling data of the vibration acceleration sensor and the temperature sensor, and encapsulate them into a multi-source sensing data frame; the multi-source sensing data frame also includes a pump start / stop status flag and a pipeline steady-state / transient flow status flag.

4. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 3, characterized in that: The first processing unit is configured to: perform wavelet packet decomposition on the signal from the vibration acceleration sensor to extract the energy proportion of the pump group bearing fault characteristic frequency band and the energy proportion of the impeller cavitation characteristic frequency band; perform empirical mode decomposition on the signal from the dynamic pressure sensor to extract the pressure pulsation dominant frequency and its harmonic energy distribution; and combine the energy proportion of the bearing fault characteristic frequency band, the energy proportion of the impeller cavitation characteristic frequency band, the pressure pulsation dominant frequency and harmonic energy distribution, and the temperature rise rate of the temperature sensor to form the short-time feature vector.

5. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 4, characterized in that: The first processing unit is further configured to: determine the impeller cavitation characteristic frequency band based on the pre-stored cavitation initiation test spectrum of the pump set at rated speed; the cavitation initiation test spectrum is obtained by: gradually reducing the inlet pressure under the rated speed condition of the pump set until the generation of cavitation bubbles on the impeller surface is detected, recording the spectral distribution of the vibration acceleration signal at this time, and determining the frequency range in the spectral distribution where the energy concentration is higher than the preset energy threshold and the frequency range is within ±5% of the impeller blade passing frequency as the impeller cavitation characteristic frequency band; The impeller cavitation characteristic frequency band is stored in the non-volatile memory of the edge computing gateway, which can be directly accessed by the first processing unit during subsequent online monitoring.

6. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 1, characterized in that: The second processing unit calculates the timing consistency coefficient in the following ways: the edge computing gateway identifies the pump group start / stop status flag in the multi-source sensor data frame; when it detects that the pump group is started by periodic self-test, it extracts the start-up transient period from the moment the motor is powered on to the moment the pipeline pressure reaches a steady-state value; it constructs a transient feature matrix from the short-time feature vectors of N consecutive first time windows within this period; it performs canonical correlation analysis on the feature sub-matrix corresponding to the vibration acceleration sensor and the feature sub-matrix corresponding to the dynamic pressure sensor, and extracts the maximum canonical correlation coefficient as the timing consistency coefficient; when the timing consistency coefficient is lower than a preset threshold, it is determined that there is a continuous anomaly.

7. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 6, characterized in that: The second processing unit is further configured to: allocate storage priorities to the multi-source sensor data frames cached in the local memory based on the anomaly level of the continuous anomaly and the pump group start / stop status flag bit in the multi-source sensor data frames; wherein, the abnormal data frames under emergency start conditions are allocated the highest priority, the abnormal data frames under periodic self-test conditions are allocated the medium priority, and the abnormal data frames under normal standby conditions are allocated the lowest priority; when the local memory capacity reaches the upper limit, the cached data is overwritten in order of priority from low to high.

8. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 1, characterized in that: The third processing unit is configured to: dynamically time-align the health time-series curve with the pre-stored historical health benchmark curve within the statutory inspection cycle of the fire equipment, and extract the cumulative deviation from the historical health benchmark curve. The equipment degradation rate is calculated based on the cumulative deviation and the remaining days of the statutory inspection cycle, and a second health adjustment signal is generated based on the equipment degradation rate.

9. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 8, characterized in that: Based on the pump start / stop status flag in the multi-source sensor data frame corresponding to the health time series curve, the health subsequence during the duration of the most recent pump start event is extracted as a monitoring sub-cycle. The monitoring sub-cycle is dynamically time-normalized and mapped to the time axis of the historical health baseline curve. The equivalent historical time span corresponding to the change in health within the monitoring sub-cycle is extracted. The ratio of the equivalent historical time span to the actual duration of the monitoring sub-cycle is calculated to obtain the time acceleration coefficient. The cumulative deviation is multiplied by the time acceleration coefficient to obtain the equivalent cumulative deviation. The equivalent cumulative deviation is divided by the remaining days of the statutory inspection cycle to obtain the equipment degradation rate. The second health adjustment signal is generated based on the equipment degradation rate.

10. The intelligent inspection system for fire-fighting equipment based on edge computing according to claim 1, characterized in that: The health fusion output unit is configured to: treat the first health adjustment signal and the second health adjustment signal as independent evidence bodies, construct an identification framework {normal, attention, abnormal, severe}; calculate the conflict coefficient between the two evidence bodies; when the conflict coefficient exceeds a preset threshold, it is determined that there is an evidence conflict between the short-term anomaly on the edge and the long-term trend in the cloud; at this time, the second health adjustment signal is used first and the weight of the first health adjustment signal is reduced, and the third processing unit is triggered to re-verify the device degradation rate; when the conflict coefficient does not exceed the preset threshold, the DS evidence theory is used to allocate confidence and output the comprehensive health score.

Images