Factory safety intelligent monitoring method and system based on multi-source data fusion

Abstract

This invention relates to the field of industrial safety monitoring, and more particularly to a method and system for intelligent factory safety monitoring based on multi-source data fusion. The method includes: processing multi-source data on equipment vibration, power, and temperature, converting them into a unified scalar sequence with the same period after filtering and preprocessing; determining the heat conduction hysteresis constant by combining it with historical health data to achieve time-series alignment; processing temperature data to obtain the thermal evolution rate, obtaining the relative thermal evolution rate after correction based on a health benchmark; removing the normal heat generation component corresponding to the load through a local regression model for operating condition partitions to obtain the pure thermal anomaly residual; performing local window adaptive normalization on vibration energy to obtain impact anomaly characteristics. The pure thermal anomaly residual and the impact anomaly characteristics are fused through a sliding window multiplication to obtain the final deviation characteristic; an adaptive alarm threshold is set to achieve early wear monitoring and alarm for equipment. This invention effectively solves the problems of signal timing misalignment and false alarms due to load fluctuations, improving the accuracy of early wear identification.

Description

Technical Field

[0001] This invention relates to the field of industrial safety monitoring. In particular, it relates to an intelligent factory safety monitoring method and system based on multi-source data fusion. Background Technology

[0002] In factories with continuous production, the operating status of key equipment such as compressors, fans, and motors directly determines the continuity of production and the safety of personnel. With the improvement of industrial intelligence, intelligent factory technologies based on data perception and intelligent analysis are gradually replacing traditional manual inspections, becoming an important support for ensuring equipment health, preventing production accidents, and improving operation and maintenance efficiency. Through real-time collection and intelligent judgment of operating parameters, automatic early warning of abnormal equipment conditions can be achieved, providing safe and continuous protection for equipment.

[0003] However, existing equipment monitoring technologies still have significant shortcomings: traditional methods often use single sensor threshold alarms. When the equipment load fluctuates frequently, vibration and temperature will rise synchronously with the load, which can easily generate a large number of false alarms. At the same time, the temperature signal has an inherent hysteresis due to heat conduction, and there is a phase misalignment with the transient vibration signal. When using traditional synchronous fusion algorithms, multi-source features are prone to cancellation, which cannot effectively identify early minor wear and cause missed fault reports. Summary of the Invention

[0004] To address the issues that traditional single-sensor threshold alarms are prone to false alarms due to load fluctuations, and that traditional multi-source data fusion methods suffer from signal phase misalignment and feature cancellation caused by temperature thermal conduction hysteresis, making it difficult to identify early minor wear and tear of equipment and resulting in missed alarms, this invention provides solutions in the following aspects.

[0005] In the first aspect, the intelligent factory safety monitoring method based on multi-source data fusion includes: collecting multi-source data such as vibration data, power data, and temperature data from equipment; performing filtering preprocessing on the multi-source data; uniformly resampling the multi-source data to the same macroscopic period to form a scalar sequence; determining the inherent thermal conduction hysteresis constant based on equipment health history data; completing the time-series alignment of the multi-source data; sequentially performing signal optimization and change feature extraction on the time-series aligned temperature data to obtain the thermal evolution rate; constructing health benchmark parameters based on equipment health history data; using the health benchmark parameters to correct the thermal evolution rate, eliminating the influence of environmental temperature drift and normal operating condition fluctuations, and obtaining the relative thermal evolution rate; based on pre-defined... A local regression model for operating condition zones, constructed using historical health data, is used to obtain the normal heat generation component corresponding to the current load. The difference between the relative thermal evolution rate and the normal heat generation component is calculated to obtain the pure thermal anomaly residual characterizing abnormal heat generation. Local window adaptive normalization processing is performed on the time-aligned vibration energy to obtain the normalized vibration impact characterizing the degree of impact anomaly. The normalized vibration impact and the pure thermal anomaly residual are multiplied and fused within a sliding window to obtain the final deviation feature used to determine the fault. An adaptive alarm threshold is set for the final deviation feature using the box plot method. Based on the comparison between the final deviation feature and the adaptive alarm threshold, it is determined whether it is early wear of the equipment and an abnormal alarm is triggered.

[0006] Preferably, the step of forming a scalar sequence includes: The vibration, power, and temperature data after filtering and preprocessing are segmented according to a preset macroscopic period duration. The root mean square value of the vibration data within the macroscopic period corresponding to each segment is used as the vibration energy scalar, the average active power is used as the power scalar, and the average temperature is used as the temperature scalar. These are then sorted sequentially according to time to obtain the corresponding vibration scalar sequence, power scalar sequence, and temperature scalar sequence. The length of the macroscopic period is selected based on the temperature response inertial characteristics.

[0007] Preferably, the relative thermal evolution rate is obtained as follows: Taking any period as the target period, calculate the average temperature data within the target period, and obtain the temperature change by the difference between the average temperature data within the previous period and the average temperature data within the target period. The time interval between the temperature change and the macroscopic period is taken as the thermal evolution rate of the target period. When the computing device is healthy and fault-free, the thermal evolution rate corresponding to all cycles is traversed, a set of thermal evolution rates is constructed, the median of the set of thermal evolution rates is obtained, and the difference between the thermal evolution rate of the target cycle and the median is taken as the relative thermal evolution rate.

[0008] Preferably, the steps for establishing the local regression model for the operating condition zoning include: Historical power data and relative thermal evolution rate data of the equipment under healthy and fault-free conditions are collected; the historical health data are divided into three power ranges: light load, medium load, and heavy load according to the rated power of the equipment; within each power range, the regression coefficients and intercept terms of the corresponding range are obtained by linear regression training with the power change rate as input and the relative thermal evolution rate as output; the linear regression models obtained by training each range independently are combined to form a local regression model of the operating condition zone applicable to all operating conditions.

[0009] Preferably, the normalized vibration and impact data are obtained as follows: Using the current period as the end point of the window, a sliding statistical window of a preset length is constructed. Vibration scalar sequences of all periods within the window are extracted, and the mean and standard deviation of the vibration scalar sequences within the window are calculated respectively. The difference between the current vibration energy and the mean of the vibration scalar sequence within the window is calculated to obtain the vibration energy offset. The vibration energy offset is then divided by the sum of the standard deviation and the minimum constant within the window to obtain the normalized vibration impact.

[0010] Preferably, the final deviation feature is calculated as follows: Using the current period as the end point of the window, obtain a series of preset periods forward. In each period corresponding to the sequence number within the window, multiply the vibration impact and the pure thermal anomaly residual to obtain the single-period fusion feature. Accumulate all single-period fusion features within the window to obtain the sum of fusion features within the window. Divide the sum of fusion features by the window length to obtain the final deviation feature of the current period.

[0011] Preferably, the step of setting an adaptive alarm threshold for the final deviation feature using the box plot method includes: Based on the final deviation characteristic data under the healthy operating conditions of the equipment, the upper quartile and interquartile range are calculated, and the adaptive alarm threshold is calculated by combining the preset empirical coefficients. When the final deviation characteristic is greater than the adaptive alarm threshold, the early wear alarm of the equipment is triggered. Conversely, when it is less than or equal to the adaptive alarm threshold, the equipment is in normal operating condition.

[0012] Secondly, a factory safety intelligent monitoring system based on multi-source data fusion includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned factory safety intelligent monitoring method based on multi-source data fusion is implemented.

[0013] The present invention has the following effects: 1. This invention addresses the heterogeneous response characteristics of vibration, power signal instantaneous response, and temperature signal with thermal conduction hysteresis. Based on the inherent thermal conduction hysteresis constant of the equipment, it performs precise time-series alignment of multi-source data, overcoming the technical defects of traditional multi-source fusion methods that cause feature cancellation and masking of effective fault features due to signal phase misalignment. Through temperature time-series change feature extraction and multi-level benchmark correction, it accurately captures the weak and slow thermal anomaly evolution trend caused by early wear of the equipment, which can accurately capture early weak wear characteristics of the equipment, significantly improve the sensitivity of minor fault identification, and effectively reduce the probability of missed fault detection.

[0014] 2. This invention constructs a local regression model for different operating conditions, and fits the normal heating law corresponding to load fluctuations for different operating conditions such as light load, medium load, and heavy load, based on the nonlinear correlation characteristics between motor load and heat generation. It accurately removes the normal temperature rise component caused by dynamic load changes. At the same time, it combines the equipment health history benchmark to eliminate the interference of ambient temperature drift and steady-state operating condition fluctuations, effectively distinguishing between normal load fluctuations and real fault anomalies, and significantly reducing the false alarm rate of monitoring in continuous industrial production scenarios from a mechanistic perspective.

[0015] 3. This invention extracts pure mechanical impact anomaly features through local window adaptive normalization, obtains pure thermal anomaly residuals through multi-level differential correction and model decoupling, and achieves joint judgment of "impact-heating" co-origin fault features based on a sliding window product fusion mechanism. It also utilizes the dual-feature symbiotic mechanism of faults to suppress interference from single-signal disturbances. Simultaneously, it employs a box plot method independent of data distribution to construct adaptive alarm thresholds, avoiding threshold failure issues caused by extreme data and distribution deviations. Ultimately, this forms a multi-dimensional, anti-interference, highly sensitive, and highly stable intelligent monitoring mechanism, effectively improving the accuracy and reliability of factory equipment safety monitoring and ensuring the stable operation of continuous production equipment. Attached Figure Description

[0016] Figure 1 This is a flowchart of steps S1-S5 in the intelligent factory safety monitoring method based on multi-source data fusion according to an embodiment of the present invention.

[0017] Figure 2 This is a structural block diagram of the intelligent factory safety monitoring system based on multi-source data fusion, according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0019] Reference Figure 1 The intelligent factory safety monitoring method based on multi-source data fusion includes steps S1-S5, as follows: S1: Collect multi-source data including vibration, power, and temperature data from the equipment. After performing filtering preprocessing on the multi-source data, resample the multi-source data to the same macroscopic period to form a scalar sequence. Determine the inherent thermal conduction hysteresis constant based on the equipment's health history data to complete the time alignment of the multi-source data.

[0020] In the production workshop, acceleration sensors are installed on core vulnerable parts such as equipment bearing housings and motor housings to collect equipment vibration signals, with the signal sampling frequency set to 50Hz; the instantaneous value of active power of the equipment is collected through frequency converters or power monitoring instruments, with the power signal sampling frequency set to 1kHz; temperature sensors are deployed in the concentrated heat areas of the equipment to collect temperature signals, with the signal sampling frequency set to 1Hz.

[0021] After data acquisition, the three types of raw signals are preprocessed using median filtering with a preset window length. For example, the preset window length is 7, which can be adjusted according to specific circumstances. This filtering method can effectively filter out isolated outliers caused by factors such as instantaneous sensor disconnection, on-site electromagnetic interference, and signal mutations, while completely preserving the true variation patterns and trends of the signal itself, providing clean and reliable basic data for subsequent time-series alignment, feature calculation, and multi-source data fusion.

[0022] The raw vibration, power, and temperature data collected during equipment operation are high-frequency continuous time-series data, and the response characteristics of these three types of data differ significantly: vibration and power data can respond instantaneously to changes in equipment operating conditions and abnormal mechanical impacts, exhibiting rich high-frequency signal characteristics; while temperature data is affected by the heat transfer characteristics of the equipment material and structure, exhibiting thermal response inertia, resulting in slow signal changes and no high-frequency abrupt changes. Directly using the raw high-frequency data for feature analysis can easily introduce interference due to the mismatch in response characteristics of the multi-source data, leading to the failure of fault feature extraction. Therefore, based on the temperature response inertia characteristics, an appropriate macroscopic period duration is set, and the multi-source raw data undergoes unified segmentation and dimensionality reduction processing.

[0023] Specifically, the vibration data, power data, and temperature data after filtering and preprocessing are segmented into equal-length time sequences according to a preset macroscopic cycle duration. The macroscopic cycle duration is selected based on the inertial characteristics of the equipment's temperature response, matching the thermal conduction hysteresis characteristics and slow change pattern of the temperature signal. This effectively smooths the random fluctuations of the temperature signal, preserves the true thermal evolution trend of the equipment, and avoids the problem of early weak fault characteristics being averaged out due to excessively long cycle durations, thus achieving physical adaptation of the multi-source data analysis scale.

[0024] Within the segmented time intervals corresponding to each macro cycle, steady-state features are extracted from three types of data: the root mean square value of vibration data within a single cycle is selected as the vibration energy scalar to quantify the overall vibration intensity and impact energy level of the equipment within a single cycle; the average value of active power within a single cycle is selected as the power scalar to characterize the actual load operating state of the equipment in the current cycle; and the average value of temperature data within a single cycle is selected as the temperature scalar to reflect the steady-state temperature rise level of the equipment within a single cycle.

[0025] Following the temporal evolution sequence, the vibration energy scalar, power scalar, and temperature scalar corresponding to each macroscopic cycle are arranged sequentially to construct continuous temporal sequences of vibration scalar, power scalar, and temperature scalar, respectively. After the above processing, the three types of homogeneous but heterogeneous data achieve temporal scale unification and feature dimensionality reduction, eliminating the invalid redundant information of the original high-frequency data and the interference caused by differences in signal response characteristics. This provides a scale-consistent basic temporal feature sequence for subsequent multi-source data temporal alignment and anomalous feature fusion analysis.

[0026] Vibration and power signals are instantaneous response signals; corresponding data changes occur instantly when equipment friction, impact, or load changes occur. Temperature signals, on the other hand, are thermally hysteretic signals; due to the thermal inertia and heat conduction hysteresis characteristics of the equipment, temperature changes significantly lag behind changes in mechanical and electrical parameters. Therefore, multi-source raw data naturally exhibit temporal misalignment and cannot be directly fused point-to-point. By quantifying a fixed inherent heat conduction hysteresis constant using historical equipment health data, and compensating for signal hysteresis differences through time-series shifting, precise time-series alignment of vibration, power, and temperature data is achieved. The specific steps are as follows: Historical operating data of the equipment, including synchronously collected vibration, power, and temperature data, is selected as the benchmark data for time hysteresis quantification. Fault data, abnormal start-up and shutdown data, and extreme environmental interference data are excluded to ensure the validity and representativeness of the benchmark data.

[0027] Based on the health history dataset, vibration scalar sequence, power scalar sequence, and temperature scalar sequence are generated by uniformly resampling according to the preset macroscopic cycle. Typical excitation moments such as step changes in equipment load and start-stop switching are selected, and multi-source feature time-series segments before and after the excitation occur are extracted as hysteresis analysis samples.

[0028] Using the power / vibration characteristic sequence with the fastest response speed as the reference sequence, a discrete cross-correlation operation is performed between the temperature characteristic sequence and the reference sequence. This process iterates through different time-series offsets, calculating the cross-correlation coefficient corresponding to each offset. The offset period corresponding to the maximum cross-correlation coefficient is taken as the inherent thermal conduction hysteresis constant of the equipment. The inherent thermal conduction hysteresis constant is an inherent property determined by the equipment's structure, materials, and heat transfer conditions. It remains fixed under healthy conditions and characterizes the fixed lag period of the temperature signal relative to mechanical and electrical signals. It should be noted that the inherent thermal conduction hysteresis constant is solved through cross-correlation calculations of historical health data. The inherent thermal conduction hysteresis constant is determined by the inherent physical properties of the equipment structure, heat exchange materials, and heat dissipation layout, and its value remains stable during the steady-state healthy operation of the equipment. When the equipment undergoes long-term operation and exhibits aging and deterioration phenomena such as shell dust accumulation, bearing wear, and blockage of heat dissipation channels, the heat exchange efficiency of the equipment changes, and the temperature hysteresis characteristics shift accordingly. The system is configured with a periodic update mechanism for the hysteresis constant: it automatically collects healthy operation samples covering the entire load range on a weekly / monthly basis and re-executes cross-correlation analysis to iteratively correct the hysteresis constant. If a significant decrease in the matching degree of multi-source features is continuously observed during online monitoring, the system triggers a temporary recalibration process to update the hysteresis compensation parameters in real time, adapting to the changes in thermal hysteresis caused by equipment aging and broadening the adaptive boundary of the algorithm for long-term operation.

[0029] Based on the inherent thermal conduction hysteresis constant, the real-time acquired temperature scalar sequence is time-shifted and compensated, or the vibration and power sequences are matched in reverse time to eliminate the time shift caused by thermal hysteresis. This allows the mechanical shock, power fluctuation, and temperature anomaly characteristics corresponding to the same fault cause to be mapped to the same time dimension, thus achieving accurate time alignment of the three types of multi-source data.

[0030] S2: The time-aligned temperature data is sequentially optimized and its change features are extracted to obtain the thermal evolution rate. Based on the equipment health history data, health benchmark parameters are constructed. The thermal evolution rate is corrected using the health benchmark parameters to eliminate the influence of ambient temperature drift and normal operating condition fluctuations, and the relative thermal evolution rate is obtained.

[0031] After constructing the temperature scalar sequence, the real-time thermal change state of the equipment is quantified based on the time-series continuous temperature scalar, and relative thermal evolution features that can eliminate interference from normal operating conditions are extracted. This invention characterizes the dynamic temperature change trend of the equipment by the difference in the mean temperature between adjacent macroscopic periods, thereby calculating the thermal evolution rate and accurately reflecting the dynamic balance between real-time heat generation and dissipation of the equipment. The specific calculation process is as follows: Using any macroscopic cycle as the target cycle, the average temperature corresponding to the target cycle and the average temperature corresponding to the previous adjacent cycle are extracted. The temperature change of the equipment during this period is obtained by the difference between the average temperatures of the two adjacent cycles. This temperature change is divided by the fixed time interval corresponding to the macroscopic cycle to eliminate the influence of time scale differences and quantify the thermal evolution rate corresponding to the target cycle. This parameter can intuitively characterize the speed of temperature change of the equipment per unit time and reflect the real-time thermal evolution status of the equipment.

[0032] To eliminate fundamental thermal evolution interference caused by normal operating conditions such as equipment start-up and shutdown, ambient temperature fluctuations, and routine load fluctuations, a benchmark reference system that closely reflects the healthy operating state of the equipment is constructed. Historical time-series data of the equipment during its entire healthy and fault-free operation with stable operating conditions are selected. Following the aforementioned thermal evolution rate calculation method, the thermal evolution rate corresponding to all healthy cycles is calculated, and a set of thermal evolution rates under healthy equipment conditions is constructed. Further statistical analysis of this set of thermal evolution rates is performed, and the median of the set is obtained as the health benchmark threshold. This median has a strong resistance to outlier interference and can accurately reflect the basic temperature rise level under normal equipment operation, avoiding contamination of the benchmark value by extreme operating conditions and incidental data.

[0033] Finally, the relative thermal evolution rate is obtained by calculating the difference between the thermal evolution rate calculated in real time for the target period and the median thermal evolution rate corresponding to the healthy state. By stripping away the inherent normal thermal evolution baseline of the equipment and completely eliminating the normal temperature change components caused by environmental drift and fluctuations in normal operating conditions, only the abnormal thermal evolution components that exceed the normal operating state are retained. This can accurately characterize abnormal temperature rise changes caused by faults such as abnormal friction and component deterioration inside the equipment, providing pure and reliable core features for the accurate identification of early thermal faults in the equipment.

[0034] Specifically, the relative thermal evolution rate satisfies the following relationship: ; In the formula, Indicates the first The relative thermal evolution rate of the period, if A positive value indicates that the current temperature rise exceeds normal fluctuations, which may be due to an abnormal acceleration of thermal evolution. If the current temperature rise rate exceeds the normal background rate, there may be abnormal frictional heat generation. If the heating rate is normal, then Approaching 0; Indicates the first The average temperature over a period of time; Indicates the first Average temperature over a period of time; The time interval representing the macroscopic cycle, for example, is 1 second, which is used as the denominator to transform the first difference into the rate. This represents the thermal evolution rate, which is calculated for each cycle under the health condition of the equipment. A set; It represents the median thermal evolution rate during the healthy phase, reflecting the average rate of temperature rise under normal equipment start-up and shutdown and environmental changes.

[0035] In other words, by dividing the first-order difference by the time interval, the temperature change is converted into a thermal evolution rate, eliminating the influence of macroscopic periodic scales; then, by subtracting the median rate of the healthy phase, the contribution of the normal background warming rate is removed, finally yielding the relative thermal evolution rate that reflects only abnormal frictional heat generation. When equipment experiences early wear, increased friction leads to additional heat generation, causing the temperature to rise at a rate exceeding normal levels. When the equipment is under normal operating conditions, The value approaches zero when the equipment experiences a decrease in load, shutdown, or deceleration. This setup directly addresses the physical nature of the fault, and through median anti-interference design, it effectively suppresses the influence of environmental noise and short-term operating condition fluctuations, providing a clean abnormal thermal signal for subsequent multi-source feature fusion.

[0036] S3: Based on the local regression model of the working condition partition constructed in advance using health history data, the normal heat generation component corresponding to the current load is obtained. The difference between the relative thermal evolution rate and the normal heat generation component is calculated to obtain the pure thermal anomaly residual characterizing abnormal heat generation.

[0037] During normal equipment operation, dynamic changes in load power cause regular changes in the equipment's heat generation power, resulting in a significant correlation between the relative thermal evolution rate and load fluctuations. This type of temperature rise falls within the normal load heat generation range and is not abnormal heat generation caused by mechanical wear, component deterioration, or other faults. Directly eliminating normal thermal interference using a fixed, uniform benchmark cannot accommodate the differentiated normal temperature rise patterns under different loads, easily leading to missed early fault detection or misjudgment of operating condition fluctuations. Therefore, this invention constructs a local regression model for operating condition zones based on historical equipment health data. This model accurately fits the correspondence between power fluctuations and normal thermal evolution under different load ranges, achieving refined separation of the normal load heat generation component. The specific modeling process is as follows: First, historical power data and corresponding time-series relative thermal evolution rate data of the equipment under long-term healthy and fault-free operation are collected to construct a healthy operating condition modeling dataset. This dataset ensures that the dataset covers the entire load range of the equipment during normal operation, thereby ensuring the model's operating condition universality and fitting accuracy.

[0038] Secondly, the health modeling dataset is divided into operating condition ranges based on the equipment's rated power. The entire operating condition range is divided into three independent power ranges: light load, medium load, and heavy load, according to the load level. In this embodiment, the light load range corresponds to 0%~33% of the rated power, the medium load range to 34%~66% of the rated power, and the heavy load range to 67%~100% of the rated power. The boundary values ​​of these regions can be adjusted according to actual conditions to achieve partitioning for different load conditions. Since the heat generation characteristics and heat dissipation balance patterns of the equipment differ significantly in different power load ranges, partitioned modeling avoids the shortcomings of a single global model that cannot adapt to the nonlinear thermal characteristics of multiple operating conditions.

[0039] Subsequently, for each power range that has been divided, local linear regression training is carried out independently to construct a mapping relationship between the power change rate and the relative thermal evolution rate, which is used to characterize the normal temperature rise pattern corresponding to normal load fluctuations. The power change rate is calculated by taking the power difference between two adjacent cycles as the power change rate.

[0040] During training, the power change rate is used as the model input feature to characterize the dynamic fluctuation state of the current load, and the relative thermal evolution rate within the corresponding interval is used as the model output label to fit the normal temperature rise evolution law corresponding to load fluctuations. The regression coefficients and constant intercept terms corresponding to each power interval are solved by a linear regression algorithm to obtain a local thermal evolution fitting model specific to each independent load interval.

[0041] Finally, the local linear regression models trained in the power ranges of light load, medium load, and heavy load are integrated and spliced ​​to form a local regression model covering all operating conditions of the equipment. This local regression model can accurately characterize the normal thermal evolution benchmark of the equipment under different load dynamic fluctuations. It can adaptively match the corresponding normal heating fitting results according to real-time operating conditions, providing accurate model support for subsequently accurately separating the normal heating component caused by load fluctuations and extracting the pure thermal anomaly residuals.

[0042] S4: Perform local window adaptive normalization on the time-aligned vibration energy to obtain normalized vibration impact that characterizes the degree of impact anomaly. Multiply and fuse the normalized vibration impact with the pure thermal anomaly residual within a sliding window to obtain the final deviation feature used to judge the fault.

[0043] During actual equipment operation, dynamic load changes cause a shift in the overall amplitude of vibration energy. If a fixed benchmark is directly used to assess the vibration state, normal fluctuations under operating conditions may be misjudged as mechanical anomalies. Therefore, a local sliding window statistical method is employed to dynamically construct a normal vibration benchmark using recent short-term operating data of the equipment, enabling adaptive extraction of impact features under operating conditions. The specific process is as follows: Using the current macroeconomic cycle as the end point of the window, a sliding statistical window of a preset fixed length is extracted, and vibration scalar sequences corresponding to all cycles within the window are extracted. The arithmetic mean and standard deviation of the vibration scalar sequences within the window are calculated respectively, where the mean reflects the average level of vibration energy under the current working condition, and the standard deviation represents the normal fluctuation range of the vibration signal under the working condition.

[0044] The difference between the current periodic vibration energy and the mean vibration value within the window is calculated to obtain the deviation of the vibration energy relative to the normal level. This deviation can preliminarily distinguish the degree to which the vibration signal deviates from the normal state. Dividing the vibration energy deviation by the sum of the window standard deviation and a preset minimum constant yields a characteristic quantity representing the degree of mechanical impact anomaly. Adding a minimum constant can avoid calculation anomalies caused by the standard deviation approaching zero when the equipment is under light load and running smoothly, ensuring the algorithm operates stably under all working conditions.

[0045] Specifically, the normalized vibration and impact satisfy the following relationship: ; In the formula, This represents the normalized vibration and impact. This represents the energy of the current periodic vibration, i.e., the original impact intensity; Represents a local window The mean value of internal vibrational energy; Represents a local window Standard deviation of internal vibration energy, local window The length is set to 10 cycles to reflect the normal fluctuation range of vibration under the current load; Represents a minimal constant. To prevent numerical distortion caused by an excessively small standard deviation under light load conditions, if the current vibration energy is higher than the recent normal fluctuation range, Impact failures such as bearing spalling and gear cracks may occur; while vibration is normal. Approaching zero.

[0046] Traditional fixed global statistical benchmarks cannot adapt to equipment operating scenarios with varying loads and conditions. Load changes directly cause the overall amplitude of the vibration baseline to shift, easily leading to a large number of false alarms or missed alarms. In this embodiment, a short-term sliding window of 10 periods is used to statistically analyze the mean and standard deviation. A normal benchmark is constructed based solely on recent equipment operating data. This allows for real-time dynamic updates of the normal vibration threshold to follow load fluctuations, achieving adaptive operating conditions and adapting to the frequent changes in operating conditions of industrial equipment.

[0047] The core characteristics of early-stage impact failures in equipment are: a sudden increase in short-term vibration energy exceeding the normal fluctuation range of the current operating conditions, rather than a global amplitude exceeding the standard. The formula uses the current value and the window mean to remove the basic vibration component caused by the operating conditions, retaining only the abnormal deviation; then, through standard deviation normalization, it maps vibration data at different amplitude scales to a unified standard, achieving quantitative comparability of the degree of abnormality under different loads and equipment conditions, and accurately focusing on the impact anomalies caused by the failure.

[0048] By dynamically updating the vibration benchmark value in real time through a local sliding window, it can adaptively match complex scenarios such as load fluctuations and operating condition switching, accurately distinguish between normal fluctuations in operating conditions and actual equipment failure impacts, and significantly reduce false alarms caused by fixed thresholds.

[0049] Using the current macroeconomic cycle as the end point of the window, a sliding statistical window is formed by truncating a preset number of consecutive cycles. For each cycle within the window, the vibration and shock characteristics corresponding to the cycle are multiplied with the residual of the pure thermal anomaly to obtain the single-cycle fused characteristics.

[0050] The sum of all single-cycle fused features within the window is obtained by sequentially summing them. Then, the sum of window fused features is divided by the period length of the sliding window to complete the time-series smoothing process, finally obtaining the final deviation feature corresponding to the current period. The sliding window cumulative averaging can filter out instantaneous sampling noise and short-term operating condition disturbances, enhance the temporal continuity of fault features, and enable the output final deviation feature to objectively and stably reflect the comprehensive abnormality of the equipment simultaneously experiencing mechanical shock and abnormal heating.

[0051] Specifically, the final deviation feature satisfies the following relationship: ; In the formula, Indicates the first The final deviation characteristic of the cycle indicates that the larger the final deviation characteristic, the higher the probability that the equipment will simultaneously experience abnormal heating and abnormal impact. Indicates the local window length, used for temporal smoothing and to enhance the signal-to-noise ratio; Indicates the first Vibration and impact of one cycle, Indicates the first The pure thermal anomaly residual for each cycle.

[0052] It should be noted that early mechanical failures of equipment, including wear, jamming, and component damage, can all lead to increased mechanical shock and abnormal heat generation. If a single feature is present, it is easy to generate false alarms. For example, if only vibration is abnormal, it may be due to instantaneous operating disturbances; if only thermal is abnormal, it may be caused by cooling system or environmental factors. The occurrence of either of these alone is not considered a fault. Only when the residual values ​​of vibration shock and pure thermal anomaly are both positive abnormal values ​​will the product result increase significantly. If only one of these is abnormal or both are normal or negative, the product result will be suppressed, filtering out interference from a single scenario from a mechanistic perspective and avoiding false alarms.

[0053] In this embodiment, the judgment logic is a joint identification of faults from the same source: early mechanical faults of the equipment, such as wear, jamming, and component damage, will simultaneously cause increased mechanical impact and abnormal heat generation. A single feature will result in a false alarm, such as abnormal vibration only, which may be due to instantaneous operating disturbances; abnormal heat only, which may be caused by the cooling system or environmental factors. The presence of either alone will not be considered a fault. Only when... , The product result will only increase significantly when both are positive outliers. If only one item is outlier, or both items are normal or negative, the product result will be suppressed. When one item is negative, the product will be negative. The negative final deviation feature is directly judged as no fault, thus filtering out interference from a single scenario from a mechanistic perspective.

[0054] Leveraging the physical characteristics of simultaneous fault impact and heat generation, interference from single signal anomalies is shielded. Combined with sliding window cumulative averaging, sampling noise and instantaneous operating disturbances are filtered out, enhancing continuous fault characteristics. The changing trend can intuitively reflect the degree of equipment wear and deterioration, supporting both instantaneous anomaly alarms and fault trend prediction. This design logic aligns with equipment fault mechanisms, has low computational complexity, facilitates engineering deployment and real-time online operation, accurately identifies early-stage common-source faults, effectively reduces the false alarm rate of the monitoring system, and provides stable and highly quantifiable output characteristic signals, meeting the needs of both online real-time monitoring and fault trend analysis.

[0055] S5: Use the box plot method to set an adaptive alarm threshold for the final deviation feature. Based on the comparison between the final deviation feature and the adaptive alarm threshold, determine whether it is early wear of the equipment and issue an abnormal alarm.

[0056] First, all final deviation characteristic data of the equipment under long-term healthy and stable operating conditions are collected. Based on this set of healthy sample data, the upper quartile and interquartile range are calculated. Then, the adaptive alarm threshold is calculated by combining the preset empirical coefficients.

[0057] Specifically, the adaptive alarm threshold satisfies the following relationship: ; In the formula, This indicates the device's adaptive alarm threshold. It represents the upper quartile of the feature dataset under healthy operating conditions, and characterizes the upper limit of normal fluctuation of the feature value under healthy conditions when the equipment is operating normally; This is an engineering experience coefficient, commonly used in classic box plot anomaly detection. However, this value has a high false alarm rate in industrial safety scenarios. To further reduce the false alarm rate and ensure sufficient sensitivity to early, subtle anomalies, we take [value missing here]. You can choose according to the actual situation. This represents the interquartile range, which is obtained by the difference between the upper and lower quartiles.

[0058] It should be noted that existing static box plot thresholds generated based on fixed healthy samples are only suitable for brand-new, undegraded equipment and cannot adapt to the transitional operating conditions of gradually worn equipment. This invention sets up an adaptive threshold rolling update mechanism: a sliding time-series sample library continuously collects the operating characteristics of the equipment throughout its entire lifecycle. The sample library simultaneously includes steady-state healthy data and data from slight degradation transitional operating conditions. After each fixed period, the quartiles and interquartile ranges are recalculated based on all valid samples in the library, iteratively refreshing the alarm threshold. As the equipment slowly degrades, the proportion of degradation features in the sample library gradually increases, and the threshold rises synchronously and smoothly. This avoids frequent false alarms due to slight equipment wear while continuously capturing gradually worsening early fault characteristics, achieving dynamic threshold control throughout the entire lifecycle and overcoming the deficiency of a single static health benchmark in covering the degradation transition stage.

[0059] During online equipment monitoring, the final deviation characteristic calculated in real time is compared with the adaptive alarm threshold. When the final deviation characteristic is greater than the adaptive alarm threshold, it is determined that the equipment has experienced synchronous mechanical shock and abnormal heating, triggering an early wear fault alarm. When the final deviation characteristic is less than or equal to the adaptive alarm threshold, it indicates that all characteristics of the equipment are within the normal fluctuation range, and the equipment is determined to be in normal operating condition.

[0060] This invention also provides an intelligent factory safety monitoring system based on multi-source data fusion. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the intelligent factory safety monitoring method based on multi-source data fusion according to the first aspect of the present invention. The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. Their configuration and functions are known in the art and will not be described further here.

[0061] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A factory safety intelligent monitoring method based on multi-source data fusion, characterized in that, include: The system collects multi-source data, including vibration, power, and temperature data from the equipment. After filtering and preprocessing the multi-source data, it resamples the data to the same macroscopic period to form a scalar sequence. Based on the equipment's health history data, it determines the inherent thermal conduction hysteresis constant and completes the time alignment of the multi-source data. The time-aligned temperature data are sequentially optimized and its change features are extracted to obtain the thermal evolution rate. Based on the equipment health history data, health benchmark parameters are constructed. The thermal evolution rate is corrected using the health benchmark parameters to eliminate the influence of ambient temperature drift and normal operating condition fluctuations, and the relative thermal evolution rate is obtained. Based on the local regression model of the working condition partition constructed in advance using health history data, the normal heat generation component corresponding to the current load is obtained. The difference between the relative thermal evolution rate and the normal heat generation component is calculated to obtain the pure thermal anomaly residual characterizing abnormal heat generation. Local window adaptive normalization is performed on the time-aligned vibration energy to obtain normalized vibration impact that characterizes the degree of impact anomaly. The normalized vibration impact and the pure thermal anomaly residual are multiplied and fused within a sliding window to obtain the final deviation feature used to judge the fault. The box plot method is used to set an adaptive alarm threshold for the final deviation feature. Based on the comparison between the final deviation feature and the adaptive alarm threshold, it is determined whether it is early wear of the equipment and an abnormal alarm is triggered.

2. The intelligent factory safety monitoring method based on multi-source data fusion according to claim 1, characterized in that, The steps for forming the scalar sequence include: The vibration, power, and temperature data after filtering and preprocessing are segmented according to a preset macroscopic period duration. The root mean square value of the vibration data within the macroscopic period corresponding to each segment is used as the vibration energy scalar, the average active power is used as the power scalar, and the average temperature is used as the temperature scalar. These are then sorted sequentially according to time to obtain the corresponding vibration scalar sequence, power scalar sequence, and temperature scalar sequence. The length of the macroscopic period is selected based on the temperature response inertial characteristics.

3. The intelligent factory safety monitoring method based on multi-source data fusion according to claim 1, characterized in that, The relative thermal evolution rate is obtained as follows: Taking any period as the target period, calculate the average temperature data within the target period, and obtain the temperature change by the difference between the average temperature data within the previous period and the average temperature data within the target period. The time interval between the temperature change and the macroscopic period is taken as the thermal evolution rate of the target period. When the computing device is healthy and fault-free, the thermal evolution rate corresponding to all cycles is traversed, a set of thermal evolution rates is constructed, the median of the set of thermal evolution rates is obtained, and the difference between the thermal evolution rate of the target cycle and the median is taken as the relative thermal evolution rate.

4. The intelligent factory safety monitoring method based on multi-source data fusion according to claim 1, characterized in that, The steps for establishing the local regression model for the aforementioned operating condition zones include: Historical power data and relative thermal evolution rate data of the equipment under healthy and fault-free conditions are collected; the historical health data are divided into three power ranges: light load, medium load, and heavy load according to the rated power of the equipment; within each power range, the regression coefficients and intercept terms of the corresponding range are obtained by linear regression training with the power change rate as input and the relative thermal evolution rate as output; the linear regression models obtained by training each range independently are combined to form a local regression model of the operating condition zone applicable to all operating conditions.

5. The intelligent factory safety monitoring method based on multi-source data fusion according to claim 1, characterized in that, The normalized vibration and impact are obtained as follows: Using the current period as the end point of the window, a sliding statistical window of a preset length is constructed. Vibration scalar sequences of all periods within the window are extracted, and the mean and standard deviation of the vibration scalar sequences within the window are calculated respectively. The difference between the current vibration energy and the mean of the vibration scalar sequence within the window is calculated to obtain the vibration energy offset. The vibration energy offset is then divided by the sum of the standard deviation and the minimum constant within the window to obtain the normalized vibration impact.

6. The intelligent factory safety monitoring method based on multi-source data fusion according to claim 1, characterized in that, The final deviation feature is calculated as follows: Using the current period as the end point of the window, a preset number of consecutive periods are obtained. Within each period corresponding to the sequence number of the window, the vibration impact and the residual of the pure thermal anomaly are multiplied to obtain the single-period fusion feature. All single-period fusion features within the window are accumulated to obtain the sum of the fusion features within the window. Dividing the sum of the fused features by the window length yields the final deviation features for the current period.

7. The intelligent factory safety monitoring method based on multi-source data fusion according to claim 1, characterized in that, The method of using box plots to set adaptive alarm thresholds for the final deviation features includes: Based on the final deviation characteristic data under the healthy operating conditions of the equipment, the upper quartile and interquartile range are calculated, and the adaptive alarm threshold is calculated by combining the preset empirical coefficients. When the final deviation characteristic is greater than the adaptive alarm threshold, the early wear alarm of the equipment is triggered. Conversely, when it is less than or equal to the adaptive alarm threshold, the equipment is in normal operating condition.

8. A factory safety intelligent monitoring system based on multi-source data fusion, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent factory safety monitoring method based on multi-source data fusion according to any one of claims 1-7.

Images