Shared tray sensor fault detection method based on multi-level fusion diagnosis model

Abstract

The application discloses a shared tray sensor fault detection method based on a multi-level fusion diagnosis model and relates to the technical field of calculation models. The shared tray sensor fault detection method based on the multi-level fusion diagnosis model comprises the following steps: S1, collecting sensor data and obtaining power supply link working electrical parameters, and performing time alignment, space registration and normalization processing; S2, constructing a power supply disturbance event window, and evaluating power supply induced coupling strength; S3, constructing a fusion diagnosis model, performing layered judgment, and judging a fault abnormal state; and S4, constructing a layered priority decision mechanism, and performing optimal evaluation on fault recovery. The application effectively improves the accuracy of shared tray sensor fault positioning and the system self-recovery capability, and solves the problem that the fault source is difficult to judge due to the coupling of power supply disturbance and sensor output abnormality in the prior art.

Description

Technical Field

[0001] This invention relates to the field of computational modeling technology, specifically to a method for fault detection of shared tray sensors based on a multi-level fusion diagnostic model. Background Technology

[0002] With the development of IoT technology, edge computing, and multi-source sensing data fusion, most existing technologies only perform simple anomaly detection on sensor output values, such as out-of-bounds errors, fluctuations, or packet loss, without providing a unified diagnosis of power supply disturbances. In existing technologies, current systems cannot distinguish between power supply link problems and sensor degradation when anomalies occur, resulting in non-targeted maintenance operations, low first-time repair rates, and high rates of incorrect component replacement. Furthermore, the reliability and stability of edge-side alarms are insufficient.

[0003] For example, the invention patent with publication number CN106447040B discloses a method for assessing the health status of mechanical equipment based on heterogeneous multi-sensor data fusion, including the following steps: First, the signals from heterogeneous multi-sensor sensors are denoised separately; then, sensitive features of various signals are extracted, a set of sensitive features is constructed, and the set of sensitive features is used as training samples for a BP neural network to establish a fault diagnosis model based on the BP neural network to achieve fault separation and diagnosis. The recognition rate of each sensor for each type of fault is obtained, constructing a DS evidence framework. Finally, the Dempster synthesis formula of DS evidence theory is used to solve for the most probable fault type. This invention, based on the fusion of DS decision layers, achieves high accuracy and efficiency in classification diagnosis, effectively improving the diagnostic effect of cracks and facilitating its use in engineering practice.

[0004] In existing technologies, when existing systems utilize battery power, low-power wake-up, and high-frequency communication simultaneously, voltage drops, instantaneous voltage drops, increased ripple, and partial power outages are prone to occur. These power supply disturbances can induce output anomalies such as sampling offsets, noise increases, short-term distortions, and communication interruptions in temperature and humidity, positioning, and data acquisition links. Without unified time alignment and correlation analysis between power supply characteristics and output anomalies, it is difficult to distinguish whether the anomaly is due to a power supply link problem or sensor degradation and malfunction, and causality may even be reversed.

[0005] Therefore, in order to address the above problems, there is an urgent need for a shared tray sensor fault detection method based on a multi-level fusion diagnostic model. Summary of the Invention

[0006] Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a shared tray sensor fault detection method based on a multi-level fusion diagnostic model, which solves the problem that it is difficult to determine the source of faults caused by abnormal coupling between power supply disturbances and sensor output in existing methods.

[0008] Technical solution

[0009] To achieve the above objectives, the present invention provides the following technical solution: a shared tray sensor fault detection method based on a multi-level fusion diagnostic model, comprising: S1, collecting sensor data and obtaining power supply link operating electrical parameters, and performing time alignment, spatial registration, and normalization processing to construct a joint monitoring dataset; S2, identifying power supply disturbance events based on the multi-source sensing dataset, constructing a power supply disturbance event window, evaluating the power supply-induced coupling strength, and constructing a power supply-induced anomaly sample set; S3, constructing a fusion diagnostic model based on the power supply disturbance event window, performing hierarchical judgment, determining the fault anomaly state, and generating fault source labels; S4, constructing a hierarchical priority decision-making mechanism based on the fault source labels, performing optimal evaluation of fault recovery, and generating a shared tray sensor fault detection report.

[0010] Further, the specific steps for collecting sensor data and acquiring power supply link operating parameters, and performing time alignment, spatial registration, and normalization processing to construct a joint monitoring dataset are as follows: Data from the built-in sensors of the shared pallet is collected throughout the entire process of warehousing, loading / unloading, transportation, and turnover scenarios. Sensor data includes temperature, humidity, acceleration, angular velocity, BeiDou positioning data, vibration intensity, and communication status. Communication status parameters include signal reception strength, packet loss rate, retransmission count, and connection status. Simultaneously, power supply link operating parameters are acquired, including battery voltage, operating current of each sensor, instantaneous voltage drop amplitude, power supply ripple amplitude, power recovery time upon wake-up, peak current pulse value and pulse width during communication transmission. Missing value detection is performed on the sensor data, using time interval criteria and frame... The system employs a dual-criteria system for serial number continuity; it completes missing data completion and labeling through classification interpolation; it removes abnormal peaks in sensor data by verifying the physical correlation of neighboring areas using a sliding window; it performs denoising based on a hierarchical denoising method adapted to different sensor types; and it corrects sampling interval jitter through linear interpolation resampling. The system also performs ripple filtering and baseline drift correction on the power supply link operating parameters, and completes event edge marking and window isolation for transient disturbances during the wake-up switching phase. For sensor data and power supply link operating parameters with different sampling periods, it uses linear interpolation to achieve unified time alignment, and performs spatial registration based on the actual installation location of the sensors and the power supply branch affiliation. Finally, it performs amplitude normalization on all processed sensor data and power supply link operating parameters, integrating them to construct a joint monitoring dataset.

[0011] Furthermore, the specific steps for identifying power supply disturbance events and constructing a power supply disturbance event window based on a multi-source sensing dataset are as follows: Based on the power supply link operating electrical parameter sequence in the joint monitoring dataset, the power supply disturbance event type is obtained by comparing the real-time monitoring power supply disturbance index with the trigger threshold. The power supply disturbance event types include voltage drop events, voltage drop events, ripple events, partial power outage events, power consumption wake-up switching events, and communication load pulse events. The power supply disturbance index includes battery voltage, instantaneous voltage drop amplitude, power supply ripple amplitude, and power supply ripple amplitude change. For each identified power supply disturbance event, a sliding window monitoring is used. When the power supply disturbance index first exceeds the threshold and remains continuously for N unified time slices, the first time slice of the N time slices is determined as the event start time. An extreme value search algorithm is used to find the time corresponding to the maximum deviation of the power supply disturbance index within the event window to obtain the event peak time. When the power supply disturbance index falls back below the threshold and continuously stabilizes within the threshold to reach a stable duration threshold, the event recovery time is obtained. A power supply disturbance event window is constructed based on the event start time, event peak time, and event recovery time.

[0012] Further, the specific steps for evaluating the power supply induced coupling strength are as follows: the current pulse interruption trigger probability is obtained by statistically analyzing the proportion of packet loss or interruption triggered by each communication pulse within the event window to the total number of pulses; the event recovery hysteresis is obtained by the time difference between the power supply disturbance recovery time and the sensor output recovery time; the voltage drop term is obtained by adding 1 to the ratio of the instantaneous voltage drop amplitude to the battery voltage within the power supply disturbance event window and then performing an exponential operation on the instantaneous voltage drop enhancement coefficient; the voltage drop term is obtained by adding 1 to the ratio of the power supply ripple amplitude change to the power supply ripple amplitude. The ripple term is obtained by performing an exponential operation on the power supply ripple enhancement coefficient after adding 1 to the current pulse interruption trigger probability and then performing an exponential operation on the pulse trigger enhancement coefficient. The adjustment term is obtained by multiplying the anomaly coefficient and the adjustment factor and then performing an exponential operation on the natural constant e. The recovery term is obtained by adding 1 to the ratio of the event recovery hysteresis difference to the event recovery time and then performing an exponential operation on the recovery hysteresis suppression coefficient. The power supply induced coupling strength evaluation value is obtained by multiplying the voltage drop contribution term, ripple contribution term, pulse contribution term, and adjustment term and then dividing by the recovery term.

[0013] Furthermore, the specific steps for constructing the power supply-induced anomaly sample set are as follows: compare the power supply-induced coupling strength evaluation value with the coupling judgment threshold in real time. When the power supply-induced coupling strength evaluation value is greater than the coupling judgment threshold, it is determined that there is a coupling relationship between the sensor output anomaly and the power supply disturbance in the current time slice. When the power supply-induced coupling strength evaluation value is less than or equal to the coupling judgment threshold, it is determined that there is no coupling relationship between the sensor output anomaly and the power supply disturbance in the current time slice.

[0014] Furthermore, the specific steps for constructing the fusion diagnostic model based on the power supply disturbance event window are as follows: Based on the power supply disturbance event window, a fusion diagnostic model is constructed, which includes a power supply link screening layer, a sensor body fault determination layer, and a composite anomaly conflict resolution layer. In the power supply link screening layer, the battery voltage drop amplitude, module operating current deviation, instantaneous voltage drop amplitude, power supply ripple amplitude, wake-up recovery delay, and power supply induced coupling strength evaluation value from the power supply disturbance event window are extracted and normalized respectively. The normalized value of each feature is incremented by 1, and the natural logarithm is taken. The logarithmic values ​​of all features are accumulated. The accumulated result is divided by the number of features and subjected to an exponential transformation to obtain the power supply fault support. In the sensor body fault determination layer, the sensor output drift, noise, and other parameters are extracted. The sensor fault support is constructed using a fuzzy comprehensive evaluation method, taking into account the acoustic energy amplification, the root mean square value of the associated residual sequence, the zero-value retention time, and the number of times the range exceeds the limit. In the composite anomaly conflict resolution layer, a basic probability assignment function is constructed based on the power supply fault support and the sensor fault support, generating power supply fault evidence bodies and sensor fault evidence bodies respectively. Each evidence body assigns probability mass to three types of propositions: power supply fault, sensor fault, and composite anomaly, forming a three-dimensional support vector. The degree of overlap is obtained by calculating the cosine similarity between the two evidence body vectors, and then fused using Dempster-Shafer evidence theory to construct the composite anomaly support. The evidence conflict degree is obtained by multiplying the basic probability assignments of all proposition pairs with no intersection in the two evidence bodies and summing them.

[0015] Further, the specific steps for stratification are as follows: the power supply contribution value obtained by exponential mapping of the power supply link fault support, the sensor contribution value obtained by exponential mapping of the sensor body fault support, and the composite anomaly support are all obtained by exponential mapping using the Sigmoid function, and the resulting composite contribution value is used as three basic contribution items; the maximum value of the three basic contribution items is used as the numerator, and the exponentially mapped values ​​of the three basic contribution items are added together, and the evidence conflict degree is added as the denominator, and a ratio operation is performed to obtain the correction item; the negative value of the degradation enhancement factor is subjected to negative exponential operation, and one is added and the reciprocal is taken to obtain the degradation effect item; the correction item and the degradation effect item are multiplied together to obtain the stratification determination value.

[0016] Further, the specific steps for determining the fault abnormality state and generating fault source labels are as follows: Real-time comparison of the stratified judgment value with the fault source threshold; if the stratified judgment value is less than or equal to the fault source threshold, the current abnormality is determined to be a controllable abnormality; if the stratified judgment value is greater than the fault source threshold, the fault source sub-classification process begins; if the power supply link abnormality support is the highest among the three support values, it is determined to be a power supply link fault state; if the sensor body fault support is the highest among the three support values, it is determined to be a sensor body fault state; if the composite abnormality support is the highest among the three support values, it is determined to be a composite abnormal state; generate fault source labels, and simultaneously set a verification flag for sensors identified as abnormal, and label the corresponding fault type.

[0017] Furthermore, the specific steps for constructing a hierarchical priority decision-making mechanism based on fault source labels are as follows: extract the generated fault source labels, construct a hierarchical priority decision-making mechanism, and select a recovery path based on power supply link, sensor body soft fault, and sensor body hard fault: when the diagnosis result is a power supply link fault, enter the power supply side handling process; when the diagnosis result is a sensor body soft fault, enter the online self-calibration process; when the diagnosis result is a sensor body hard fault, enter the data reconstruction and degradation operation process; when the diagnosis result is a composite anomaly, enter the composite handling process.

[0018] Furthermore, the specific steps for optimizing fault recovery and generating a shared tray sensor fault detection report are as follows: The fault suppression gain value is obtained by calculating the difference in output anomaly scores after the execution of the current candidate recovery action; the energy consumption value is obtained by calculating the increased sensor sampling power consumption, communication power consumption, and processor load power consumption during the recovery action execution; the improvement is obtained by statistically analyzing the increase in sampling period before and after the recovery action execution; the edge computing load is obtained by calculating the increase in CPU utilization during the recovery action execution and normalizing it by dividing it by the rated CPU utilization; the success probability is obtained by statistically analyzing the proportion of successful recovery action executions to the total number of attempts; and the additional communication overhead is obtained by statistically analyzing the amount of new data packets or the communication bandwidth occupancy ratio during the recovery action. The fault suppression gain value is incremented by 1 to obtain the fault suppression item. The following steps are performed: First, add 1 to the boost amount to obtain the additional item; add 1 to the success probability to obtain the probability item; add 1 to the energy consumption value to obtain the energy consumption item; add 1 to the edge computing load to obtain the edge load item; add 1 to the communication overhead to obtain the communication load item; multiply the fault suppression item, the additional item, and the probability item to obtain the performance gain item; multiply the energy consumption item, the edge load item, and the communication load item to obtain the resource overhead item; divide the performance gain item by the resource overhead item to obtain the preferred evaluation value; compare the preferred evaluation value with the recovery trigger threshold in real time; if the preferred evaluation value is greater than the recovery trigger threshold, execute the current candidate recovery action; if the preferred evaluation value of the recovery strategy is less than or equal to the recovery trigger threshold, do not execute the recovery operation; structurally bind the recovery result, fault source category, handling action, and review status after execution to generate a shared tray sensor fault detection report.

[0019] Beneficial effects

[0020] The present invention has the following beneficial effects:

[0021] (1) This invention constructs a unified multi-source sensing dataset by jointly collecting the output of multi-source sensors on a shared tray and the working electrical parameters of the power supply link. This achieves unified time alignment and spatial registration of power supply disturbances and sensor output anomalies, improving data traceability and consistency. It also achieves effective decoupling between power supply anomalies and sensor anomalies, improving the accuracy of fault source determination.

[0022] (2) This invention, by constructing a hierarchical fusion diagnostic model based on a power supply link screening layer, a sensor body fault determination layer, and a composite anomaly conflict resolution layer, realizes hierarchical determination of anomaly causality and dynamic priority screening, thereby improving the accuracy of edge-side fault location. It also realizes quantitative evaluation of multi-dimensional feature fusion, providing quantifiable indicators for hierarchical determination.

[0023] (3) This invention achieves closed-loop self-recovery of sensor soft faults by combining online self-calibration process with virtual reference values ​​and real-time correlation residual verification, thereby improving the system's adaptive repair capability and reducing manual intervention. It also achieves continuous data approximate output under sensor hard faults, ensuring the system's critical monitoring capability in the event of partial sensor failure.

[0024] (4) This invention achieves intelligent triggering and adaptive delayed execution of recovery actions by using power supply-induced coupling strength, recovery strategy optimization evaluation value, and threshold closed-loop judgment, avoiding blind recovery and improving recovery efficiency and system stability. It realizes accurate location, consistent alarm, and optimized maintenance handling of sensor faults under low power consumption conditions of shared tray, significantly improving overall operation and maintenance efficiency and first-time repair rate.

[0025] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0026] Figure 1 This invention relates to a shared tray sensor fault detection method based on a multi-level fusion diagnostic model.

[0027] Figure 2 This is a voltage waveform diagram for detecting power supply disturbance events in this invention;

[0028] Figure 3 This is an abnormal coupling analysis diagram of the present invention;

[0029] Figure 4 This is a support graph for each fault type in this invention;

[0030] Figure 5 This is a preferred evaluation value region map for the recovery strategy of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-5This invention provides a technical solution: a shared tray sensor fault detection method based on a multi-level fusion diagnostic model, comprising: S1, collecting sensor data and obtaining power supply link operating electrical parameters, and performing time alignment, spatial registration, and normalization processing to construct a joint monitoring dataset; S2, identifying power supply disturbance events based on the multi-source sensing dataset, constructing a power supply disturbance event window, evaluating the power supply-induced coupling strength, and constructing a power supply-induced anomaly sample set; S3, constructing a fusion diagnostic model based on the power supply disturbance event window, performing hierarchical judgment, determining the fault abnormal state, and generating fault source labels; S4, constructing a hierarchical priority decision-making mechanism based on the fault source labels, performing optimal evaluation of fault recovery, and generating a shared tray sensor fault detection report.

[0033] Specifically, the steps for collecting sensor data and acquiring power supply link operating parameters, and constructing a joint monitoring dataset through time alignment, spatial registration, and normalization are as follows: Data from the built-in sensors of the shared pallet is collected throughout the entire process of warehousing, loading / unloading, transportation, and turnover scenarios. Sensor data includes temperature, humidity, acceleration, angular velocity, BeiDou positioning data, vibration intensity, and communication status. Signal reception strength is obtained from the instantaneous value reported in real-time by the communication module; packet loss rate and retransmission count are calculated based on a sliding time window; connection status is indicated by online and offline Boolean indicators. Communication status parameters include signal reception strength, packet loss rate, retransmission count, and connection status. The power supply link operating parameters are simultaneously acquired. The power supply link's operating electrical parameters include battery voltage, operating current of each sensor, instantaneous voltage drop amplitude, power supply ripple amplitude, power recovery time during wake-up, and peak value and pulse width of the current pulse during the communication transmission phase. Battery voltage and operating current are acquired through the channel ADC; instantaneous voltage drop amplitude is obtained by detecting the falling edge of the voltage waveform and calculating the maximum drop per unit time; power supply ripple amplitude is calculated by extracting the ripple component through a bandpass filter and then calculating the peak-to-peak value; power recovery time during wake-up is obtained by recording the time difference between the wake-up signal edge and the voltage stabilization threshold; peak value and pulse width of the current pulse during the communication transmission phase are captured by a current detection amplifier in conjunction with the ADC, and the pulse width is measured by a hardware counter.

[0034] Missing value detection is performed on sensor data using a dual criterion of time interval and frame sequence number continuity. The time interval criterion determines the maximum allowable interval based on the sampling period, while the frame sequence number continuity criterion identifies missing frames by checking the ascending order of frame numbers. Missing data is filled and marked using categorical interpolation. The categorical interpolation employs linear interpolation, spline interpolation, or nearest neighbor interpolation depending on the sensor type, while simultaneously generating missing value markers. A sliding window with a fixed-length neighborhood physical correlation verification is used to remove abnormal spikes in the sensor data. The neighborhood physical correlation verification uses constraints between sensors of the same type or related physical quantities for consistency checks. Denoising is performed using a layered denoising method with sensor type-specific adaptive filtering. Layered denoising includes a first layer of median filtering to remove isolated outliers and a second layer of low-pass filtering to smooth high-frequency peaks. For dynamic data, frequency noise and third-layer Kalman filtering are applied, and sampling interval jitter correction is achieved through linear interpolation resampling. Ripple filtering uses a band-stop filter or moving average method, and baseline drift correction removes trend terms through high-pass filtering. Ripple filtering and baseline drift correction are applied to the power supply link operating parameters, and event edge marking and window isolation are performed for transient disturbances during the wake-up switching phase. For sensor data and power supply link operating parameters with different sampling periods, time alignment is uniformly achieved using linear interpolation. At the same time, spatial registration is performed according to the actual installation location of the sensor and the power supply branch affiliation. Spatial registration establishes a mapping matrix from sensor data to the physical location of the tray using calibrated installation coordinates. All processed sensor data and power supply link operating parameters are normalized in amplitude and integrated to construct a joint monitoring dataset.

[0035] In this implementation plan, by comprehensively collecting sensor data and power supply link operating parameters of shared pallets in various logistics scenarios, filling in missing values, removing anomalies, performing layered noise reduction, ripple filtering, and baseline drift correction, and combining unified time alignment and spatial registration processing, a unified amplitude normalized joint monitoring dataset is constructed. This achieves accurate synchronization and fusion of multi-source heterogeneous data, providing a high-quality, traceable, and standardized data foundation for fault diagnosis, anomaly detection, and energy consumption analysis, and effectively improving the integrity, reliability, and real-time analysis capabilities of monitoring data.

[0036] Specifically, the steps for identifying power supply disturbance events and constructing a power supply disturbance event window based on a multi-source sensing dataset are as follows: Based on the power supply link operating electrical parameter sequence in the joint monitoring dataset, the power supply disturbance event type is obtained by comparing the real-time monitoring power supply disturbance index with the trigger threshold. The power supply disturbance event types include voltage drop events, voltage drop events, ripple events, partial power outage events, power consumption wake-up switching events, and communication load pulse events; the power supply disturbance index includes battery voltage, instantaneous voltage drop amplitude, power supply ripple amplitude, and power supply ripple amplitude change; when the battery voltage shows a downward trend within N consecutive uniform time slices and the voltage drop amplitude exceeds the voltage descent threshold, and the duration of the downward state exceeds the continuous If the duration threshold is exceeded, it is identified as a voltage drop event; if the instantaneous voltage drop amplitude exceeds the instantaneous voltage drop threshold, it is identified as a voltage drop event; if the power supply ripple amplitude exceeds the ripple threshold and the duration exceeds the duration threshold, it is identified as a ripple event; if the battery voltage is lower than the device's minimum operating voltage threshold and the duration exceeds the duration threshold, it is identified as a partial power failure event; if the detected operating current spike exceeds the wake-up current spike threshold and the detected battery voltage drop exceeds the wake-up voltage drop threshold, it is identified as a power consumption wake-up switching event; if it is in the communication transmission phase and the detected current pulse peak value exceeds the pulse peak value range and the pulse width exceeds the pulse width range, it is identified as a communication load pulse event.

[0037] For each identified power supply disturbance event, a sliding window monitoring method is used. When the power supply disturbance index first exceeds the threshold and remains there for N consecutive time slices, the first time slice of the N time slices is determined as the start time of the event. N is obtained by dividing the dejittering time by the period of the unified time slice and rounding up. The dejittering time is used to eliminate instantaneous noise interference and ensure the reliability of event triggering. An extreme value search algorithm is used to find the moment corresponding to the maximum deviation of the power supply disturbance index within the event window to obtain the event peak moment. When the power supply disturbance index falls back below the threshold and remains stable within the threshold for a stable duration, the event recovery moment is obtained. A power supply disturbance event window is constructed based on the event start time, the event peak moment, and the event recovery moment.

[0038] like Figure 2The voltage waveform diagram for power supply disturbance events shown illustrates the voltage waveform of the shared tray power supply link and the detection results of various power supply disturbance events. The horizontal axis represents time, and the vertical axis represents voltage value. The blue curve represents the real-time battery voltage change, and the red dashed line represents the minimum operating voltage threshold. The diagram uses different colors to label various types of power supply disturbance events, including voltage drop events, voltage decrease events, partial power failure events, ripple events, function wake-up switching events, and communication load pulse events. As can be seen from the diagram, voltage drop events occur multiple times, with the voltage dropping briefly but still within the device's allowable range, showing the impact of instantaneous power supply fluctuations on the system. Voltage decrease events occur when the voltage is below the safety threshold, indicating a possible triggering of power supply side protection or causing abnormal sensor output. Partial power failure events are characterized by a sudden drop in voltage to a significantly low value that lasts for a certain period, showing the characteristics of a partial power supply interruption. Ripple events are indicated by rapid voltage fluctuations, reflecting a decrease in power supply stability and potential interference with high-frequency communication and sensitive sensor output. Function wake-up switching events and communication load pulse events are represented by short-term peaks and drops in the voltage curve, reflecting the transient impact of low-power wake-up and high-frequency communication loads on the voltage. Overall, it not only reflects the real-time fluctuations of battery voltage, but also intuitively presents the timing and intensity of different types of power supply disturbance events through event markers, providing basic data support for the construction of power supply disturbance event windows and the decoupling analysis of sensor output anomalies.

[0039] In this implementation plan, the method effectively improves the reliability and accuracy of power supply disturbance event detection, providing a highly reliable data foundation for subsequent sensor output anomaly decoupling, fault cause tracing, and system steady-state protection. It also supports multi-event overlay analysis and historical event archiving, providing technical support for safety monitoring and intelligent operation and maintenance under low-power supply conditions of shared trays.

[0040] Specifically, the steps for evaluating the power supply-induced coupling strength are as follows: the probability of current pulse interruption is obtained by statistically analyzing the proportion of packet loss or interruption triggered by each communication pulse within the event window to the total number of pulses; the event recovery hysteresis is obtained by the time difference between the power supply disturbance recovery time and the sensor output recovery time; the statistical analysis of the current pulse interruption trigger probability uses a sliding time window, with the window length dynamically set according to the communication frequency, and a minimum sample size threshold of 30 pulses. When the pulse sample size within the window is lower than the threshold, the mean is used as a substitute, and an exponentially weighted moving average is used for smoothing; the sensor output recovery time in the event recovery hysteresis is defined as the time point when the output abnormal characteristics fall back to the baseline range and remain stable for a continuous duration threshold.

[0041] The voltage drop term is obtained by adding 1 to the ratio of the instantaneous voltage drop amplitude to the battery voltage within the power supply disturbance event window and then performing an exponential operation on the instantaneous voltage drop enhancement coefficient. The ripple term is obtained by adding 1 to the ratio of the change in power supply ripple amplitude to the power supply ripple amplitude and then performing an exponential operation on the power supply ripple enhancement coefficient. The pulse term is obtained by adding 1 to the current pulse interruption trigger probability and then performing an exponential operation on the pulse trigger enhancement coefficient. The regulation term is obtained by multiplying the anomaly coefficient and the adjustment factor and then performing an exponential operation on the natural constant e. The recovery term is obtained by adding 1 to the ratio of the event recovery hysteresis difference to the event recovery time and then performing an exponential operation on the recovery hysteresis suppression coefficient. The power supply induced coupling strength evaluation value is obtained by multiplying the voltage drop contribution term, ripple contribution term, pulse contribution term, and regulation term and then dividing by the recovery term.

[0042] The specific formula for calculating the power supply induced coupling strength assessment value is as follows:

[0043] ;

[0044] In the formula, This indicates the intensity at which power supply disturbances induce abnormal sensor output. It represents the instantaneous voltage drop amplitude, used to characterize the voltage disturbance amplitude of the power supply link; This represents the battery voltage and is used to normalize the instantaneous voltage drop. It represents the change in the amplitude of the power supply ripple, used to reflect fluctuations in the power supply output; This represents the power supply ripple amplitude and is used to standardize the variation in the power supply ripple amplitude. This indicates the probability of a current pulse interruption, used to quantify the probability of a communication pulse inducing an output anomaly. The anomaly coefficient is calculated by measuring the Pearson correlation coefficient between sensor data and power supply link operating parameters within the event window. The value ranges from -1 to 1 and is used to reflect the degree of linear and nonlinear correlation between power supply disturbances and output anomalies. This represents the adjustment factor, obtained through the sensitivity analysis of the correlation between power supply and output within the event window. Its value ranges from 0 to 10 and is used to control the amplification or reduction effect of the correlation coefficient on the coupling strength assessment. This indicates the event recovery lag difference, used to suppress spurious coupling caused by output recovery lag; Indicates the event recovery time, used to standardize the recovery lag; This represents the instantaneous voltage drop enhancement coefficient, obtained through event data statistical analysis. Its value ranges from 0 to 5, and it is used to increase or decrease the weight of voltage drop in the coupling assessment. This represents the power supply ripple enhancement coefficient, which is obtained through event data playback analysis. Its value ranges from 0 to 5 and is used to control the amplification effect of ripple on the determination of power supply-induced anomalies. This represents the pulse trigger enhancement coefficient, obtained through statistical analysis of the probability of communication pulse triggering anomalies. Its value ranges from 0 to 5, and it is used to reflect the importance of communication pulse triggering anomalies. This represents the recovery hysteresis suppression coefficient, which is obtained by a sensitivity analysis method that measures the effect of output recovery hysteresis on coupling strength. The value ranges from 0 to 5 and is used to reduce the spurious amplification effect of output recovery hysteresis on coupling strength.

[0045] In this implementation scheme, by uniformly quantifying the intensity of disturbances such as voltage drop, ripple and communication pulses within the power supply disturbance window, and introducing recovery hysteresis suppression and correlation adjustment mechanisms, measurable discrimination and graded evaluation of whether power supply disturbances induce sensor output anomalies are achieved. This avoids misjudging occasional noise, short-time hysteresis or small sample pulses as strong coupling, and improves the accuracy of anomaly diagnosis and the reliability of handling of shared trays in scenarios where low-power power supply and high-frequency communication coexist.

[0046] Specifically, the steps for constructing the power supply-induced anomaly sample set are as follows: The power supply-induced coupling strength assessment value is compared with the coupling judgment threshold in real time. When the power supply-induced coupling strength assessment value is greater than the coupling judgment threshold, it is determined that the sensor output anomaly in the current time slice is coupled with the power supply disturbance. The current time slice is marked as a power supply-induced anomaly time slice and written into the power supply-induced mark field. Starting from the current time slice, power supply disturbance event windows are continuously recorded within a fixed number of time slices to construct the power supply-induced anomaly sample set. When the power supply-induced coupling strength assessment value is less than or equal to the coupling judgment threshold, it is determined that the sensor output anomaly in the current time slice is not coupled with the power supply disturbance. The current time slice is marked as a non-power supply independent anomaly candidate time slice, and the writing of the time slice into the power supply-induced anomaly sample set is paused until the next time slice is reassessed and judged again.

[0047] like Figure 3The anomaly coupling analysis diagram shown is used to visually demonstrate the impact and coupling degree of power supply disturbances on sensor output anomalies. The first sub-plot is the power supply disturbance intensity curve, with the horizontal axis representing time slices and the vertical axis representing power supply disturbance intensity. The blue curve represents the instantaneous disturbance intensity of the battery and power supply link, and the fluctuation characteristics of power supply disturbances caused by events such as low-power wake-up, voltage drop, and ripple can be observed. The second sub-plot is the sensor anomaly intensity curve, with the horizontal axis representing time slices and the vertical axis representing sensor anomaly intensity. The green curve represents the amplitude of sensor output anomalies in the corresponding time slice, showing the abnormal response of different sensors under power supply disturbance events. The third sub-plot is the power supply-induced coupling strength analysis diagram, with the horizontal axis representing time slices and the vertical axis representing coupling strength. The red curve represents the degree of coupling of power supply disturbances to sensor anomalies. The background color indicates independent anomalies (yellow), coupled anomalies (red), and fusion judgment thresholds (orange dashed lines), which can visually demonstrate the inducing effect of various power supply disturbance events on sensor output anomalies in different time slices. Overall, it provides a visual basis for analyzing power supply anomalies-induced sensor anomalies and hierarchical judgment.

[0048] In this implementation scheme, the correlation between power supply disturbances and sensor output anomalies is transformed into landable time slice annotation and sample set construction rules by online discrimination based on coupling strength evaluation value and judgment threshold. This can stably and continuously accumulate control samples of power supply-induced anomalies and non-power supply independent anomalies, avoid mislearning and misdiagnosis caused by mixed abnormal samples, and improve the reliability of power supply anomaly induction judgment and the pertinence of handling strategies.

[0049] Specifically, the steps for constructing a fusion diagnostic model based on the power supply disturbance event window are as follows: A fusion diagnostic model is constructed based on the power supply disturbance event window. The fusion diagnostic model includes a power supply link screening layer, a sensor body fault determination layer, and a composite anomaly conflict resolution layer. In the power supply link screening layer, the battery voltage drop amplitude, module operating current deviation, instantaneous voltage drop amplitude, power supply ripple amplitude, wake-up recovery delay, and power supply induced coupling strength evaluation value from the power supply disturbance event window are extracted and normalized respectively. The normalized value of each feature is incremented by 1 and the natural logarithm is taken. The logarithms of all features are accumulated. The accumulated result is divided by the number of features and subjected to exponential transformation to obtain the power supply fault support. In the sensor body fault determination layer, the sensor output drift, noise energy increase, root mean square value of the associated residual sequence, zero-value holding time, and number of range exceedances are extracted and fuzzy logic is used. A comprehensive evaluation method is used to construct sensor fault support. In the compound anomaly conflict resolution layer, basic probability assignment functions are constructed based on power supply fault support and sensor fault support, generating power supply fault evidence bodies and sensor fault evidence bodies respectively. Each evidence body assigns probability mass to three types of propositions: power supply fault, sensor fault, and compound anomaly, forming a three-dimensional support vector. That is, each evidence body outputs a triple, corresponding to power supply fault support, sensor fault support, and compound anomaly support respectively. The degree of overlap is obtained by calculating the cosine similarity between the vectors of two evidence bodies. The closer the cosine similarity is to 1, the more consistent the judgments of the two evidence bodies regarding the source of the fault are. Dempster-Shafer evidence theory is used for fusion to construct compound anomaly support. The evidence conflict degree is obtained by multiplying the basic probability assignments of all proposition pairs with no intersection in the two evidence bodies and then summing them. This method can effectively handle the evidence contradiction problem when power supply anomalies and sensor degradation coexist, improving the accuracy and robustness of compound anomaly judgment.

[0050] In this implementation plan, by performing hierarchical modeling and fusion judgment of the power supply disturbance characteristics, sensor degradation characteristics and the evidence conflict relationship between the two within the same power supply disturbance event window, unified quantitative output and conflict resolution are achieved, avoiding misjudgment caused by a single piece of evidence dominating in the scenario of superimposed power supply fluctuation and sensor degradation.

[0051] Specifically, the steps for stratification are as follows: The power supply contribution value obtained by exponentially mapping the power supply link fault support, the sensor contribution value obtained by exponentially mapping the sensor body fault support, and the composite anomaly support are all obtained by exponentially mapping the composite contribution value using the Sigmoid function, and are used as three basic contribution items. The maximum value of the three basic contribution items is used as the numerator, and the exponentially mapped values ​​of the three basic contribution items are added together, with the evidence conflict degree added as the denominator, which is always greater than zero. A ratio operation is performed to obtain a correction item, and the negative value of the degradation enhancement factor is subjected to a negative exponential operation, then one is added and the reciprocal is taken to obtain the degradation impact item. The correction item and the degradation impact item are multiplied together to obtain the stratification determination value.

[0052] The specific formula for calculating the stratification determination value is as follows:

[0053] ;

[0054] In the formula, This represents the hierarchical judgment value, used to quantify the likelihood that the sensor output anomaly within the current time slice is caused by a power supply link failure, a sensor body failure, or a combination of anomalies. This indicates the support for power supply failures, used to quantify the contribution of power supply link anomalies to output anomalies. This indicates the sensor failure support level, used to quantify the contribution of sensor-related abnormalities to output abnormalities. This indicates the support of composite anomalies, used to quantify the importance of composite anomalies in stratification. The degree of evidence conflict is indicated by calculating the ratio or difference between the maximum and minimum values ​​of the combined anomaly intensity of the signal layer, correlation layer, and residual layer, and is used to reduce the credibility of conflicting data. The degradation enhancement factor is obtained by the time cumulative growth rate of operating current deviation, noise increase, and output drift accumulation. Its value ranges from 0 to positive infinity and is used to enhance the weight of long-term degradation in the stratification determination.

[0055] Table 1, showing the fault stratification diagnostic data, records the degree of evidence support, conflict level, and final judgment results for power supply, sensor, and composite anomalies at different time slices. For time slice 3: power supply fault support is 0.2463, sensor fault support is 0.1277, composite anomaly support is 0.0399, evidence conflict level is 0.1020, and the stratification judgment value is 13.0186, with power supply link fault as the dominant factor. For time slice 10: power supply fault support is 0.0791, sensor fault support is 0.2060, composite anomaly support is 0.0307, ​​evidence conflict level is 0.0456, and the stratification judgment value is 14.3890, with sensor body fault as the dominant factor. For time slice 43: power supply fault support is 0.0... Time slice 232: Sensor fault support is 0.0504, composite anomaly support is 0.1053, evidence conflict is 0.2478, stratification judgment value is 433.9715, and the dominant factor is composite anomaly; Time slice 51: Power supply fault support is 0.8648, sensor fault support is 0.0035, composite anomaly support is 0.0122, evidence conflict is 0.0915, stratification judgment value is 14.5802, and the dominant factor is power supply link fault; Time slice 81: Power supply fault support is 0.0856, sensor fault support is 0.1479, composite anomaly support is 0.0207, evidence conflict is 0.2599, stratification judgment value is 8.7661, and the dominant factor is sensor body fault.

[0056] Table 1 Fault Stratification Diagnosis Data Table

[0057]

[0058] like Figure 4 The support graphs for each fault type are shown to visually illustrate the changes in support for various fault sources across different time slices. The horizontal axis represents the time slice, and the vertical axis represents the fault support value. The blue curve represents the support for power supply link faults, quantifying the inducing strength of power supply anomalies on sensor output anomalies; the green curve represents the support for sensor-specific faults, reflecting the contribution of sensor-specific anomalies to output anomalies; and the red curve represents the support for combined anomalies, describing the combined anomaly strength when power supply anomalies and sensor-specific anomalies coexist. The graphs show that fluctuations in support across different time slices reflect the dominance of different fault types in their respective time slices. When the blue curve is dominant, it indicates that power supply link faults are the primary cause; when the green curve is dominant, it indicates that sensor-specific faults dominate the anomaly; and when the red curve is dominant, it indicates that power supply disturbances and sensor anomalies work together to form a combined anomaly. The overall graph provides a quantitative basis for stratified judgment and recovery strategy selection, visually demonstrating the dynamic evolution of various fault types over time.

[0059] In this implementation scheme, by uniformly mapping the support of power supply failure, the support of sensor failure, and the support of composite anomaly into comparable contribution items, and introducing joint correction of evidence conflict degree and degradation enhancement factor, a stable output of hierarchical judgment of anomaly source on the time slice scale is achieved. This enables the consistency and distinguishability of the judgment results in complex scenarios of power supply disturbance, sensor degradation, and the superposition of the two.

[0060] Specifically, the steps for determining fault and abnormal states and generating fault source tags are as follows: Real-time comparison of the stratified judgment value with the fault source threshold. If the stratified judgment value is less than or equal to the fault source threshold, the current abnormality is determined to be a controllable abnormality, which can be automatically recovered through online calibration or parameter adjustment. If the stratified judgment value is greater than the fault source threshold, the fault source is further classified. This classification process includes ranking and comparing the support for power supply link abnormalities, sensor body faults, and composite abnormalities. If the support for power supply link abnormalities is the highest among the three support values, it is determined to be a power supply link fault state. If the support for sensor body faults is the highest among the three support values, it is determined to be a sensor body fault state. If the support for composite abnormalities is the highest among the three support values, it is determined to be a composite abnormal state. The identified fault state is structurally bound to the current pallet number, sensor number, logistics node information, sampling time information, and fault source category to generate a fault source tag. Simultaneously, a verification flag is set for sensors identified as abnormal. The verification flag indicates that subsequent maintenance cycles require focused inspection and labels the corresponding fault type. Fault types include subcategories such as voltage drop, transient voltage drop, ripple anomaly, drift, noise aging, and complete failure.

[0061] In this implementation plan, by comparing the hierarchical judgment value with the fault source threshold online and triggering the fine classification and sorting decision, the hierarchical identification and stable classification of abnormal states from controllable abnormalities to clear fault sources are realized, avoiding unclear sources or fluctuating judgment results under mixed conditions of power supply disturbances and sensor degradation.

[0062] Specifically, the steps for constructing a hierarchical priority decision-making mechanism based on fault source labels are as follows: extract the generated fault source labels, construct a hierarchical priority decision-making mechanism, and select a recovery path based on power supply link, sensor body soft fault, and sensor body hard fault: when the diagnosis result is a power supply link fault, enter the power supply side handling process; when the diagnosis result is a sensor body soft fault, enter the online self-calibration process; when the diagnosis result is a sensor body hard fault, enter the data reconstruction and degradation operation process.

[0063] In the power supply side handling process, based on the current power supply anomaly intensity and remaining power status, the following actions are performed: sensor sampling duty cycle adjustment, communication transmission frequency reduction, non-critical acquisition tasks postponement, backup power supply branch switching, battery maintenance mark generation, and power module inspection mark generation. When the power supply induced coupling strength assessment value falls back to within the threshold after power supply is restored, normal acquisition is resumed.

[0064] In the online self-calibration process, the shared tray is identified as being stationary, without environmental fluctuations, and within the calibration window of the relevant sensors. The deviation of the faulty sensor's current output is calculated, and the zero-point and gain parameters of the sensor are updated using the recursive least squares method. The correlation residuals are then monitored to ensure they return to normal after calibration. Normal correlation residual recovery is defined as the residuals from other reliable sensors being below a residual threshold for five consecutive sampling periods.

[0065] In the data reconstruction and degradation operation process, when the diagnosis result is that the sensor is completely failed or has an unrecoverable fault, the fault data reconstruction mode is activated: using normal sensor data that has physical or statistical correlation, a multiple linear regression model is adopted, with normal sensor readings as input features, and regression coefficients are trained through data to reconstruct the approximate output data of the faulty sensor in real time; at the same time, the degradation operation mode is switched and uncertainty labels are added to the reconstructed data.

[0066] When the diagnosis result is a complex anomaly, the complex handling process is initiated. The process prioritizes handling the power supply side, and the sensor status is reassessed after power is restored. If the sensor anomaly persists after power is restored, the handling process corresponding to the sensor body fault is initiated. If power cannot be restored, the power supply side handling is maintained and the complex anomaly status is reported.

[0067] In this implementation plan, a hierarchical priority decision-making mechanism based on fault source tags is used to achieve accurate response to anomalies in shared tray sensors and power supply links, forming a hierarchical, traceable, and closed-loop priority decision-making processing path, and realizing automatic diversion, reconstruction, and state recovery of multi-source anomalies in dynamic scenarios.

[0068] Specifically, the steps for optimizing fault recovery and generating a shared tray sensor fault detection report are as follows: The fault suppression gain value is obtained by calculating the difference in output anomaly scores after the execution of the current candidate recovery action; the energy consumption value is obtained by calculating the increased sensor sampling power consumption, communication power consumption, and processor load power consumption during the recovery action execution; the improvement is obtained by statistically analyzing the increase in sampling period before and after the recovery action execution; the edge computing load is obtained by calculating the increase in CPU utilization during the recovery action execution and normalizing it by dividing it by the rated CPU utilization; the success probability is obtained by statistically analyzing the proportion of successful recovery action executions to the total number of attempts; and the additional communication overhead is obtained by statistically analyzing the amount of new data packets or the communication bandwidth occupancy ratio during the recovery action. To ensure that the dimensions of each indicator are consistent and the final optimized evaluation value is dimensionless, the original dimensional indicators need to be normalized. The above normalized benchmark values ​​are all generated based on the device's rated parameters or operational statistics and are embedded in the edge computing unit. Through normalization, all indicators involved in the calculation are converted into dimensionless relative values ​​in the range of [0,1]. After adding 1, the value range is [1,2], ensuring that the results of multiplication and division operations are still dimensionless numbers, thereby guaranteeing the comparability and stability of the optimal evaluation values.

[0069] The fault suppression gain is increased by 1 to obtain the fault suppression term; the boost is increased by 1 to obtain the additional term; the success probability is increased by 1 to obtain the probability term; the energy consumption is increased by 1 to obtain the energy consumption term; the edge computing load is increased by 1 to obtain the edge load term; the communication overhead is increased by 1 to obtain the communication load term; the fault suppression term, additional term, and probability term are multiplied together to obtain the performance gain term; the energy consumption term, edge load term, and communication load term are multiplied together to obtain the resource overhead term; the performance gain term is divided by the resource overhead term to obtain the optimal evaluation value.

[0070] The specific formula for calculating the preferred evaluation value is as follows:

[0071] ;

[0072] In the formula, This represents the preferred evaluation value, used to quantify the overall advantages and disadvantages of the current candidate recovery action in terms of improving system stability, data continuity, and success probability. This represents the fault suppression gain value, used to measure the contribution of recovery actions to reducing the impact of faults; This indicates the energy consumption value and its intended use. This indicates the increase, used to measure the contribution of recovery actions to ensuring data continuity; This represents the edge computing load and is obtained by normalization. It is used to control the impact of recovery actions on edge computing resources. This indicates the probability of successful execution, reflecting the executability and reliability of the recovery action; This indicates additional communication overhead, used to reflect the consumption of network resources by recovery actions.

[0073] The system compares the optimal evaluation value with the recovery trigger threshold in real time. When the optimal evaluation value is greater than the recovery trigger threshold, the current candidate recovery action is executed. When the optimal evaluation value of the recovery strategy corresponding to the online self-calibration action is the largest, the online self-calibration process is triggered first. When the optimal evaluation value of the recovery strategy corresponding to the data reconstruction action is the largest, the data reconstruction and degradation operation process is initiated first. When the optimal evaluation value of the recovery strategy corresponding to the power supply adjustment action is the largest, the power supply strategy adjustment and communication duty cycle reduction process is executed first. When the optimal evaluation value of the recovery strategy is less than or equal to the recovery trigger threshold, no recovery operation is executed. The system keeps the current sensor data acquisition status unchanged, continuously monitors the optimal evaluation value of the time slice, and recalculates and compares the recovery trigger threshold in the next sampling period. The system structurally binds the recovery results, fault source category, handling action, and review status after execution to generate a shared tray sensor fault detection report.

[0074] like Figure 5 The diagram showing the optimal evaluation value region of the recovery strategy is used to visually display the optimal evaluation value of the recovery strategy in different time slices and its relationship with the recovery trigger threshold. The horizontal axis represents the time slice, and the vertical axis represents the optimal evaluation value of the recovery strategy. The blue curve represents the optimal evaluation value of the recovery strategy in each time slice, the red dashed line represents the recovery trigger threshold, and the orange highlighted area represents the time slice in which the system determines that the recovery conditions are met. As can be seen from the diagram, when the blue curve exceeds the red trigger threshold, the system determines that the recovery strategy conditions for the current time slice are met and enters the recovery action trigger state; the highlighted areas correspond to these trigger periods. When the blue curve is below the threshold, the system determines that the recovery conditions are not met, and the recovery action is paused or delayed. The overall diagram clearly reflects the dynamic optimization effect of the recovery strategy over time, providing a quantitative basis for intelligent triggering and adaptive delayed recovery, and verifying the effectiveness of the recovery strategy and the closed-loop decision-making of the system state.

[0075] In this implementation scheme, a dynamic response to shared tray sensor anomalies is achieved by determining the recovery strategy based on the preferred evaluation value, and a shared tray sensor fault detection report is generated. By visually comparing the preferred evaluation value with the threshold in a time series, the triggering period and recovery effect can be intuitively displayed, providing the system with a quantitative basis for intelligent triggering, adaptive delayed execution and closed-loop optimization, and realizing accurate response and automated management of multi-source sensor anomalies.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A fault detection method for shared tray sensors based on a multi-level fusion diagnostic model, characterized in that: Includes the following steps: S1 collects sensor data and obtains the operating electrical parameters of the power supply link, and performs time alignment, spatial registration and normalization processing to construct a joint monitoring dataset; S2, Identify power supply disturbance events based on multi-source sensing dataset, construct a power supply disturbance event window, evaluate the power supply-induced coupling strength, and construct a power supply-induced anomaly sample set; S3, based on the power supply disturbance event window, constructs a fusion diagnostic model, performs hierarchical judgment, determines the fault abnormal state, and generates fault source labels; S4, based on the fault source label, constructs a hierarchical priority decision-making mechanism, performs optimal evaluation of fault recovery, and generates a shared tray sensor fault detection report.

2. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for collecting sensor data and obtaining power supply link operating parameters, and performing time alignment, spatial registration, and normalization processing to construct a joint monitoring dataset are as follows: The system collects data from the built-in sensors of the shared pallet throughout the entire process of warehousing, loading and unloading, transportation, and turnover. The sensor data includes temperature, humidity, acceleration, angular velocity, BeiDou positioning data, vibration intensity, and communication status. Communication status parameters include signal reception strength, packet loss rate, retransmission count, and connection status. The system also synchronously acquires the operating electrical parameters of the power supply link, including battery voltage, operating current of each sensor, instantaneous voltage drop amplitude, power supply ripple amplitude, power recovery time during wake-up, and peak current pulse and pulse width during the communication transmission phase. Missing value detection is performed on sensor data using a dual criterion of time interval and frame sequence continuity. Missing data is filled in and marked through classification interpolation. Abnormal spikes in sensor data are removed by verifying the physical correlation of the neighborhood using a sliding window. Denoising is performed using a hierarchical denoising method that adapts to the filtering of different types of sensors. Sampling interval jitter is corrected by linear interpolation resampling. Ripple filtering and baseline drift correction are performed on the power supply link operating parameters. For transient disturbances during the wake-up switching phase, event edge marking and window isolation are performed. Time alignment is uniformly completed for sensor data and power supply link operating parameters with different sampling periods using linear interpolation. At the same time, spatial registration is performed according to the actual installation location of the sensor and the power supply branch affiliation. All processed sensor data and power supply link operating parameters are normalized in amplitude and integrated to construct a joint monitoring dataset.

3. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for identifying power supply disturbance events based on multi-source sensing datasets and constructing a power supply disturbance event window are as follows: Based on the power supply link working electrical parameter sequence in the joint monitoring dataset, the power supply disturbance event type is obtained by comparing the real-time monitoring power supply disturbance index with the trigger threshold. The power supply disturbance event types include voltage drop event, voltage drop event, ripple event, partial power failure event, power consumption wake-up switching event, and communication load pulse event. Power supply disturbance indicators include battery voltage, instantaneous voltage drop amplitude, power supply ripple amplitude, and power supply ripple amplitude change. For each identified power supply disturbance event, a sliding window monitoring method is used. When the power supply disturbance index first exceeds the threshold and remains at a constant level for N consecutive time slices, the first time slice of the N time slices is determined as the start time of the event. An extreme value search algorithm is used to find the moment corresponding to the maximum deviation of the power supply disturbance index within the event window to obtain the event peak moment. When the power supply disturbance index falls back below the threshold and remains stable within the threshold for a stable duration, the event recovery moment is obtained. A power supply disturbance event window is constructed based on the event start time, the event peak moment, and the event recovery moment.

4. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for evaluating the power supply-induced coupling strength are as follows: The probability of current pulse interruption is obtained by statistically analyzing the proportion of packet loss or interruption triggered by each communication pulse within the event window to the total number of pulses. The event recovery hysteresis is obtained by the time difference between the power supply disturbance recovery time and the sensor output recovery time; The voltage drop term is obtained by adding 1 to the ratio of the instantaneous voltage drop amplitude to the battery voltage within the power supply disturbance event window and then performing an exponential calculation of the instantaneous voltage drop enhancement coefficient. The ripple term is obtained by adding 1 to the ratio of the change in power supply ripple amplitude to the power supply ripple amplitude and then performing an exponential operation on the power supply ripple enhancement coefficient; the pulse term is obtained by adding 1 to the current pulse interruption trigger probability and then performing an exponential operation on the pulse trigger enhancement coefficient. The adjustment term is obtained by multiplying the anomaly coefficient by the adjustment factor and then exponentially operating on the natural constant e. The recovery term is obtained by adding 1 to the ratio of the event recovery hysteresis difference to the event recovery time and then exponentially operating on the recovery hysteresis suppression coefficient. The power supply induced coupling strength evaluation value is obtained by multiplying the voltage drop contribution term, ripple contribution term, pulse contribution term, and adjustment term and then dividing by the recovery term.

5. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for constructing the power supply-induced anomaly sample set are as follows: The power supply induced coupling strength assessment value is compared with the coupling judgment threshold in real time. When the power supply induced coupling strength assessment value is greater than the coupling judgment threshold, it is determined that the sensor output abnormality in the current time slice is coupled with the power supply disturbance. When the power supply induced coupling strength assessment value is less than or equal to the coupling judgment threshold, it is determined that the sensor output abnormality in the current time slice is not coupled with the power supply disturbance.

6. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for constructing the fusion diagnostic model based on the power supply disturbance event window are as follows: Based on the power supply disturbance event window, a fusion diagnostic model is constructed, which includes a power supply link screening layer, a sensor body fault determination layer, and a composite anomaly conflict resolution layer. In the power supply link screening layer, the battery voltage drop amplitude, module operating current deviation, instantaneous voltage drop amplitude, power supply ripple amplitude, wake-up recovery delay, and power supply induced coupling strength evaluation value are extracted from the power supply disturbance event window. These values ​​are then normalized. The normalized value of each feature is incremented by 1 and the natural logarithm is taken. The logarithms of all features are then summed. The summation result is divided by the number of features and subjected to an exponential transformation to obtain the power supply fault support. In the sensor body fault determination layer, the sensor output drift, noise energy increase, root mean square value of the associated residual sequence, zero-value holding time, and number of range exceedances are extracted. A fuzzy comprehensive evaluation method is used to construct the sensor fault support. In the composite anomaly conflict resolution layer, a basic probability assignment function is constructed based on the support of power supply failure and sensor failure, generating power supply failure evidence bodies and sensor failure evidence bodies respectively. Each evidence body assigns probability mass to three types of propositions: power supply failure, sensor failure, and composite anomaly, forming a three-dimensional support vector. The degree of overlap is obtained by calculating the cosine similarity between the two evidence body vectors, and the Dempster-Shafer evidence theory is used for fusion to construct the composite anomaly support. The evidence conflict degree is obtained by multiplying the basic probability assignments of all proposition pairs with no intersection in the two evidence bodies and summing them.

7. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for performing the stratification determination are as follows: The power supply contribution value obtained by exponentially mapping the power supply link fault support, the sensor contribution value obtained by exponentially mapping the sensor body fault support, and the composite anomaly support are all obtained by exponentially mapping the composite contribution value using the Sigmoid function. The resulting composite contribution value is used as three basic contribution terms. The maximum value of the three basic contribution terms is used as the numerator, and the exponentially mapped values ​​of the three basic contribution terms are added together. The evidence conflict degree is added as the denominator, and the ratio is calculated to obtain the correction term. The negative value of the degradation enhancement factor is subjected to negative exponential operation, and the reciprocal of the result is added to obtain the degradation effect term. The stratification determination value is obtained by multiplying the correction term by the degradation effect term.

8. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for determining the fault abnormality state and generating the fault source tag are as follows: The hierarchical judgment value is compared with the fault source threshold in real time. When the hierarchical judgment value is less than or equal to the fault source threshold, the current anomaly is determined to be a controllable anomaly. If the stratification judgment value is greater than the fault source threshold, the process of detailed fault source classification will begin. When the support for a power supply link anomaly is the largest among the three supports, it is determined to be a power supply link fault state; when the support for a sensor body fault is the largest among the three supports, it is determined to be a sensor body fault state; when the support for a composite anomaly is the largest among the three supports, it is determined to be a composite anomaly state; generate a fault source label, and set a verification flag for the sensors identified as abnormal, and label the corresponding fault type.

9. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for constructing a hierarchical priority decision-making mechanism based on fault source labels are as follows: Extract the generated fault source tags, construct a hierarchical priority decision-making mechanism, and select the recovery path based on power supply link, sensor body soft fault, and sensor body hard fault: when the diagnosis result is a power supply link fault, enter the power supply side handling process; When the diagnostic result indicates a soft fault in the sensor itself, the online self-calibration process is initiated. When the diagnosis result is a hard fault in the sensor itself, the data reconstruction and degradation operation process is initiated. When the diagnosis result is a complex abnormality, the complex treatment procedure is initiated.

10. The shared tray sensor fault detection method based on a multi-level fusion diagnostic model according to claim 1, characterized in that: The specific steps for optimizing fault recovery and generating a shared tray sensor fault detection report are as follows: The fault suppression gain is obtained by calculating the difference in output anomaly scores after the current candidate recovery action is executed. The energy consumption is obtained by calculating the increased sensor sampling power consumption, communication power consumption, and processor load power consumption during the recovery action execution. The improvement is obtained by statistically analyzing the increase in sampling period before and after the recovery action execution. The edge computing load is obtained by calculating the increase in CPU utilization during the recovery action execution and normalizing it by dividing it by the rated CPU utilization. The success probability is obtained by statistically analyzing the proportion of successful recovery action executions to the total number of attempts. The additional communication overhead is obtained by statistically analyzing the amount of new data packets or the communication bandwidth occupancy ratio during the recovery action. The fault suppression gain is increased by 1 to obtain the fault suppression term; the boost is increased by 1 to obtain the additional term; the success probability is increased by 1 to obtain the probability term; the energy consumption is increased by 1 to obtain the energy consumption term; the edge computing load is increased by 1 to obtain the edge load term; the communication overhead is increased by 1 to obtain the communication load term; the fault suppression term, additional term, and probability term are multiplied together to obtain the performance gain term; the energy consumption term, edge load term, and communication load term are multiplied together to obtain the resource overhead term; the performance gain term is divided by the resource overhead term to obtain the optimal evaluation value. The system compares the preferred evaluation value with the recovery trigger threshold in real time. When the preferred evaluation value is greater than the recovery trigger threshold, the current candidate recovery action is executed. When the preferred evaluation value of the recovery strategy is less than or equal to the recovery trigger threshold, no recovery operation is executed. The system then structurally binds the recovery result, fault source category, handling action, and review status to generate a shared tray sensor fault detection report.

Images