A multi-source industrial equipment data collection method

Abstract

The application discloses a multi-source industrial equipment data acquisition method, and relates to the technical field of data fusion processing, which comprises the following steps: enhancing a package header for data encapsulation on the edge side and transmitting; reordering and aligning the data by an upper computer, evaluating transmission quality and generating a label; meanwhile, evaluating the health of sensors and generating a label through a multi-sensor consistency model and signal analysis; based on the joint evaluation of transmission quality and sensor health, implementing intelligent reconstruction and correction on missing or suspicious data; finally, giving the data a credibility weight for weighted fusion and diagnosis, and outputting the diagnosis result and confidence degree fused with data quality information; the method integrates the evaluation of data quality and sensor state, and performs resilience data processing accordingly, which significantly improves the accuracy and reliability of industrial equipment fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of data fusion processing technology, and in particular to a method for acquiring data from multi-source industrial equipment. Background Technology

[0002] In the context of modern industrial internet and intelligent manufacturing, condition monitoring and fault prediction of critical equipment rely on high-quality, highly consistent data collected in real time from multiple sensors (such as vibration, current, and temperature). However, in complex factory environments, the acquisition of multi-source data faces a series of inherent challenges. First, during the transmission of data from edge sensors to the host computer via industrial networks, network jitter and bandwidth contention often lead to out-of-order delivery, packet loss, and non-deterministic delays, disrupting the strict time synchronization between multi-source signals and introducing fundamental errors into subsequent time-series-based fusion analysis. Second, the unreliability of the transmission network is only one aspect; sensors themselves may also experience performance degradation or even failure due to aging, drift, abnormal power supply, or loose installation, resulting in distorted output signals that traditional acquisition systems struggle to distinguish from actual equipment status changes. The intertwining of transmission layer issues and sensor reliability problems makes it difficult to assess the quality of the raw dataset, leading to low confidence and a high risk of false alarms and missed alarms in fault diagnosis based directly on such data.

[0003] To address these challenges, existing technologies typically handle transmission quality or sensor health issues in isolation. For example, they ensure transmission reliability through timestamp synchronization and retransmission mechanisms, or identify obviously faulty sensors through preset thresholds and periodic calibration. However, these methods have significant limitations: First, they lack a joint, online, and refined evaluation framework for data quality, failing to dynamically distinguish between data loss due to network issues and data unreliability due to sensor malfunctions. Second, when data problems are detected, traditional approaches often involve simple discarding or interpolation, failing to utilize the inherent physical correlations and redundancy information between multi-source signals to intelligently and faithfully reconstruct and repair suspicious or missing data. Third, in the final fault diagnosis, there is a lack of a fusion mechanism that quantifies the data's inherent reliability as a weight; the diagnostic model treats all inputs equally, making its conclusions susceptible to contamination by low-quality data. Summary of the Invention

[0004] This application provides a multi-source industrial equipment data acquisition method that can comprehensively assess the data transmission quality and sensor health, and thereby realize a comprehensive acquisition and processing method for data resilience reconstruction and reliable fusion diagnosis.

[0005] This application provides a method for acquiring data from multi-source industrial equipment, including:

[0006] S1, at the edge of data acquisition, encapsulates an enhanced header for each sensor data packet containing a global serial number, a high-precision timestamp, and auxiliary sensor readings, and transmits it through the industrial network; S2, the host computer reorders and aligns the received data packets according to the timestamp, detects packet loss and severe delay, and generates a preliminary transmission quality label for each data packet; S3 runs a lightweight multi-sensor consistency model to calculate the deviation in the relationship between signals, and analyzes the long-term statistical characteristics and auxiliary readings of each sensor signal to generate sensor health scores and labels. S4 integrates transmission quality tags and sensor health tags, generates repairs for lost but sensor-healthy data, downweights and reconstructs signals to replace data that is complete but sensor-susceptible, and marks channels that are completely unusable. S5 calculates the comprehensive confidence weight for each sensor data, performs weighted processing during the feature fusion stage, and outputs the fault type and diagnostic confidence based on model probability and average data confidence simultaneously after being input into the diagnostic model.

[0007] Preferably, in step S1, during data acquisition, the edge computing unit periodically reads the core measurement values ​​from the connected physical sensors, the core measurement values ​​including device status monitoring data; the enhanced packet header specifically includes: an independent, continuously increasing global sequence number for detecting packet loss for each sensor data stream; a high-precision global timestamp generated by the edge side based on a precision clock synchronization protocol with microsecond-level accuracy; and a snapshot of the sensor's original readings, including the instantaneous power supply voltage and ambient temperature of the sensor channel recorded by the edge computing unit, reflecting its own operating status.

[0008] Preferably, the formula for calculating the Kangdu score is: Where C is the consistency health component, P is the ontology feature health component, and α and β are weighting coefficients, satisfying α + β = 1; the formula for calculating C is: ,in This is the deviation value calculated from other related signals by the consistency model at the current moment. It is a preset deviation threshold; the formula for calculating P is: Where n is the total number of features monitored. Let i be the long-term normal value obtained statistically from the sensor's health baseline period for the i-th feature. The value of the i-th feature is obtained statistically within the current monitoring window. This represents the maximum allowed degradation threshold for this feature.

[0009] Preferably, in step S4, for each sensor data at each time moment, a corresponding processing path is triggered based on its quality-health label combination, including: Path 1: When data is marked as transmission lost and the sensor health label is healthy, a pre-trained conditional generation model is invoked to generate an estimated value for the corresponding data point marked as transmission lost using other healthy and complete sensor data at the same time moment; Path 2: When data arrives completely and the sensor health label is warning or fault, a low initial confidence weight is assigned to the data, and the signal estimate that the sensor should have is reconstructed through the generation model using other highly healthy sensor data at the same time moment to partially correct or replace the original suspicious signal; Path 3: When data is marked as lost and the sensor health label is also fault, the channel data is determined to be completely unusable, its weight is set to zero, and a maintenance alarm is generated.

[0010] Preferably, the step of generating and repairing the missing but healthy data includes: S41, for the various processing markers and sensor health labels generated in S4, predefining their corresponding influence decay functions, and maintaining a time-varying mass footprint vector for each monitored physical quantity; S42, continuously recording the health score timeline of all sensors and all mass label events to form an audit log; when the sensor health status is clearly corrected, automatically triggering the backtracking process for historical time periods; S43, for the original data within the affected time period, using the corrected sensor health information, re-execute steps S1 to S5 in sequence.

[0011] Preferably, maintaining a time-varying mass footprint vector includes: maintaining a mass footprint vector for each independently monitored physical quantity, where each dimension of the vector corresponds to a predefined mass defect type, with an initial value of zero; at each processing moment, generating an influence injection vector based on the processing marker determined for the current mass point and the sensor health label, where the dimension of this vector is the same as the mass footprint vector, and the value of each dimension represents the newly generated influence intensity of the corresponding type of defect at the current moment; and calculating the mass footprint vector at the current moment through iterative updates, with the update formula being: ,in, Let be the quality footprint vector at the current moment. Let this be the mass footprint vector from the previous time step. It is a diagonal matrix, where each element on the diagonal is the attenuation coefficient corresponding to the defect type. Inject a vector into the influence generated at the current moment.

[0012] Preferably, the process of triggering the backtracking of a historical time period includes: when the state of any sensor is detected to meet a preset explicit correction condition, capturing the state of the sensor at the time of triggering the correction, including: recording the sensor's identifier, state change type, exact time T of the change, and specific state values ​​before and after the change; determining the corresponding backtracking time depth Δ based on the state correction event and the sensor type, and accordingly determining the historical time window [T-Δ, T] that needs to be re-evaluated; and generating a backtracking task instance based on the recorded state correction event and the determined historical time window.

[0013] Preferably, the automatic triggering of the historical time period backtracking process specifically includes: S421, by defining the physical connection topology and coupling transfer function between production line equipment, a graph model describing the mutual influence relationship of physical quantities across equipment is constructed; S422, after completing the diagnosis of a single equipment, all its upstream equipment is traced back, and the quality footprint of the upstream equipment is converted and superimposed through the corresponding coupling transfer function to calculate an additional cross-equipment quality pollution index for the target equipment; S423, the simulation program is run regularly to virtually repair each unhealthy sensor, and its improvement effect on the reliability of the whole network diagnosis is evaluated through the coupling model, a repair priority list is generated, and the production line quality status and key pollution paths are dynamically displayed in the panoramic view; S424, after the historical diagnosis conclusion of a certain equipment is corrected, all its physical coupling downstream equipment is automatically identified, and the collaborative backtracking and production line-level reliability reassessment of the historical data of these downstream equipment in the same time period is triggered.

[0014] Preferably, the graph model of the interaction relationship between physical quantities across devices includes: defining directed edges between nodes based on the actual physical connection relationship between devices, and labeling each edge with the coupling type it represents; associating each directed edge with a coupling transfer function, which describes the relationship between a specific physical state from the source device to the target device propagating through the physical path and affecting the sensor readings of the target device; and constructing the nodes, directed edges and their associated coupling transfer functions into a computable directed graph data structure.

[0015] Preferably, the step of tracing all upstream devices after completing the diagnosis of a single device specifically includes: establishing a frequency response function for each directed edge in the coupling graph to describe the transmission characteristics of the signal as the frequency changes, and generating a predicted transmission signal as a feature template of the pollution source of the path based on this function and the estimated source signal of the upstream device; performing time-frequency analysis on the observed signal of the downstream device and modeling it as a linear combination of its own state components and each upstream pollution feature template, obtaining the contribution coefficient of each component by solving the optimization problem, thereby quantifying the specific contribution percentage of each upstream pollution source; dynamically highlighting the main pollution paths in the situation map based on the real-time calculated pollution contribution of each path and generating precise suggestions for repairing specific sensors, while monitoring abrupt changes in the contribution time series to warn of state changes of upstream devices or physical paths; and embedding a contribution analysis process when conducting collaborative backtracking and reassessment of historical data to accurately quantify the pollution contribution ratio of upstream devices to downstream signals at historical moments.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: By encapsulating high-precision timestamps and serial numbers at the edge, precise synchronization and transmission quality monitoring of multi-source data are achieved. Through parallel execution of consistency models and statistical analysis, independent assessment of sensor health is realized. Its core innovation lies in the intelligent, multi-source correlation-based condition generation and reconstruction of lost or suspicious data based on a joint assessment of transmission quality and sensor health, thereby enhancing transmission resilience at the data level. Subsequently, weighted fusion and diagnosis are performed using confidence weights, outputting a diagnostic confidence score that incorporates data quality information. Ultimately, this significantly improves the accuracy and reliability of fault diagnosis in complex industrial environments, reducing false alarms and missed alarms caused by data transmission problems or sensor malfunctions.

[0017] By establishing a quality footprint model, the risk of persistent interference from historical data quality issues with current diagnostic conclusions is quantified. This model automatically triggers reassessment and calibration of historical diagnostic results when sensor conditions are subsequently corrected. This mechanism effectively eliminates the risk of long-term misjudgment caused by slow sensor degradation or calibration delays, significantly improving the longitudinal reliability of diagnostic conclusions over time. Simultaneously, the panoramic timeline tracing function provides complete quality evidence and correction records for every decision, enhancing operational trust and providing high-quality, interpretable closed-loop data for continuous model optimization. Ultimately, this enables the diagnostic system to achieve self-verification, self-correction, and continuous evolution in complex industrial environments.

[0018] By constructing a physical coupling map of production line equipment and calculating cross-equipment quality contamination indices, this method quantifies for the first time the indirect interference risk of upstream equipment data quality issues on downstream diagnostic conclusions, thereby extending diagnostic reliability assessment from a single point to a collaborative level across the entire production line. This approach not only identifies and warns of contaminated diagnoses and avoids erroneous maintenance, but also intelligently recommends optimal sensor repair paths through simulation analysis, maximizing the overall diagnostic reliability of the production line at minimal cost. Simultaneously, its panoramic visualization and collaborative backtracking mechanism allows historical conclusions to be dynamically updated as the state of the source sensors is corrected, ultimately forming a self-evolving diagnostic ecosystem capable of understanding and optimizing production line-level data health.

[0019] By introducing a frequency-domain selective coupling feature library and multi-source signal contribution analysis, the percentage contribution of each upstream device to the downstream signal contamination can be quantified, thereby accurately locating the main interference sources and specific sensors, greatly improving the targeting and efficiency of maintenance. Simultaneously, by monitoring abrupt changes in contamination contribution, it is also possible to proactively warn of abnormal changes in the state of upstream equipment or its physical connection paths, providing unprecedented data-driven insights for the precise maintenance and health management of complex industrial production lines. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a multi-source industrial equipment data acquisition method according to an embodiment of the present invention. Detailed Implementation

[0021] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terminology used herein includes any and all combinations of one or more of the associated listed items.

[0023] Example 1: Figure 1 This is a flowchart illustrating a multi-source industrial equipment data acquisition method according to an embodiment of the present invention.

[0024] like Figure 1 As shown, a method for acquiring data from multi-source industrial equipment includes the following steps: S1 encapsulates an enhanced header containing a global serial number, a high-precision timestamp, and auxiliary sensor readings for each sensor data packet at the edge of data acquisition, and transmits it over the industrial network.

[0025] During data acquisition, the edge computing unit periodically reads core measurement values ​​from the connected physical sensors. These include equipment condition monitoring data such as vibration acceleration / velocity, current, voltage, temperature, pressure, and noise.

[0026] The enhanced packet header specifically consists of an independent, continuously increasing sequence number for each sensor data stream, used for detecting packet loss. A unified time base, generated by the edge side based on a precision clock synchronization protocol, requires a high-precision global timestamp at the microsecond level. The edge computing unit records snapshots of the sensor's raw readings, including instantaneous power supply voltage, ambient temperature, and other auxiliary readings that may reflect its operating status.

[0027] S2, the host computer reorders and aligns the received data packets based on the timestamp, detects packet loss and severe delay, and generates a preliminary transmission quality label for each data packet.

[0028] Specifically, the host computer (a computer or server responsible for centralized processing, analysis, and control of the data uploaded by the slave devices) continuously listens to the network port and stores incoming data packets from various sensors into a dynamic time window buffer (e.g., overwriting the data from the most recent 2 seconds). Data packets are retrieved from the buffer, and their enhanced headers contain a high-precision global timestamp. Based on this timestamp, all data packets from different sources are uniformly sorted and aligned to the same microsecond-level timeline to correct for out-of-order issues caused by network transmission.

[0029] For each independent sensor data stream, check the continuity of the global sequence number in its data packets. If a sequence number interval is found to be greater than 1, it is determined that data packet loss has occurred within that interval, and the range of lost sequence numbers is recorded. For each data packet with aligned timestamps, calculate the difference between its timestamp and the host computer system time at the time of data reception. If this difference exceeds a preset threshold, the data packet is determined to be severely delayed. Based on the detection results, assign a status label to each sensor data at each time point: complete (data packets arrive in sequence, no packet loss, and delay within the threshold), lost (data packets were not received at the expected time point (sequence numbers are not continuous)), and delayed (data packets arrived, but their timestamps indicate that they experienced network delays exceeding the threshold).

[0030] S3 runs a lightweight multi-sensor consistency model, calculates the deviation in the relationship between signals, analyzes the long-term statistical characteristics and auxiliary readings of each sensor signal, and generates sensor health scores and labels.

[0031] Specifically, multi-source sensor data from the same cross-section at the same time point, aligned using the S2 time-series algorithm, are acquired. This aligned multi-source data (such as three-phase current and multiple vibration signals) is then input into a pre-trained lightweight multi-sensor consistency model. This model is trained using supervised learning. First, a large number of synchronized multi-sensor data segments under healthy conditions are collected as positive samples. The model architecture can employ a lightweight graph neural network or an attention mechanism network, using all sensor signals at a given time point as input to learn and predict the ideal relationships between the sensor signals. The training objective is to minimize the reconstruction error between the model output and the actual healthy signal. After training, the model can encode the normal physical or statistical correlations between signals, used to calculate the deviation between the input signal and the ideal relationship.

[0032] The model calculates the deviation between the actual relationship of the current multi-source signals and the ideal relationship embedded in the model (the imbalance of three-phase current and the energy ratio error of the associated signals). Independent online analysis is performed on the signal sequences of each sensor, including: statistical characteristic analysis, continuously tracking and updating the long-term statistical characteristics of the signals, such as mean (zero point), standard deviation (noise level), amplitude distribution, etc., and monitoring their slow changing trends. Auxiliary reading monitoring checks the instantaneous power supply voltage and ambient temperature readings of the sensors attached to each data packet to determine if they exceed the normal operating range.

[0033] By fusing consistency deviation results (whether a sensor continuously disrupts overall consistency) and ontological feature analysis results (whether its own signal characteristics have degraded, and whether auxiliary readings are abnormal), a sensor health label (e.g., healthy, warning, fault) or a quantified health score is generated for each sensor through preset rules or a lightweight scoring model.

[0034] The formula for calculating the health score is as follows: H is the final health score, typically ranging from 0.0 (complete failure) to 1.0 (complete health); C is the consistency health component, calculated from the output of the multi-sensor consistency model. , This is the deviation value of the sensor signal from other correlated signals calculated using a consistency model at the current moment. This is a preset deviation threshold. When When C=0, it indicates that the sensor has completely failed in terms of consistency. This formula maps the original deviation of the model output to a health score between 0 and 1; the smaller the deviation, the higher the score. P is the ontological characteristic health component, which is derived from long-term statistical analysis of the sensor's own signals. , It is the long-term normal value obtained statistically from the i-th feature during the sensor's health baseline period. It is the value obtained by statistical analysis of the i-th feature within the current monitoring window. This refers to the maximum allowable degradation threshold for this feature (an empirical or calibrated value determined based on historical data statistical analysis for specific application scenarios; for example, for zero-point drift characteristics, if the mean value of a vibration sensor during its healthy reference period is 0mV, engineering practice considers a drift exceeding ±10mV as severe degradation, then a threshold can be set). =10mV; For noise level variation characteristics, if the standard deviation of the baseline period is 0.5mV, then an increase in noise exceeding 3 times (i.e., 1.5mV) is considered abnormal, and a setting can be made. =1.0mV (increment of 1.0mV)), n is the total number of monitored features. Common features include: zero drift, |current signal mean - reference signal mean|; noise level change, |current signal standard deviation - reference signal standard deviation|; auxiliary reading anomalies, such as |current supply voltage - rated voltage|; signal saturation rate is the percentage of the signal that exceeds the range; α and β are weighting coefficients that satisfy α+β=1, used to adjust the relative importance of the two components. For example, they can be set to α=0.6 and β=0.4.

[0035] S4 integrates transmission quality and sensor health labels, generates repairs for lost but healthy data, downweights and reconstructs signals to replace complete but questionable data, and marks completely unusable channels.

[0036] Specifically, it receives a queue of raw data from multiple sources of sensors aligned by timestamps, and synchronously acquires the transmission quality label (complete / lost / delay) and sensor health label (healthy / early warning / fault, etc. or rating) corresponding to each data point.

[0037] For each sensor data point at each time moment, the corresponding processing path is triggered based on its quality-health label combination: Path 1 handles data that is lost during transmission but whose sensors are healthy. When a data point is detected as missing, but its corresponding sensor health label is healthy, conditional generation and imputation are triggered. At this point, a pre-trained conditional generation model is invoked, using all other healthy and complete sensor data at the same time as input conditions to generate estimated values ​​for sensor data points that did not reach the host computer due to network transmission issues, thus filling the data gaps. The conditional generation model employs supervised training. A large number of multi-sensor synchronous data segments are collected under healthy device conditions, and data from some sensor channels are randomly masked as input conditions, using the real data from the masked channels as the training target. The model architecture can use a conditional variational autoencoder or a conditional generative adversarial network to learn the mapping from known channel data to missing channel data. The training objective is to minimize the mean square error or sensing loss between the generated signal and the real signal. After training, the model can generate estimated values ​​for missing sensors at any given time based on healthy sensor data.

[0038] Path two processes data that is complete but indicates a suspected sensor malfunction. When complete data arrives, but the sensor's health label is warning or faulty, a weighting and reconstruction process is triggered. First, a very low initial confidence weight is assigned to this data. Simultaneously, using data from other highly healthy sensors at the same time, a generative model is used to reconstruct the expected signal of the sensor under the current operating conditions. In subsequent processing, this reconstructed signal will be used to partially correct or replace the original suspicious signal.

[0039] Path 3 handles data transmission loss and sensor failure. When data is marked as lost and its sensor health label is also faulty, the channel is considered completely failed at that moment. Data in this channel will not be filled, its weight will be set to zero, and a high-priority maintenance alert will be generated, indicating the faulty sensor and link.

[0040] After the above processing, a set of cleaned data that has been identified and may be repaired or replaced is output for each sensor channel. At the same time, a preliminary processing label (such as original, generated, corrected, unavailable) determined by this step and an initial confidence weight coefficient are attached to each data point for use in subsequent fusion steps.

[0041] S5 calculates the comprehensive confidence weight for each sensor data, performs weighted processing during the feature fusion stage, and outputs the fault type and diagnostic confidence based on model probability and average data confidence simultaneously after being input into the diagnostic model.

[0042] Specifically, for each sensor's data at each time moment, a comprehensive reliability weight between 0 and 1 is calculated based on its transmission quality label (complete / lost / delayed) and sensor health label (or score) through a preset rule mapping. This weight comprehensively reflects the reliability of the data point in terms of both transmission reliability and sensor reliability.

[0043] The time-series data from all sensors are input into a feature extraction network (such as a convolutional neural network) for feature extraction. In the feature fusion layer, the feature vector extracted from each sensor channel is multiplied by its corresponding comprehensive confidence weight, and then aggregated (such as by concatenation or summation) to form the final weighted fused feature vector.

[0044] The weighted fused feature vector is input into the trained fault diagnosis model (trained using supervised learning. Extensive multi-sensor time-series data of the equipment under various known fault modes and health states are collected, processed by S1-S4 to obtain a weighted fused feature vector as input, and the corresponding fault type label as output. The model architecture uses LSTM or GRU, and the training objective is to minimize the cross-entropy loss function. After training, the model can output the probability distribution of each fault type based on the input weighted fused features). The model outputs the diagnostic result for the equipment status (fault type probability distribution). Simultaneously, the overall confidence level of this diagnosis is calculated based on two key pieces of information: the highest category probability value output by the diagnostic model, and the average overall confidence weight of all sensor data involved in this diagnosis.

[0045] Based on the calculated overall confidence level, a tiered decision-making process is executed and output: If the confidence level is high, a warning for the diagnosed fault type is directly issued. If the confidence level is medium, explanatory notes regarding the potential impact of data quality on the results are attached along with the diagnostic results. If the confidence level is low, a device fault warning is not issued temporarily; instead, a self-inspection suggestion report for the sensing subsystem targeting a specific sensor or data link is generated. Finally, the diagnostic conclusion (or self-inspection report) and the overall confidence level value of this diagnosis are output simultaneously.

[0046] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By encapsulating high-precision timestamps and serial numbers at the edge, precise synchronization and transmission quality monitoring of multi-source data are achieved. Through parallel execution of consistency models and statistical analysis, independent assessment of sensor health is realized. Its core innovation lies in the intelligent, multi-source correlation-based condition generation and reconstruction of lost or suspicious data based on a joint assessment of transmission quality and sensor health, thereby enhancing transmission resilience at the data level. Subsequently, weighted fusion and diagnosis are performed using confidence weights, outputting a diagnostic confidence score that incorporates data quality information. Ultimately, this significantly improves the accuracy and reliability of fault diagnosis in complex industrial environments, reducing false alarms and missed alarms caused by data transmission problems or sensor malfunctions.

[0047] Example 2: In Example 1, by real-time evaluation of transmission quality and sensor health, and accordingly performing resilient reconstruction and weighted fusion of data, instant purification and reliable diagnosis of multi-source data under complex field conditions were achieved. However, considerations for data quality issues are still mainly limited to fragmented evaluations at the current or near-real-time level. The continuous and cumulative impact of various quality defects in historical data (such as slow sensor drift and data generated after occasional packet loss) on current diagnostic conclusions cannot be effectively quantified and tracked. More importantly, when the health status of a sensor is clearly corrected afterward (e.g., its zero-drift fault is confirmed through on-site calibration), there is a lack of an automatic retrospective, re-evaluation, and correction mechanism for diagnostic conclusions within the affected historical period. This may lead to diagnostic conclusions based on erroneous historical data continuously misleading maintenance decisions, and it is also impossible to utilize the critical event of state correction to optimize the model. To quantify the longitudinal impact of historical quality issues and achieve dynamic calibration and continuous optimization of diagnostic conclusions during post-sensor state correction, further optimization and improvement are needed based on Example 1.

[0048] In some embodiments, to repair the generation of conditions for initiating lost but healthy data, step S4 further includes: S41 predefines the corresponding influence decay function for the various processing markers and sensor health labels generated in S4, and maintains a time-varying mass footprint vector for each monitored physical quantity.

[0049] Specifically, for the various processing markers and sensor health labels output by step S4, their corresponding quality defect types are clearly defined. For each defect type, combined with the characteristics of key operators (filters, LSTM forget gates) in the downstream signal processing chain, an influence decay function is predefined. This function (e.g., exponential decay form) quantifies the law that the influence of a single defect event on subsequent data points weakens over time, and a decay coefficient is determined from it.

[0050] For each independently monitored physical quantity (e.g., X-axis vibration acceleration at the drive end of motor #1), initialize a mass footprint vector. The number of dimensions of this vector is the same as the number of defined mass defect types, with each dimension corresponding to one defect type. The initial value is usually set to zero vector.

[0051] At each data processing moment, based on the processing marker and sensor health label determined for the current data point in step S4, an influence injection vector is generated through a preset rule mapping table. This vector has the same dimension as the quality footprint vector, and the value of each dimension represents the intensity of the new influence of the corresponding type of defect at the current moment (positive values ​​indicate an enhanced negative impact, and zero or negative values ​​indicate no impact or a positive impact).

[0052] The mass footprint vector at the current time t is calculated using an iterative update method. The updated formula is: .in, It is the footprint vector from the previous moment. It is a diagonal matrix, where each element on the diagonal is a decay coefficient defined for the corresponding defect type. This is the generated influence injection vector for the current moment. This step aims to dynamically track the residual cumulative impact of various historical quality defects on the current data point.

[0053] S42 continuously records the health score timeline of all sensors and all quality label events to form an audit log; when the health status of a sensor is clearly corrected, it automatically triggers the backtracking process for historical time periods.

[0054] Specifically, a timestamped continuous audit log is created to persistently record two core time-series data: first, the detailed health scores and status labels of all sensors at each moment; and second, all events related to data quality, including the transport quality labels generated in step S2 and the data processing labels generated in step S4.

[0055] Establish a set of explicit rules to determine when a sensor's health status has undergone a clear correction. These include: a sudden change in the health score from a stable range and its continuous transition to another range; a specific transition in the status label; or receiving and confirming sensor parameter update instructions from external calibration tools or maintenance work orders.

[0056] Continuously monitor the health scores and tag streams of all sensors. When the state of any sensor meets the defined correction conditions, immediately capture the event and record the sensor ID, state change type, exact time T when the change occurred, and the specific state values ​​before and after the change.

[0057] Based on the sensor type and fault mode of the triggered state correction event, the corresponding backtracking time depth Δ is obtained from the preset configuration or calculated by rules. For example, for a slow drift fault, Δ may be several hours; for a sudden fault, Δ may be shorter. Based on this, the historical time window [T-Δ, T] that needs to be reassessed is determined. A backtracking task instance is generated, which includes the identifier of the triggering sensor, details of the state change, and the determined time window [T-Δ, T]. The rules for calculating the corresponding backtracking time depth Δ can refer to: a lookup rule based on fault mode: a default Δ value or range is preset for each fault mode (e.g., slow drift, sudden impact, intermittent anomaly). A calculation rule based on degradation rate: Δ is calculated based on the rate of change of the sensor health score (e.g., the score decreases per minute). If the degradation rate is slow (e.g., a decrease of 0.01 points per hour), Δ takes a larger value; if the degradation rate is fast (e.g., a decrease of 0.1 points per minute), Δ takes a smaller value. Rules based on signal statistical characteristics: For example, dynamically adjusting Δ based on the autocorrelation time or noise level of the sensor signal. These can be further defined according to specific application scenarios.

[0058] S43, For the raw data within the affected time period, use the corrected sensor health information to re-execute steps S1 to S5 in sequence.

[0059] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By establishing a quality footprint model, the risk of persistent interference from historical data quality issues with current diagnostic conclusions is quantified. This model automatically triggers reassessment and calibration of historical diagnostic results when sensor conditions are subsequently corrected. This mechanism effectively eliminates the risk of long-term misjudgment caused by slow sensor degradation or calibration delays, significantly improving the longitudinal reliability of diagnostic conclusions over time. Simultaneously, the panoramic timeline tracing function provides complete quality evidence and correction records for every decision, enhancing operational trust and providing high-quality, interpretable closed-loop data for continuous model optimization. Ultimately, this enables the diagnostic system to achieve self-verification, self-correction, and continuous evolution in complex industrial environments.

[0060] Example 3: In Example 2, by establishing a quality footprint model and a historical backtracking mechanism, the longitudinal impact of data quality issues within a single device on its own diagnostic conclusions was effectively quantified and corrected. However, this method still treats each device as an independent analysis unit, ignoring the strong physical coupling formed between devices in an actual production line through mechanical connections, fluid pipelines, or structural foundations. Under such coupling, sensor data quality issues in one device not only affect its own diagnosis but also contaminate the data and diagnostic results of downstream related devices through physical paths. Example 2 cannot identify and quantify this cross-device quality risk penetration. This may lead to misjudgments of downstream device faults or make it difficult for maintenance personnel to locate the true source of interference. To assess the comprehensive credibility of diagnostic conclusions at the production line level and achieve optimal allocation of maintenance resources in the coupled network, it is necessary to introduce device coupling modeling and a production line-level collaborative evaluation mechanism based on Example 2.

[0061] In some embodiments, step S42, which automatically triggers the backtracking process for historical time periods, further includes: S421 constructs a graphical model describing the mutual influence of physical quantities across equipment by defining the physical connection topology and coupling transfer function between production line equipment.

[0062] The graphical model of the interaction between physical quantities across equipment abstracts all key equipment involved in monitoring in the production line (such as motors, pumps, and fans) as independent nodes. Based on the actual physical connections between equipment (such as rigid bolt connections, series pipe connections, and shared bases), directed edges are defined between nodes, and the coupling type represented by each edge is labeled (such as mechanical vibration transmission, fluid pressure transmission, and heat conduction). One or more coupling transfer functions are associated with each directed coupling edge (from source equipment A to target equipment B). This function describes how a specific physical state of equipment A (such as vibration acceleration or pressure fluctuations in a specific frequency band) propagates through this physical path and is reflected in the sensor readings of equipment B. These functions can be obtained through initial experiments (such as applying a known excitation to equipment A and measuring the response in equipment B) or through system identification methods based on long-term normal operation data (an engineering method that uses the system's input and output observation data to establish its mathematical model; specifically, by collecting multi-sensor data during long-term normal operation of the equipment, algorithms such as subspace methods, ARX models, or neural networks are used to automatically fit the transfer function relationship between the physical quantities of the source equipment and the sensor readings of the target equipment). The defined nodes, directed edges, and associated coupling transfer functions are constructed into a computable directed graph (i.e., a coupling graph) data structure. Each node in the graph stores a device identifier, and each edge stores the coupling type and the corresponding transfer function model parameters.

[0063] S422, after completing the diagnosis of a single device, traces all its upstream devices, and calculates and superimposes the mass footprints of the upstream devices through the corresponding coupling transfer function to calculate an additional cross-device mass contamination index for the target device.

[0064] The specific calculation method for the cross-device quality pollution index is as follows: The target device (the device currently undergoing fault diagnosis analysis) is set as device B, which has been constructed through step S421 and includes data from any device. Directed edges to device B and their corresponding coupling transfer functions .

[0065] At the calculation time, device B and all its upstream devices All already have the latest local mass footprint vectors maintained by the method in Example 2. and Each dimension of this vector corresponds to a predefined type of quality defect, and its value represents the cumulative influence intensity of that type of defect.

[0066] For each upstream device that can directly affect device B through a physical path (i.e., there are edges in the graph) ): From the equipment mass footprint vector In the process, extract the dimension components that are directly negatively correlated with data reliability (corresponding to labels indicating data problems). Let the extracted vector be... .

[0067] Will Coupled transfer function input to this path The core function of this feature is to attenuate the impact of upstream defects, reflecting the reduction in impact due to factors such as physical distance and connection methods. The output is a attenuated impact vector. Its formula is The specific form of the transfer function T can be simplified to an attenuation coefficient matrix related to the coupling strength.

[0068] All upstream devices The converted influence vector Perform a weighted summation to obtain an aggregated cross-device influence vector. Its formula is Among them, weight Can be based on the edge The coupling type strength or physical distance is set to distinguish the importance of the influence of different upstream devices. The aggregated cross-device influence vector is then used. It is converted into a scalar value, namely the cross-equipment quality pollution index.

[0069] S423 runs a simulation program periodically to virtually repair each unhealthy sensor. It evaluates the improvement effect on the reliability of the entire network diagnosis through a coupled model, generates a repair priority list, and dynamically displays the production line quality status and key contamination paths in a panoramic view.

[0070] S424 After the historical diagnostic conclusion of a certain device is corrected, all its physically coupled downstream devices are automatically identified, and a collaborative backtracking and production line-level credibility reassessment of the historical data of these downstream devices in the same period is triggered.

[0071] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By constructing a physical coupling map of production line equipment and calculating cross-equipment quality contamination indices, this method quantifies for the first time the indirect interference risk of upstream equipment data quality issues on downstream diagnostic conclusions, thereby extending diagnostic reliability assessment from a single point to a collaborative level across the entire production line. This approach not only identifies and warns of contaminated diagnoses and avoids erroneous maintenance, but also intelligently recommends optimal sensor repair paths through simulation analysis, maximizing the overall diagnostic reliability of the production line at minimal cost. Simultaneously, its panoramic visualization and collaborative backtracking mechanism allows historical conclusions to be dynamically updated as the state of the source sensors is corrected, ultimately forming a self-evolving diagnostic ecosystem capable of understanding and optimizing production line-level data health.

[0072] Example 4: In Example 3, by constructing a physical coupling map of the production line and calculating the cross-equipment quality pollution index, the interference of upstream equipment data quality on downstream diagnostic conclusions can be effectively identified. However, this method can only assess the overall existence and risk level of pollution, and cannot quantify the specific contribution of different upstream pollution sources. When downstream equipment signals are affected by multiple upstream equipment, the system struggles to accurately locate the most critical and priority pollution source, leading to potentially inaccurate operation and maintenance decisions and room for improvement in resource allocation efficiency. To achieve quantitative decomposition of pollution sources and provide precise maintenance guidelines for specific equipment and sensors, frequency-domain selective coupling analysis and multi-source signal contribution decoupling techniques need to be introduced based on Example 3.

[0073] In some embodiments, step S422, after completing the single-device diagnostics, traces all its upstream devices, and further includes: 2A establishes a frequency response function for each directed edge in the coupling graph to describe the transmission characteristics of the signal as the frequency changes, and generates a predicted transmission signal as a feature template of the pollution source in the path based on this function and the estimated source signal of the upstream device.

[0074] Specifically, for each directed edge in the coupling graph pointing from the source device X to the target device Y, its corresponding frequency response function is established through offline experiments (hammer impact test, injection of a specific frequency excitation signal) or high-fidelity simulation based on physical mechanisms. The function uses frequency The independent variable is X, and the output is a complex number that represents the amplitude gain and phase shift of the physical path from X to Y for signals of different frequencies.

[0075] In real-time operation, for the upstream device A that needs to be analyzed, the estimated source signal, which has been evaluated and corrected by its local quality status, is invoked from its real-time processing flow. This signal is designed to approximate the actual physical state of device A as closely as possible, and has largely eliminated the influence of its own sensor defects and transmission problems.

[0076] In the frequency domain, the estimated source signal of upstream device A is... Convert to frequency domain The path frequency response function from A to target device B Complex multiplication is performed to obtain the predicted frequency domain conduction signal. Then, it is inversely transformed back to the time domain to obtain the predicted conduction signal. This signal is the path. The pollution characteristic template at the current moment predicts the theoretical form that the state signal of device A should exhibit on the sensor of device B when it is transmitted to device B via this physical path.

[0077] 2B. Time-frequency analysis is performed on the observation signals of downstream equipment, and the equipment is modeled as a linear combination of its own state components and the characteristic templates of each upstream pollution source. The contribution coefficients of each component are obtained by solving the optimization problem, thereby quantifying the specific contribution percentage of each upstream pollution source.

[0078] Specifically, the raw sensor signals of target device B. Performing a short-time Fourier transform yields its two-dimensional observation time-frequency energy distribution. .

[0079] Observation signal This can be expressed as a linear combination model. The model consists of the following components added together: a component to be estimated representing the true state signal of device B, and all upstream devices... Corresponding predicted transmission signal (i.e., pollution feature template) multiplied by the corresponding unknown contribution coefficient. , and a residual term representing noise and unmodeled parts.

[0080] The linear model described above is solved using a numerical optimization algorithm. The goal of this optimization process is to find a set of non-negative contribution coefficients (including the coefficients of device B's own component). and the coefficients of each upstream template This makes the sum of the components weighted by these coefficients most closely approximate the observed signal in the least squares sense. During the solution process, sparsity constraints are usually imposed to make the solution more consistent with the physical assumption that there are only a few major pollution sources.

[0081] Based on the optimal coefficients obtained from the solution, calculate the values ​​for each upstream device. Percentage of contamination contribution to the current observed signal of downstream device B The specific calculation method is as follows: The contribution coefficient of the upstream equipment is... Divide by all contribution coefficients (including and all The sum of these components is multiplied by 100%. Simultaneously, the theoretical proportion of device B's own signal component in the current mixed signal can be calculated.

[0082] 2C, based on the real-time calculated pollution contribution of each path, dynamically highlights the main pollution paths in the situation map and generates precise suggestions for the repair of specific sensors. At the same time, it monitors abrupt changes in the contribution time series to provide early warning of changes in the status of upstream equipment or physical paths.

[0083] In 2D, when conducting collaborative backtracking and reassessment of historical data, a contribution analysis process is embedded to accurately quantify the proportion of pollution contribution of upstream devices to downstream signals at historical moments.

[0084] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By introducing a frequency-domain selective coupling feature library and multi-source signal contribution analysis, the percentage contribution of each upstream device to the downstream signal contamination can be quantified, thereby accurately locating the main interference sources and specific sensors, greatly improving the targeting and efficiency of maintenance. Simultaneously, by monitoring abrupt changes in contamination contribution, it is also possible to proactively warn of abnormal changes in the state of upstream equipment or its physical connection paths, providing unprecedented data-driven insights for the precise maintenance and health management of complex industrial production lines.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for acquiring data from multi-source industrial equipment, characterized in that, include: S1, at the edge of data acquisition, encapsulates an enhanced header for each sensor data packet containing a global serial number, a high-precision timestamp, and auxiliary sensor readings, and transmits it through the industrial network; S2, the host computer reorders and aligns the received data packets according to the timestamp, detects packet loss and severe delay, and generates a preliminary transmission quality label for each data packet; S3 runs a lightweight multi-sensor consistency model to calculate the deviation in the relationship between signals, and analyzes the long-term statistical characteristics and auxiliary readings of each sensor signal to generate sensor health scores and labels. S4 integrates transmission quality tags and sensor health tags, generates repairs for lost but sensor-healthy data, downweights and reconstructs signals to replace data that is complete but sensor-susceptible, and marks channels that are completely unusable. S5 calculates the comprehensive confidence weight for each sensor data, performs weighted processing during the feature fusion stage, and outputs the fault type and diagnostic confidence based on model probability and average data confidence after being input into the diagnostic model.

2. The multi-source industrial equipment data acquisition method according to claim 1, characterized in that, In S1, during data acquisition, the edge computing unit periodically reads the core measurement values ​​from the connected physical sensors, and the core measurement values ​​include the device's status monitoring data. The enhanced packet header specifically includes: an independent, continuously incrementing global sequence number for detecting packet loss for each sensor data stream; A high-precision global timestamp with microsecond-level accuracy generated by the edge side based on a precision clock synchronization protocol; and a snapshot of the sensor's original readings, including the instantaneous power supply voltage and ambient temperature of the sensor channel recorded by the edge computing unit, reflecting its own operating status.

3. The multi-source industrial equipment data acquisition method according to claim 1, characterized in that, The formula for calculating the Kangdu score is as follows: Where C is the consistency health component, P is the ontology feature health component, and α and β are weighting coefficients, satisfying α + β = 1; the formula for calculating C is: ,in This is the deviation value calculated from other related signals by the consistency model at the current moment. It is a preset deviation threshold; the formula for calculating P is: Where n is the total number of features monitored. Let i be the long-term normal value obtained statistically from the sensor's health baseline period for the i-th feature. The value of the i-th feature is obtained statistically within the current monitoring window. This represents the maximum allowed degradation threshold for this feature.

4. The multi-source industrial equipment data acquisition method according to claim 1, characterized in that, In S4, for each sensor data at each time moment, a corresponding processing path is triggered based on its quality-health label combination, including: Path 1: When data is marked as transmission lost and the sensor health label is healthy, a pre-trained conditional generation model is invoked to generate an estimated value for the corresponding data point marked as transmission lost using other healthy and complete sensor data at the same time moment; Path 2: When data arrives completely and the sensor health label is warning or fault, a low initial confidence weight is assigned to the data, and the signal estimate that the sensor should have is reconstructed through the generation model using other highly healthy sensor data at the same time moment to partially correct or replace the original suspicious signal; Path 3: When data is marked as lost and the sensor health label is also fault, the channel data is determined to be completely unusable, its weight is set to zero, and a maintenance alarm is generated.

5. The multi-source industrial equipment data acquisition method according to claim 1, characterized in that, The repair process for generating conditions to initiate the process for lost but healthy data specifically includes: S41, for the various processing markers and sensor health labels generated in S4, predefining their corresponding influence decay functions, and maintaining a time-varying mass footprint vector for each monitored physical quantity; S42, continuously recording the health score timeline of all sensors and all mass label events to form an audit log; when the sensor health status is clearly corrected, automatically triggering the backtracking process for historical time periods; S43, for the original data within the affected time period, using the corrected sensor health information, re-execute steps S1 to S5 in sequence.

6. The multi-source industrial equipment data acquisition method according to claim 5, characterized in that, Maintaining a time-varying mass footprint vector includes: maintaining a mass footprint vector for each independently monitored physical quantity, where each dimension of the vector corresponds to a predefined mass defect type, with an initial value of zero; at each processing moment, generating an influence injection vector based on the processing marker determined for the current mass point and the sensor health label, where the dimension of the vector is the same as the mass footprint vector, and the value of each dimension represents the newly generated influence intensity of the corresponding type of defect at the current moment; and calculating the mass footprint vector at the current moment through iterative updates, with the update formula being: ,in, Let be the quality footprint vector at the current moment. Let this be the mass footprint vector from the previous time step. It is a diagonal matrix, where each element on the diagonal is the attenuation coefficient corresponding to the defect type. Inject a vector into the influence generated at the current moment.

7. The multi-source industrial equipment data acquisition method according to claim 5, characterized in that, The process of triggering a backtracking of a historical time period includes: when the state of any sensor is detected to meet a preset explicit correction condition, capturing the state of the sensor at the time of the correction, including: recording the sensor's identifier, state change type, exact time T of the change, and specific state values ​​before and after the change; determining the corresponding backtracking time depth Δ based on the state correction event and the sensor type, and accordingly determining the historical time window [T-Δ, T] that needs to be re-evaluated; and generating a backtracking task instance based on the recorded state correction event and the determined historical time window.

8. The multi-source industrial equipment data acquisition method according to claim 5, characterized in that, The automatic triggering process for backtracking historical time periods is as follows: S421, by defining the physical connection topology and coupling transfer function between production line equipment, a graph model describing the mutual influence relationship of physical quantities across equipment is constructed; S422, after completing the diagnosis of a single equipment, all its upstream equipment is traced back, and the quality footprint of the upstream equipment is converted and superimposed through the corresponding coupling transfer function to calculate an additional cross-equipment quality pollution index for the target equipment; S423, the simulation program is run regularly to virtually repair each unhealthy sensor, and its improvement effect on the reliability of the whole network diagnosis is evaluated through the coupling model, a repair priority list is generated, and the production line quality status and key pollution paths are dynamically displayed in the panoramic view; S424, after the historical diagnosis conclusion of a certain equipment is corrected, all its physically coupled downstream equipment are automatically identified, and the collaborative backtracking and production line-level reliability reassessment of the historical data of these downstream equipment in the same time period is triggered.

9. The multi-source industrial equipment data acquisition method according to claim 8, characterized in that, The graphical model of the interaction between physical quantities across devices includes: defining directed edges between nodes based on the actual physical connection between devices, and labeling each edge with the coupling type it represents; associating each directed edge with a coupling transfer function, which describes the relationship between a specific physical state from the source device to the target device propagating through the physical path and affecting the sensor readings of the target device; and constructing the nodes, directed edges, and their associated coupling transfer functions into a computable directed graph data structure.

10. The multi-source industrial equipment data acquisition method according to claim 8, characterized in that, The process of tracing all upstream devices after completing single-device diagnosis specifically includes: establishing a frequency response function for each directed edge in the coupling graph to describe the transmission characteristics of the signal as the frequency changes, and generating a predicted transmission signal as a feature template of the pollution source for that path based on this function and the estimated source signal of the upstream device; performing time-frequency analysis on the observed signal of the downstream device and modeling it as a linear combination of its own state components and each upstream pollution feature template, obtaining the contribution coefficient of each component by solving an optimization problem, thereby quantifying the specific contribution percentage of each upstream pollution source; dynamically highlighting the main pollution paths in the situation map based on the real-time calculated pollution contribution of each path and generating precise suggestions for repairing specific sensors, while monitoring abrupt changes in the contribution time series to warn of state changes of upstream devices or physical paths; and embedding a contribution analysis process during the collaborative retrospective reassessment of historical data to accurately quantify the proportion of pollution contribution of upstream devices to downstream signals at historical moments.

Images