Multi-point positive pressure monitoring system based on IoT data acquisition
By establishing a dynamic pressure cause-effect graph to optimize the data acquisition of the multi-point positive pressure monitoring system, the problem of unprocessed correlation between monitoring points was solved, and the communication load was optimized and the data transmission continuity was achieved.
Patent Information
- Application Number
- CN202611211808.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-06
AI Technical Summary
Existing multi-point positive pressure monitoring systems lack a joint processing mechanism for the sequential relationship of changes between monitoring points, response time delays, and network reporting scheduling, resulting in increased communication load, data loss, and difficulty in reconstructing the propagation process.
By establishing a dynamic pressure causal graph, based on the prediction residuals and historical sequences, the set of associated nodes is determined, and staggered sampling phases and ordered reporting time slots are configured to optimize the data acquisition process.
During the stable phase, reduce duplicate data uploads, maintain the temporal continuity of the anomaly propagation process, reduce communication load, and avoid data loss and delays.
Smart Images
Figure CN123055604A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet of Things (IoT) communication and networked data acquisition technology, specifically a multi-point positive pressure monitoring system based on IoT data acquisition. Background Technology
[0002] Existing multi-point positive pressure monitoring systems typically consist of distributed positive pressure data acquisition nodes, a communication network, and a central processing platform. Each acquisition node acquires positive pressure data according to a preset sampling period and encapsulates the sampled value, node identifier, and local time information into an IoT data acquisition message, which is then periodically uploaded to the central processing platform via the network. The central processing platform stores, sorts, and performs threshold comparisons on the data from each node. When the monitored value exceeds a set range or the rate of change reaches a set condition, an abnormal event is generated, and control commands are issued to the relevant nodes to increase the sampling frequency or shorten the reporting period. To reduce the amount of communication data, some solutions calculate the mean, extreme values, or rate of change of continuous sampled values at the node side and only upload the statistical results; other solutions use historical data to build a single-point prediction model, reducing the upload of raw data when the deviation between the measured value and the predicted value is small, and restoring high-frequency transmission when the deviation increases. For situations where multiple monitoring points trigger anomalies simultaneously, existing platforms typically process them sequentially according to the order of message arrival, or uniformly configure the same acquisition period and reporting period for all triggering nodes, without continuously modeling the sequence of changes, response delays, and propagation directions among the monitoring points.
[0003] Data acquisition control is based independently on the numerical threshold or prediction deviation of a single monitoring point, failing to utilize the correlation paths between multiple positive pressure changes to determine the acquisition targets, acquisition timing, and reporting order. When the positive pressure is stable, each node continues to repeatedly upload data with minor changes according to a fixed cycle; when the local pressure state changes, nodes related to and unrelated to anomaly propagation may simultaneously switch to high-frequency acquisition, resulting in multiple messages entering the same communication link and causing queuing, delays, or loss. If only the reporting volume is limited, continuous change segments between the anomaly source node and subsequent response nodes are easily lost, making it difficult to reconstruct the propagation sequence of positive pressure changes; if high-granularity acquisition is continuously applied to all nodes, the communication load increases in the long term. The fundamental reason is that the existing system lacks an acquisition mechanism that jointly processes prediction residuals, response delays between monitoring points, and network reporting scheduling, failing to retain the correlation data required for the anomaly propagation process while controlling redundant data volume. Summary of the Invention
[0004] The purpose of this invention is to provide a multi-point positive pressure monitoring system based on Internet of Things (IoT) data acquisition, which can solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A multi-point positive pressure monitoring system based on IoT data acquisition includes multiple positive pressure data acquisition nodes, an IoT data acquisition coordination terminal, and a data processing terminal. The positive pressure data acquisition node generates a status summary based on the current positive pressure value, change slope, fluctuation intensity and short-term trend, and obtains the prediction residual based on the prediction model; The data processing terminal establishes a dynamic pressure causal graph based on the historical sequence of each positive pressure data acquisition node, representing the sequence of changes, response time delay, and residual propagation direction. The IoT data acquisition coordination terminal receives a status summary when the prediction residual is within the confidence interval. The data processing terminal reconstructs the unuploaded segment based on the status summary. When the prediction residual exceeds the confidence interval, the associated node set is determined based on the dynamic pressure cause-effect graph. The terminal then issues an acquisition task containing a unified event window, staggered sampling phase, and ordered reporting time slot to the associated node set.
[0007] Preferably, the data processing terminal performs sliding alignment of the historical sequence according to the candidate time delay between monitoring point pairs, calculates the directional consistency, time delay stability and repetitive event overlap between the change segment of the preceding sequence and the prediction residual of the following sequence, and forms a directed correlation weight based on the three. The monitoring point pairs whose directed correlation weights meet the edge establishment conditions are written into the dynamic pressure cause-effect graph. For the monitoring point pairs that simultaneously meet the edge establishment conditions in opposite directions, the direction with small residual dispersion is retained, and the corresponding candidate time delay is recorded as the response time delay.
[0008] Preferably, the positive pressure data acquisition node divides the continuous sampling sequence into adjacent state segments, extracts the endpoint value, segment slope, local fluctuation amount, sequence continuity identifier and model version number for each state segment, and combines them to form the state summary; The predicted residual is formed by weighting the differences between the measured endpoint values of the state segment and the output endpoint values of the prediction model, the differences in the segmented slope, and the differences in local fluctuations. The confidence interval is updated on a rolling basis according to the residual distribution of continuous stable state segments under the same model version.
[0009] Preferably, the IoT data acquisition coordination terminal uses the positive pressure data acquisition node whose predicted residual exceeds the confidence interval as the event source node, expands layer by layer along the directed edge in the dynamic pressure causal graph that is consistent with the direction of residual propagation, and stops expanding when the cumulative response delay exceeds the event window length or the directed association weight is lower than the expansion condition, so as to obtain the set of associated nodes. Then, staggered sampling phases are allocated according to the graph path hierarchy of each associated node relative to the event source node, and orderly reporting time slots are configured according to the sampling phase order.
[0010] Preferably, the data processing terminal maintains an edge state record for each directed edge, consisting of the number of valid events, time delay dispersion, number of direction reversals, and residual propagation consistency. After a new event ends, the measured change segments of each monitoring point within the event are matched with the edge state records. For directed edges that match, the response delay and the directed association weight are updated. For directed edges that show a reversal of direction or a continuous increase in delay dispersion, the weight is decayed. When the weight decays to the edge removal condition, the historical edge identifier of the corresponding monitoring point pair is retained and the real-time extension qualification is deleted.
[0011] Preferably, after receiving the state summary, the data processing terminal calls the corresponding prediction model according to the model version number, and reconstructs the unuploaded state fragments using the endpoint values and segment slopes of adjacent state summaries as boundary constraints. When the sequence continuity identifiers are discontinuous, the model version numbers of adjacent state summaries are inconsistent, or the verification residual between the reconstructed fragment and the subsequent measured endpoint value exceeds the verification interval, an original data supplementation instruction pointing to the missing sequence range is generated, and the reconstructed fragment, the supplementation fragment, and their respective source identifiers are associated and written into the same event data chain.
[0012] Preferably, the IoT data acquisition coordination terminal constructs a sampling phase sequence based on the graph path level, expected message length, and current link pending transmission volume of each positive pressure data acquisition node in the associated node set, so that the sampling time of adjacent graph path levels keeps covering the overlapping interval of the corresponding response delay; The sampled phase sequence is then mapped to the cyclic reporting time slot. Nodes whose expected message length exceeds the capacity of a single time slot are split into multiple messages with the same event number and consecutive fragment numbers, and the multiple messages are configured in non-adjacent time slots.
[0013] Preferably, the IoT data acquisition coordination terminal continuously receives the prediction residuals and sequence continuity identifiers of the associated node set within the unified event window, and updates the event window boundary based on the residual fallback status of the event source node, the residual arrival status of the downstream node, and the number of incomplete reporting messages. When the residual of the event source node falls back and the downstream node has not yet reached the corresponding response time delay, the event window is extended. When the residuals of all associated nodes return to the reliable range and the number of unreported messages is 0, the event window ends and the status summary reporting method of the end node is restored.
[0014] Preferably, the data processing terminal includes the monitoring points whose real-time expansion eligibility has been deleted into the candidate recovery set according to the historical edge identifier, and extracts the change segments that are consistent with the original residual propagation direction in subsequent events; When the monitoring point pair in the candidate recovery set forms a stable response time delay in multiple discontinuous events and the number of direction reversals does not increase, the directed association weight is reconstructed based on the residual propagation consistency of the new events, and the reconstructed directed edge is set as an observation edge that only participates in the verification of the associated node set. After the observed edge passes event verification consecutively, it is then written into the real-time expansion path of the dynamic pressure cause-effect graph.
[0015] Preferably, the IoT data acquisition coordination terminal maintains an event control record for each event window, including the node entry order, sampling phase version, message completion identifier, and window boundary version. When the event window is extended, the sampling phase is recalculated only for the associated nodes that have not yet covered the corresponding response delay, and the original time slot mapping of the completed message is retained; Before the event window ends, the event control record and the event data chain received by the data processing terminal are checked for sequence closure. For nodes with node entry sequence breakpoints or missing message completion identifiers, a supplementary sampling task with a limited sequence range is issued.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By establishing a dynamic pressure causal graph that includes the sequence of changes, response delays, and residual propagation directions from the historical sequences of each positive pressure data acquisition node, and using the predicted residuals at the node side as the basis for switching acquisition modes, the system can receive state summaries and reconstruct unuploaded segments when the positive pressure is stable. When the predicted residuals exceed the confidence interval, the system only determines the set of associated nodes along the causal graph. Associated nodes are configured with staggered sampling phases and ordered reporting time slots according to the graph path hierarchy, ensuring that the data from the anomaly source node and subsequent response nodes cover the corresponding propagation delays, while simultaneously avoiding unassociated nodes from synchronously increasing the acquisition granularity. This forms an adaptive IoT data acquisition process based on pressure association paths, reducing duplicate messages during the stable phase while maintaining the temporal continuity of anomaly propagation segments, and reducing message queuing, delays, and missing messages caused by centralized reporting.
[0017] 2. The state summary sets endpoint values, segment slopes, local fluctuations, sequence continuity identifiers, and model version numbers. This enables the data processing end to reconstruct unuploaded state segments using a consistent model version and perform verification using adjacent segment boundaries and subsequent measured endpoints. If the verification fails, the system generates a data acquisition task based on the missing sequence range and writes the reconstructed segment, the acquired segment, and the source identifier into the same event data chain. Directed edges in the dynamic pressure cause-effect graph are updated, decayed, or restored based on event matching results. The unified event window adjusts its boundaries based on the event source node residuals, downstream node response status, and message completion status, and performs sequence closure verification before completion. This process maintains consistency between the associated model, acquisition tasks, and event data, reducing interference from model version mismatches, sequence breakpoints, and incomplete messages on subsequent data analysis. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the overall process of multi-point positive pressure monitoring based on IoT data acquisition. Figure 2 To predict the variation curves of residuals and confidence intervals; Figure 3 Reconstruct the curve for boundary constraints of unuploaded state fragments; Figure 4 This is a diagram showing the allocation of sampling phase and reporting time slots. Detailed Implementation
[0019] In one embodiment, a multi-point positive pressure monitoring system based on IoT data acquisition includes multiple positive pressure data acquisition nodes, an IoT data acquisition coordination terminal, and a data processing terminal. The positive pressure data acquisition nodes generate a state summary based on the current positive pressure value, change slope, fluctuation intensity, and short-term trend, and obtain prediction residuals based on a prediction model. The data processing terminal establishes a dynamic pressure causality graph representing the sequence of changes, response time delay, and residual propagation direction based on the historical sequences of each positive pressure data acquisition node. The IoT data acquisition coordination terminal receives the state summary when the prediction residual is within a confidence interval, and the data processing terminal reconstructs the unuploaded segment based on the state summary. When the prediction residual exceeds the confidence interval, the data processing terminal... The set of associated nodes is determined based on the dynamic pressure causal graph, and a collection task containing a unified event window, staggered sampling phase, and ordered reporting time slots is issued to the set of associated nodes. Each positive pressure data collection node corresponds to a positive pressure monitoring location, and each node continuously forms a positive pressure sequence with node identifier, collection sequence number, and local time index. The IoT data collection coordination terminal is responsible for establishing linkage control between the node-side collection method and the network-side reporting method. The data processing terminal is responsible for saving the original data, status summary, reconstructed data, event data chain, and dynamic pressure causal graph, so that the summary reporting in the stable phase, the associated additional collection in the abnormal phase, and the model update after the event ends constitute a continuous data processing closed loop.
[0020] In this embodiment, please refer to Figure 1 The positive pressure data acquisition node performs duplicate removal, continuity checks, and time index correction on continuous sampled values. When removing duplicates, it retains the earliest arriving record with consecutive sequence numbers. The continuity check is based on whether adjacent sequence numbers are increasing and whether adjacent time indices remain monotonic. The time index correction only handles local reversals within a node without changing the relative propagation relationship between different nodes. The processed positive pressure sequence is divided into multiple adjacent state segments, each with a unified start and end index. The node calculates the endpoint value, segment slope, local fluctuation amount, and short-term trend direction for each state segment and combines the above information with the model version number, sequence continuity identifier, and node identifier to form a state summary. After receiving the state summary, the IoT data acquisition coordination terminal first checks the model version number against the version mapping currently stored at the data processing terminal. If the versions match, it enters the reconstruction judgment; if the versions do not match, it retains the summary but does not perform reconstruction and returns the model identifier of the specified version to the corresponding node to prevent summaries generated by different models from being spliced into the same continuous sequence.
[0021] Furthermore, the data processing end performs sliding alignment of the historical sequence according to the candidate time delays between monitoring point pairs, calculates the directional consistency, time delay stability, and repetitive event overlap between the change segment of the preceding sequence and the predicted residual of the following sequence, and forms a directed correlation weight based on the three. Monitoring point pairs whose directed correlation weights meet the edge establishment conditions are written into the dynamic pressure causal graph. For monitoring point pairs that meet the edge establishment conditions in opposite directions, the direction with smaller residual dispersion is retained, and the corresponding candidate time delay is recorded as the response time delay. Specifically, the data processing end extracts the change segments that exceed their respective local fluctuation baselines from the historical sequences of two monitoring points, takes the earlier change segment as the preceding candidate, and takes the change segments with the same or opposite directions but stable morphology in the subsequent time range as the following candidate. Then, the directed edge is determined by whether the order of events in multiple events is repeated, whether the candidate time delays are concentrated, and whether the change segments fall into the same event window, so as to avoid establishing a fixed correlation based on a single synchronous change.
[0022] The directed association weights are calculated according to the following formula:
[0023] ; in, This represents the directed association weight from monitoring point i to monitoring point j. This indicates the degree of consistency in the direction of change among multiple historical events, where monitoring point i changes first and monitoring point j changes later. This indicates the degree to which candidate response delays are concentrated around the median delay. This indicates the overlap of repeated events where the change segments of two monitoring points are grouped into the same event window. , and The combination coefficients of the three indicators remain unchanged within the same graph version; for example, when the directional consistency is 0.8, the time delay stability is 0.7, and the repetitive event overlap is 0.9, and the three combination coefficients are 0.4, 0.3, and 0.3 respectively, the directed association weight is 0.8. If the reverse association weight is 0.62 and the reverse residual dispersion is larger, then only the direction from monitoring point i to monitoring point j is retained, so that the subsequent association node expansion proceeds along the recurring propagation direction.
[0024] In a preferred embodiment, candidate time delays are not determined all at once over the entire historical sequence. Instead, time delay samples are generated separately for each event window, and the median value of the time delay samples is used as the response time delay. The degree to which the time delay samples deviate from the median value forms the time delay stability. When the directed correlation weights in both directions meet the edge establishment conditions, the data processing end calculates the predicted residual dispersion of the follow-up sequence in both directions. The direction with smaller predicted residual dispersion indicates that the follow-up change can be explained more stably by the preceding change. Therefore, this direction is retained, and the other direction is registered as a suppressed candidate edge. If the predicted residual dispersion of the two directions is close, the real-time extended edge is not established for the time being. Instead, the monitoring point pair is placed in the observation set and recalculated after subsequent events are added to prevent short-term common disturbances from being mistaken for stable propagation paths.
[0025] In one embodiment, the positive pressure data acquisition node divides the continuous sampling sequence into adjacent state segments, extracts the endpoint value, segment slope, local fluctuation amount, sequence continuity identifier, and model version number for each state segment, and combines them to form the state summary; the prediction residual is formed by weighting the differences between the measured endpoint value and the predicted model output endpoint value of the state segment, the differences in segment slope, and the differences in local fluctuation amount; the confidence interval is updated on a rolling basis according to the residual distribution of continuous stable state segments under the same model version; wherein, the division boundary of the state segment is triggered by any of the following conditions: change direction switching, local fluctuation abrupt change, or preset segment length upper limit; change direction switching is used to maintain the consistency of the trend within a single segment; local fluctuation abrupt change is used to isolate transient disturbances; segment length upper limit is used to avoid long-term stable sequences forming excessively large summary spans; the sequence continuity identifier is composed of the start sequence number, the end sequence number, and the set of missing positions.
[0026] The predicted residual is calculated according to the following formula:
[0027] ; in, This represents the prediction residual corresponding to the k-th state segment of monitoring i. and These represent the measured end value and the predicted end value of the state segment, respectively. and These represent the measured piecewise slope and the predicted piecewise slope, respectively. and These represent the measured local fluctuation and the predicted local fluctuation, respectively. , and These are fixed residual combination coefficients under the same model version; for example, when the three absolute differences are 0.6, 0.2 and 0.1 respectively, and the three combination coefficients are 0.5, 0.3 and 0.2 respectively, the prediction residual is 0.38. If the upper bound of the confidence interval corresponding to the same model version is 0.45, the node maintains the state summary report. If the prediction residual of subsequent segments becomes 0.57, the node enters the candidate state for uploading raw data.
[0028] Please refer to Figure 2 , Figure 2 The horizontal axis represents the display sequence number of consecutive state segments, showing the most recent 120 summary segments; the vertical axis represents the prediction residuals and their confidence interval boundaries. The black solid line is the curve of the prediction residual change for each segment, the gray dashed line is the upper boundary of the confidence interval, and the gray dotted line is the lower boundary of the confidence interval. The three curves together represent the comparison result of the deviation between the node-side state summary and the prediction model at the data processing end.
[0029] from Figure 2 As can be seen, the minimum predicted residual within the display interval is approximately 0.0336, the maximum is approximately 0.4091, and the average is approximately 0.1814. The upper boundary of the confidence interval rolls with the residual distribution of stable segments under the same model version, ranging from approximately 0.4045 to 0.4473, while the lower boundary remains near 0.0, reflecting the engineering constraint of residuals as a non-negative deviation.
[0030] Figure 2 A local peak appears at the 18th displayed segment, with the residual slightly higher than the corresponding upper boundary; most other segments remain within the confidence interval, indicating that most changes in barotropic conditions can be compressed and expressed through summary reporting and model prediction, while a few out-of-bounds segments need to be registered as event source candidates by the coordination end. The segments shown cover multiple barotropic acquisition nodes such as PZ-08, PZ-09, PZ-10, PZ-11, and PZ-12, reflecting the residual comparison method when different nodes share the same model version in a multi-point monitoring scenario.
[0031] Preferably, the confidence interval is maintained separately by the data processing end for each monitoring point and each model version. The data processing end extracts prediction residual samples from continuous stable state segments. A stable state segment refers to a segment without sequence missing, model version switching, or cross-point event markers. Then, the upper and lower boundaries are formed according to the median value and median absolute deviation of the residual samples, and updated with a sliding sample set when a new stable segment arrives. If the model version changes, the confidence interval of the old version continues to be bound to the old version summary, and the new version establishes an independent confidence interval from the initialization sample. The reconstructed data of the old version must not enter the residual sample set of the new version. When the prediction residual is close to the upper boundary of the confidence interval and multiple consecutive segments are in an increasing state, the node keeps reporting the summary and adds a trend warning mark. The coordination end checks the availability of relevant directed edges in advance based on this, but does not immediately switch the collection method for all nodes.
[0032] In one embodiment, the IoT data acquisition coordination terminal uses positive pressure data acquisition nodes whose predicted residuals exceed the confidence interval as event source nodes. It expands layer by layer along the directed edges in the same direction as the residual propagation in the dynamic pressure causal graph, and stops expanding when the cumulative response delay exceeds the event window length or the directed association weight is lower than the expansion condition, to obtain the set of associated nodes. Then, it allocates staggered sampling phases according to the graph path hierarchy of each associated node relative to the event source node, and configures orderly reporting time slots according to the sampling phase order. In specific implementation, the coordination terminal records the event source node as layer 0, the nodes that can be directly reached by the layer 0 nodes through available directed edges as layer 1, and the nodes that can only be reached by the layer 1 nodes as layer 2. The expansion is repeated until all candidate paths meet the stopping condition. When the same node is reached by multiple paths, the path with the smaller cumulative response delay and the higher path association weight is retained to avoid the same node being repeatedly issued acquisition tasks.
[0033] The path determination of associated nodes is performed according to the following formula:
[0034] ; in, This represents the path affinity from the event source node s to the candidate node v. This represents the directed path from the event source node to the candidate node. This represents the directed association weight from monitoring point a to monitoring point b in the path. This indicates the cumulative response time delay of the path. This represents the response delay of the corresponding directed edge record. For example, a path includes two directed edges with directed association weights of 0.8 and 0.75, a path association degree of 0.6, and two response delays of 3 and 4 sampling intervals, respectively. The cumulative response delay is 7 sampling intervals. When the event window length covers 8 sampling intervals and the expansion condition allows a path association degree of 0.6 to enter, the candidate node is included in the associated node set. When the cumulative response delay exceeds the event window length after adding a third edge, the expansion stops.
[0035] Furthermore, after the set of associated nodes is formed, the coordinating end generates a unified event number. The event number is bound to the event source node, graph version, model version set, and event starting index. The target segment range, sampling phase, reporting time slot, and data level of each associated node are written in the acquisition task. The data level at least distinguishes between state summary, original continuous segment, and missing segment supplementation. The event source node uploads the original continuous segment by default. Nodes that are close to the event source node in the path level and have a high path correlation upload the original continuous segment. Nodes located at the end of the extension and whose current prediction residual is still within the confidence interval can first upload the state summary with increased density. If its prediction residual subsequently exceeds the confidence interval, it switches to the original continuous segment, so that the acquisition granularity is adjusted according to the propagation state.
[0036] In one embodiment, the data processing terminal maintains an edge state record for each directed edge, consisting of the number of valid events, time delay dispersion, number of direction reversals, and residual propagation consistency. After a new event ends, the measured change segments of each monitoring point within the event are matched with the edge state record. For directed edges that match, the response time delay and directed association weights are updated. For directed edges that exhibit direction reversals or continuously increasing time delay dispersion, weight decay is performed. When the weight decays to the edge removal condition, the historical edge identifier of the corresponding monitoring point pair is retained while the real-time extension qualification is deleted. Among these, the number of valid events only counts events with consecutive sequences and consistent model versions. The time delay dispersion is formed by the deviation of each event's response time delay relative to the median time delay. The number of direction reversals counts events where the subsequent node changes before the preceding node. The residual propagation consistency is used to record whether the subsequent node exhibits a same-direction residual change within the response time delay range after the preceding node's residual change.
[0037] Preferably, directed edge updates adopt an event-level incremental recording method. The data processing end does not directly overwrite the old response delay and directed association weights. Instead, it writes the direction consistency, delay samples, and event overlap identifiers of the new events into the edge state record, and then recalculates based on the most recent set of consecutive events. When a directed edge is removed from the real-time expansion qualification, the historical edge identifier continues to save the monitoring point pair, the original propagation direction, the historical median delay, the last available graph version, and the reason for deletion. The dynamic pressure cause-effect graph does not pass through the directed edge during online expansion, but the offline matching after the event ends still checks the corresponding monitoring point pair, which facilitates the subsequent identification of propagation relationship recovery or changes in operating status, and avoids permanent edge removal due to a single abnormal event.
[0038] In one embodiment, after receiving the state summary, the data processing terminal calls the corresponding prediction model according to the model version number, and reconstructs the unuploaded state segments using the endpoint values and segment slopes of adjacent state summaries as boundary constraints. When the sequence continuity identifier is discontinuous, the model version numbers of adjacent state summaries are inconsistent, or the verification residual between the reconstructed segment and the subsequent measured endpoint value exceeds the verification interval, an original data supplementation instruction pointing to the missing sequence range is generated, and the reconstructed segment, the supplemented segment, and their respective source identifiers are associated and written into the same event data chain. The prediction model can be a segmented trend model trained from historical stable segments of the same monitoring point. The model input includes the end value, end slope, local fluctuation amount, and time span of the previous state segment, and the model output includes the predicted end value, predicted slope, and predicted fluctuation amount of the next state segment. The model update only uses data whose source identifier is the original measured data and which has passed the sequence continuity verification.
[0039] Unuploaded state fragments are reconstructed using boundary constraint interpolation:
[0040] ; in, This represents the reconstructed positive pressure value of monitoring point i at time t before data upload. and These represent the start and end endpoint values of the segment, respectively. and These represent the initial slope and the final slope of the segment, respectively. Indicates the start time of the segment. The duration of the segment is represented by u, and the normalized time is represented by . , , , The boundary constraint basis function is used; for example, the starting and ending values of the segment are 30 and 32, the starting slope and ending slope are 0.2 and 0.1, the segment duration is 4 sampling intervals, and when u is 0.5 at the middle time, the reconstruction result is constrained by both the endpoint values and the endpoint slopes, avoiding ignoring the trend changes at both ends of the segment by simply connecting them with a straight line.
[0041] Please refer to Figure 3 , Figure 3 The horizontal axis represents the sampling time sequence number within the unuploaded state segment, ranging from 0 to 80; the vertical axis represents the reconstructed positive pressure value. The solid black line is the boundary constraint reconstruction curve, the dashed gray line is the linear reference curve, and the hollow circle indicates the start and end point verification positions of the segment. The legend is used to distinguish between the reconstruction result, the linear reference, and the end point constraints.
[0042] from Figure 3 As can be seen, the reconstructed curve is equal to 30.0 Pa at the starting point and 32.0 Pa at the ending point, which is consistent with the endpoint values in the state summary; the reconstructed value at the midpoint is about 31.05 Pa, which is slightly higher than the linear reference value of 31.00 Pa, indicating that the curve is not a simple straight line connection, but is constrained by the boundary information of the starting slope of 0.2 and the ending slope of 0.1.
[0043] The maximum difference between the reconstructed curve and the linear reference is approximately 0.1693 Pa, occurring near the 61st sampling time. While small, this difference is sufficient to reflect the smooth transition of pressure changes within the state segment. Because the curve simultaneously satisfies endpoint value and trend direction constraints at both ends, the data processing unit can recover the continuous positive pressure trajectory used for trend analysis, model updates, and event chain verification even without receiving the complete original sequence.
[0044] In a preferred embodiment, after the data processing end completes the reconstruction, it does not immediately mark the reconstructed segment as complete data. Instead, it waits for the subsequent original measured endpoints to arrive and calculates the verification residual. The verification residual is formed by the difference between the predicted value at the end of the reconstructed segment and the value of the subsequent measured endpoint, and the difference between the slope at the end of the reconstruction and the slope of the subsequent measured endpoint. If the verification residual is within the verification interval and the sequence is continuously closed, the reconstructed segment is marked as usable for event correlation analysis but not for model training. If the verification residual exceeds the verification interval, a supplementary acquisition instruction containing the missing start sequence number, the missing end sequence number, the model version number, and the event number is generated. After receiving the supplementary acquisition instruction, the node extracts the original data from the local rolling cache. If the cache range is insufficient, it returns the range that can be supplemented and the range that cannot be recovered. Based on this, the data processing end clearly records the data boundaries in the event data chain.
[0045] In one embodiment, the IoT data acquisition coordination terminal constructs a sampling phase sequence based on the graph path level, expected message length, and current link pending transmission volume of each positive pressure data acquisition node in the associated node set, so that the sampling time of adjacent graph path levels keeps covering the overlapping interval of the corresponding response delay; then the sampling phase sequence is mapped to the cyclic reporting time slot, and nodes whose expected message length exceeds the capacity of a single time slot are split into multiple messages with the same event number and consecutive fragment sequence number, and multiple messages are configured in non-adjacent time slots; specifically, the event source node obtains the earliest sampling phase, the sampling phase of the first layer node is expanded around its response delay with the event source node, and the second layer and subsequent nodes start sampling before the sampling window of the previous layer node ends, so as to retain the overlapping part of the change fragment of adjacent levels, and the link pending transmission volume is used to avoid configuring nodes with large message backlog in adjacent reporting time slots.
[0046] The sampling phase and reporting time slot are formed according to the following formula:
[0047] ; in, This represents the sampling phase of the associated node v relative to the start time of the event. This represents the cumulative response time lag from the event source node to the associated nodes. This represents the overlap time reserved to cover the frontier of transmission. This indicates the cyclic reporting time slot sequence number corresponding to the associated node. This indicates the scheduling offset based on the expected message length and the current amount of data to be sent on the link. This indicates the time width of a single cyclic reporting slot. This indicates rounding down; for example, if the cumulative response delay is 7 sampling intervals and the overlap duration is 2 sampling intervals, then the sampling phase starts from the 5th sampling interval. If the scheduling offset is 1 sampling interval and the width of a single time slot is 2 sampling intervals, then the time slot number is 3, so that the sampling window covers the data before and after the predicted propagation arrives and avoids the concentrated transmission period of the preceding messages.
[0048] Please refer to Figure 4 , Figure 4 The horizontal axis represents the associated positive pressure acquisition node entering the event control record, while the vertical axis displays both the sampling phase and the reporting time slot number. Light-colored diagonal bars represent the sampling phase obtained based on the cumulative response delay and the propagation overlap section, while the black broken line uses square markers to represent the mapped reporting time slot. Figure 4 The CCP displayed 15 node-level scheduling records, covering three event windows: EV-20260728-001, EV-20260728-002, and EV-20260728-003.
[0049] from Figure 4 As can be seen, the sampling phase ranges from 0 to 5, the reporting time slot number ranges from 0 to 4, and the single time slot width is 2. In EV-20260728-001, the sampling phase of the event source node PZ-01 is 0, and the reporting time slot is 0; the sampling phase of the first-level associated node PZ-02 is 1, and the reporting time slot is 1; the sampling phase of subsequent nodes PZ-03, PZ-04, and PZ-05 is 5, and they fall into different reporting time slots according to the scheduling offset.
[0050] The bar chart phase and the reported time slot on the broken line do not completely coincide, indicating that the coordinating end not only determines the sampling start position based on the response delay, but also allocates the collected results to non-collision time slots by combining the cyclic reporting interval, scheduling offset, and the amount of data to be sent on the link. This allows adjacent propagation layers to retain necessary time overlap, while avoiding multiple associated nodes from uploading in the same communication window, reducing the risk of instantaneous queuing and message loss in IoT links.
[0051] Furthermore, the expected message length is determined by the target fragment range, data level, and fragment header information. The current link pending transmission quantity is obtained by the coordinator accumulating the event messages that have not yet been confirmed. Each fragment message carries an event number, node identifier, original message sequence number range, fragment sequence number, and total number of fragments. The coordinator configures multiple fragments of the same node in non-adjacent time slots and reserves retransmission positions in subsequent time slots. When a fragment does not return a completion flag, only the missing fragment is retransmitted without retransmitting the complete message. The data processing end splices fragments according to the original message sequence number range. If the fragments arrive in different orders, they are rearranged according to the fragment sequence number. Only when the number of fragments is consistent with the total number of fragments and the original sequence range is continuous will the corresponding fragment be written into the event data chain.
[0052] In one embodiment, the IoT data acquisition coordination terminal continuously receives the predicted residuals and sequence continuity identifiers of the associated node set within a unified event window, and updates the event window boundary based on the residual fallback status of the event source node, the residual arrival status of the downstream node, and the number of incomplete reporting messages. When the residual of the event source node falls back but the downstream node has not yet reached the corresponding response delay, the event window is extended. When the residuals of all associated nodes return to the reliable interval and the number of incomplete reporting messages is zero, the event window ends, and the status summary reporting method of the end node is restored. The initial length of the event window is jointly determined by the maximum cumulative response delay in the associated node set, the current abnormal segment length of the event source node, and the reserved verification interval. When updating the window, the end boundary is only moved backward, and the event start boundary is not rolled back, so as to maintain the stability of the data index within the same event.
[0053] The event window end boundary is updated according to the following formula: ; in, Indicates the end time of the updated event window. Indicates the end time of the event window before the update. This indicates the start time of the residual out-of-bounds error at the event source node. Represents the current set of associated nodes. Indicates the relationship between the event source node and related nodes. Cumulative response time delay, This indicates the end time of the data corresponding to the most recent unfinished message. This indicates the additional observation time required for closure verification. For example, if the original end time is the 20th sampling time, the start time of the event source node is the 10th sampling time, the maximum cumulative response delay is 8 sampling intervals, the end time of the data corresponding to the most recently incomplete message is the 19th sampling time, and the additional observation time is 3 sampling intervals, then the updated end time is the 22nd sampling time to cover messages that have not yet been closed and their subsequent verification data.
[0054] Preferably, the coordinator maintains an in-event state for each associated node. The in-event state includes waiting for propagation, collecting data, reporting data, waiting for closure, and completed. After the event source node enters the collecting data state, subsequent nodes in the path enter the waiting for propagation state based on the cumulative response time delay. When its sampling phase arrives, it enters the collecting data state. After the target segment is completed, it enters the reporting data state. After all segments are completed and the prediction residuals return to the confidence interval, it enters the waiting for closure state. If the residuals of the event source node have fallen back but a downstream node is still in the waiting for propagation state, the coordinator extends the event window based on its cumulative response time delay. If all nodes enter the waiting for closure or completed state and the number of incomplete messages is zero, then the event data link closure verification is performed and the status summary reporting is restored.
[0055] In one embodiment, the data processing end includes monitoring point pairs whose real-time expansion eligibility has been deleted according to historical edge identifiers into the candidate recovery set, and extracts change segments consistent with the original residual propagation direction in subsequent events; when the monitoring point pairs in the candidate recovery set form a stable response delay in multiple discontinuous events and the number of direction reversals does not increase, the directed association weights are reconstructed based on the residual propagation consistency of the new events, and the reconstructed directed edges are set as observation edges that only participate in the verification of the associated node set; after the observation edges pass the event verification consecutively, they are written into the real-time expansion path of the dynamic pressure causal graph; wherein, discontinuous events refer to events whose event numbers are not adjacent or whose event windows have stable state segments, in order to avoid the same long event being split and counted repeatedly, the observation edges are only allowed to add candidate nodes to the verification set, and are not allowed to continue to expand from the candidate nodes to deeper levels.
[0056] Furthermore, the observation edge being verified by the event means that after the event source node triggers along the existing real-time expansion path, the candidate nodes connected by the observation edge show a consistent change segment in the direction near the historical median time delay, and the data sequence returned by the candidate nodes is continuous, the model version is consistent, and the change in the prediction residual is consistent with the original propagation direction. If the observation edge reverses direction or the time delay dispersion increases again during the verification period, the observation count is cleared and the historical edge identifier is retained. If multiple consecutive independent events pass the verification, the observation edge is restored as a real-time expansion edge and the path correlation is recalculated. The restored directed edge is set with a new graph version number, and the old graph version continues to be bound to the existing event data chain to avoid historical events changing the original propagation interpretation after the graph is updated.
[0057] In one embodiment, the IoT data acquisition coordination terminal maintains an event control record for each event window, including the node entry order, sampling phase version, message completion identifier, and window boundary version. When the event window is extended, the sampling phase is recalculated only for associated nodes that have not yet covered the corresponding response delay, and the original time slot mapping of completed messages is retained. Before the event window ends, the event control record is checked for sequence closure with the event data chain received by the data processing terminal. For nodes with node entry order breakpoints or missing message completion identifiers, a supplementary sampling task with a limited sequence range is issued. The node entry order is determined by the time when the associated node first enters the data acquisition state. The sampling phase version increments each time the window is extended and causes phase adjustment. The message completion identifier is bound to the total number of fragments, the set of received fragments, and the original sequence range. The window boundary version records the basis for each end boundary update.
[0058] In a preferred embodiment, the sequence closure check examines the event data chain one by one according to the node entry order in the event control record. The check includes whether the event source node has a complete out-of-bounds start fragment, whether each associated node covers the target range corresponding to its cumulative response delay, whether all fragmented messages are completed, whether the reconstructed fragment has a source identifier, and whether the supplemented fragment is consistent with the original missing range. If a node entry order breakpoint is found, the coordinator determines the supplemented range based on the end sequence number of the node before the breakpoint and the start sequence number of the node after the breakpoint. If a message completion identifier is found to be missing, the missing fragment is determined based on the fragment sequence number set. If the original cache cannot be recovered, the data processing end writes the unrecoverable range as a closure gap into the event data chain and prohibits the corresponding segment from participating in the dynamic pressure cause-effect graph update.
[0059] In a preferred embodiment, when multiple positive pressure data acquisition nodes experience prediction residual out-of-bounds errors within adjacent time ranges, the coordinating end establishes event source candidates for each node and checks whether the reachable paths, cumulative response time delay ranges, and initial event windows of different event source candidates overlap in the dynamic pressure causal graph. If two event source candidates are connected along the same propagation direction and the subsequent out-of-bounds time falls within the cumulative response time delay range of the earlier event source, the two candidates are merged into the same event control record. The earlier node is retained as the primary event source, and the later node is registered as the secondary event source. The set of associated nodes is the union of the two extended paths, and only the earlier sampling phase is retained for the same nodes. If there is no available directed path between the two event source candidates, or the change direction is inconsistent with the historical propagation direction, independent event windows and independent reporting time slot sets are maintained. The coordinating end interleaves the messages of the two events during link scheduling to avoid mistakenly merging independent positive pressure changes into a single propagation event.
[0060] Preferably, the data processing end adopts a unified index structure for state summaries, reconstructed fragments, original continuous fragments, and supplementary fragments. The unified index structure includes at least node identifier, event number, start sequence number, end sequence number, model version number, graph version number, source type, and verification status. When original data and reconstructed data exist within the same sequence range, the original data is used as the input for event analysis, and the verification record of the reconstructed data is retained. When multiple supplementary data exist within the same sequence range, closed data is selected based on sequence continuity and message completion identifier. The dynamic pressure cause-effect graph update only reads data whose source type is original data or supplementary data and whose verification status is passed. The prediction model update further excludes fragments with event anomaly markers. Event playback restores the original processing conditions according to the model version and graph version bound when the event occurred, so that subsequent version changes do not rewrite the node relationships and data boundaries of existing events.
[0061] Furthermore, the prediction model update is performed on a per-monitoring-point basis. The data processing end selects continuous stable segments from outside the event data chain, and uses the endpoint values, segment slopes, local fluctuations, and time spans of adjacent segments to form training samples, which are then divided into update samples and validation samples according to time order. The updated model calculates the endpoint value deviation, slope deviation, and fluctuation deviation respectively. A new model version is generated only when all three types of deviations do not exceed the corresponding deviations of the old model on the same validation sample. After the new model version is generated, it does not immediately replace the version in the currently executing event, but is issued after the corresponding node exits all event windows, and a new confidence interval is re-established from the stable segments of the new version to ensure that the model switching process and the event acquisition process are isolated from each other.
[0062] In this embodiment, the complete working process is manifested as follows: each positive pressure data acquisition node continuously generates state summaries and prediction residuals; the data processing end maintains a dynamic pressure causal graph using historical measured sequences; the coordination end maintains summary reporting and performs boundary constraint reconstruction when the prediction residuals are within the confidence interval; when the prediction residuals of any node exceed the confidence interval, a unified event window is created and the set of associated nodes is determined along the dynamic pressure causal graph; then, staggered sampling phases are configured according to path hierarchy and cumulative response time delay; ordered reporting time slots are configured according to expected message length and link pending transmission volume; during the event, the window boundary is adjusted according to the residual propagation status and message completion status; sequence closure verification is performed before the event ends; after the event ends, directed edges, confidence intervals, and prediction models are updated with measured data that have passed continuity and version consistency verification. This forms a continuous implementation process from stable state summary acquisition, abnormal correlation additional acquisition, network ordered transmission, data closure verification to model and graph structure updates.
Claims
1. A multi-point positive pressure monitoring system based on Internet of Things data acquisition, characterized in that, This includes multiple positive pressure data acquisition nodes, an IoT data acquisition coordination terminal, and a data processing terminal; The positive pressure data acquisition node generates a status summary based on the current positive pressure value, change slope, fluctuation intensity and short-term trend, and obtains the prediction residual based on the prediction model; The data processing terminal establishes a dynamic pressure causal graph based on the historical sequence of each positive pressure data acquisition node, representing the sequence of changes, response time delay, and residual propagation direction. The IoT data acquisition coordination terminal receives a status summary when the prediction residual is within the confidence interval. The data processing terminal reconstructs the unuploaded segment based on the status summary. When the prediction residual exceeds the confidence interval, the associated node set is determined based on the dynamic pressure cause-effect graph. The terminal then issues an acquisition task containing a unified event window, staggered sampling phase, and ordered reporting time slot to the associated node set.
2. The multi-point positive pressure monitoring system based on Internet of Things data acquisition according to claim 1, characterized in that, The data processing terminal performs sliding alignment of the historical sequence according to the candidate time delay between monitoring point pairs, calculates the directional consistency, time delay stability and repetitive event overlap between the change segment of the preceding sequence and the prediction residual of the following sequence, and forms a directed correlation weight based on the three. The monitoring point pairs whose directed correlation weights meet the edge establishment conditions are written into the dynamic pressure cause-effect graph. For the monitoring point pairs that simultaneously meet the edge establishment conditions in opposite directions, the direction with small residual dispersion is retained, and the corresponding candidate time delay is recorded as the response time delay.
3. The multi-point positive pressure monitoring system based on Internet of Things data acquisition according to claim 1, characterized in that, The positive pressure data acquisition node divides the continuous sampling sequence into adjacent state segments, extracts the endpoint value, segment slope, local fluctuation amount, sequence continuity identifier and model version number for each state segment, and combines them to form the state summary. The predicted residual is formed by weighting the differences between the measured endpoint values of the state segment and the output endpoint values of the prediction model, the differences in the segmented slope, and the differences in local fluctuations. The confidence interval is updated on a rolling basis according to the residual distribution of continuous stable state segments under the same model version.
4. The multi-point positive pressure monitoring system based on Internet of Things data acquisition according to claim 2, characterized in that, The IoT data acquisition coordination terminal uses the positive pressure data acquisition node whose predicted residual exceeds the confidence interval as the event source node, expands layer by layer along the directed edge in the dynamic pressure causal graph that is consistent with the direction of residual propagation, and stops expanding when the cumulative response delay exceeds the event window length or the directed association weight is lower than the expansion condition, so as to obtain the set of associated nodes. Then, based on the graph path hierarchy of each associated node relative to the event source node, staggered sampling phases are allocated, and orderly reporting time slots are configured according to the sampling phase order.
5. The multi-point positive pressure monitoring system based on Internet of Things data acquisition according to claim 2, characterized in that, The data processing terminal maintains an edge state record for each directed edge, consisting of the number of valid events, time delay dispersion, number of direction reversals, and residual propagation consistency. After a new event ends, the measured change segments of each monitoring point within the event are matched with the edge state records. For directed edges that match, the response time delay and the directed association weight are updated. For directed edges that show a reversal of direction or a continuous increase in time delay dispersion, the weight is decayed. When the weight decays to the edge removal condition, the historical edge identifier of the corresponding monitoring point pair is retained and the real-time extension qualification is deleted.
6. The multi-point positive pressure monitoring system based on IoT data acquisition according to claim 3, characterized in that, After receiving the state summary, the data processing terminal calls the corresponding prediction model according to the model version number, and reconstructs the unuploaded state fragments using the endpoint values and segment slopes of adjacent state summaries as boundary constraints. When the sequence continuity identifiers are not continuous, the model version numbers of adjacent state summaries are inconsistent, or the verification residual between the reconstructed fragment and the subsequent measured endpoint value exceeds the verification interval, an original data supplementation instruction pointing to the range of missing sequences is generated, and the reconstructed fragment, the supplemented fragment, and their respective source identifiers are associated and written into the same event data chain.
7. The multi-point positive pressure monitoring system based on IoT data acquisition according to claim 4, characterized in that, The IoT data acquisition coordination terminal constructs a sampling phase sequence based on the graph path level, expected message length, and current link pending transmission volume of each positive pressure data acquisition node in the associated node set, so that the sampling time of adjacent graph path levels keeps covering the overlapping interval of the corresponding response delay. The sampled phase sequence is then mapped to the cyclic reporting time slot. Nodes whose expected message length exceeds the capacity of a single time slot are split into multiple messages with the same event number and consecutive fragment numbers, and the multiple messages are configured in non-adjacent time slots.
8. The multi-point positive pressure monitoring system based on IoT data acquisition according to claim 1, characterized in that, The IoT data acquisition coordination terminal continuously receives the prediction residuals and sequence continuity identifiers of the associated node set within the unified event window, and updates the event window boundary based on the residual fallback status of the event source node, the residual arrival status of the downstream node, and the number of incomplete reporting messages. When the residual of the event source node falls back and the downstream node has not yet reached the corresponding response time delay, the event window is extended. When the residuals of all associated nodes return to the reliable range and the number of unreported messages is 0, the event window ends and the status summary reporting method of the end node is restored.
9. The multi-point positive pressure monitoring system based on IoT data acquisition according to claim 5, characterized in that, The data processing terminal includes the monitoring points whose real-time expansion eligibility has been deleted into the candidate recovery set according to the historical edge identifier, and extracts the change segment that is consistent with the original residual propagation direction in subsequent events. When the monitoring point pair in the candidate recovery set forms a stable response time delay in multiple discontinuous events and the number of direction reversals does not increase, the directed association weight is reconstructed based on the residual propagation consistency of the new events, and the reconstructed directed edge is set as an observation edge that only participates in the verification of the associated node set. After the observed edge passes event verification consecutively, it is then written into the real-time expansion path of the dynamic pressure cause-effect graph.
10. The multi-point positive pressure monitoring system based on IoT data acquisition according to claim 8, characterized in that, The IoT data acquisition coordination terminal maintains an event control record for each event window, including the node entry order, sampling phase version, message completion identifier, and window boundary version. When the event window is extended, the sampling phase is recalculated only for the associated nodes that have not yet covered the corresponding response delay, and the original time slot mapping of the completed message is retained; Before the event window ends, the event control record and the event data chain received by the data processing terminal are checked for sequence closure. For nodes with node entry sequence breakpoints or missing message completion identifiers, a supplementary sampling task with a limited sequence range is issued.