Edge computing-based dangerous chemical delivery state mutation early warning method

Abstract

The present application relates to the technical field of dangerous chemical product transportation alarm signal processing and state early warning, in particular to a dangerous chemical product transportation state mutation early warning method based on edge computing. The method obtains pressure, flow, temperature, pump frequency, valve position and pump valve control instructions at the edge computing node, time-aligns and divides the working condition segments of the process data, calculates the pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position-pressure direction residual and temperature compensation residual according to the working condition, performs multi-scale change point identification on the process data and residual, generates a mutation fingerprint containing the change point time, residual direction, duration and amplitude level, and performs explainability matching with the pump valve control instruction. The method can distinguish between normal pump valve adjustment disturbance and transportation imbalance mutation that cannot be explained by the control instruction, identify risks such as leakage, blockage and air blockage before the fixed limit is exceeded, and form a traceable edge side alarm basis.

Description

Technical Field

[0001] This invention relates to the field of hazardous chemical transport alarm signal processing and status early warning technology, specifically a method for early warning of sudden changes in the status of hazardous chemical transport based on edge computing. Background Technology

[0002] Hazardous chemical transportation typically relies on pumps, valves, pipelines, and control systems to form a continuous transportation process. The operational status is mainly reflected by process data such as pressure, flow rate, temperature, pump frequency, valve position, and pump / valve control commands. Existing transportation alarm systems often set fixed thresholds, rate-of-change thresholds, and over-limit durations at the field acquisition or control terminals. When pressure exceeds the upper limit, flow rate falls below the lower limit, temperature exceeds the set range, or a certain state quantity changes too rapidly within a short period, an audible and visual alarm, control room alert, or a shutdown interlock is triggered. For relatively stable conventional transportation scenarios, this method can provide basic over-limit alerts, but its judgment mainly relies on comparing a single state quantity with preset boundaries, failing to fully utilize the correlations between pressure and flow rate, pump frequency and flow rate, valve position and upstream / downstream pressure, and temperature and pressure fluctuations. Hazardous chemical media possess characteristics such as volatility, corrosiveness, flammability, or toxicity. Sudden changes during transportation often show early signs before the state quantity crosses the danger boundary, and single-point threshold alarms struggle to reliably identify such early state imbalances. Especially when the conveying load changes, the pipeline back pressure fluctuates, the medium temperature drifts slowly, or the pump valve is in a state of frequent adjustment, the boundary of the state quantity itself will change with the change of operating conditions. If the fixed threshold is still triggered by the uniform standard, it is easy to cause the alarm conclusion to deviate from the actual conveying state.

[0003] In relatively recent technologies, to reduce on-site alarm lag, some systems upload sensor data to a central platform or perform simple filtering, trend calculation, and anomaly classification on an edge controller. These solutions typically aggregate pressure, flow, and temperature sequences according to a fixed sampling period, perform mean filtering, sliding window comparison, historical average deviation calculation, or curve abrupt change detection on the data, and then output early warning information based on the anomaly score or classification results. Their implementation still relies primarily on the shape of state variable curves, using control actions such as pump start-up, pump shutdown, valve opening adjustment, and flow setting changes as auxiliary markers or post-event records, without incorporating the sequential, directional, and hysteresis relationships between control commands and state variable responses into alarm criteria. Because normal operating condition switching during hazardous chemical transportation can cause pressure rebound, flow rate rise, temperature compensation fluctuations, and short-term oscillations, relying solely on trend curves or change point detection means that the explainable fluctuations generated by normal control actions are similar in appearance to dangerous abrupt changes such as leaks, blockages, and gas locks. This necessitates a trade-off between false alarm suppression and early warning in the alarm system. If the central platform undertakes the main judgment task, the uploading, cleaning and summarizing of raw data will introduce communication latency; if the edge controller only performs simple threshold comparison, it will be unable to subdivide the causes of anomalies under different operating conditions.

[0004] The main technical problem with existing early warning systems for sudden changes in the status of hazardous chemical transport lies in the lack of a closed-loop discrimination mechanism at the edge side that addresses the inherent correlations within the transport process. This makes it difficult to distinguish between normal operating disturbances caused by pump and valve control actions and genuine hazardous changes caused by imbalances in the transport status. Specifically, when pump frequency adjustments, valve position switching, or start / stop commands cause short-term jumps in pressure and flow, the system is prone to triggering false alarms due to state variables exceeding limits or having excessive rates of change. Furthermore, when leaks, blockages, or gas locks are in their early stages, individual state variables may not have reached the fixed alarm threshold, and the system may experience a delayed response. The fundamental technical reason is that existing solutions fail to integrate operating condition segment division, physical property constraint residuals, multi-scale change point results, and the interpretability matching of pump and valve control commands into a unified criterion at the edge side. This makes it impossible to determine whether a sudden change in status simultaneously violates pressure-flow correlation, pump frequency-flow hysteresis, valve position pressure direction, and temperature compensation relationships, resulting in alarm information lacking technical support for the source of the change. Under continuous transportation conditions of hazardous chemicals, this discriminative gap will cause the alarm system to only see the surface changes of the state quantity and cannot confirm whether the changes exceed the interpretable range of the current operating conditions. Therefore, it is difficult to form a verifiable and traceable early warning conclusion of a sudden change at the edge side, and it is also difficult to ensure that the alarm triggering and the actual transportation imbalance process are consistent in time and direction. Summary of the Invention

[0005] The purpose of this invention is to provide an early warning method for sudden changes in the status of hazardous chemical transportation based on edge computing, which can solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing includes: acquiring hazardous chemical transportation process data at an edge computing node, wherein the process data includes pressure, flow rate, temperature, pump frequency, valve position and pump control commands; The process data is time-aligned and divided into operating condition segments; Calculate the pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual, and temperature compensation residual based on the aforementioned operating condition segment. Multi-scale change point identification is performed on the process data and the residuals to generate a mutation fingerprint including the change point time, residual direction, duration and magnitude. The mutation fingerprint is matched with the pump and valve control commands for interpretability, and a warning information on the mutation of hazardous chemical transportation status is generated based on the matching result.

[0007] Preferably, the process data is time-aligned and divided into operating condition segments, including: using the triggering time of the pump valve control command as the segment anchor point, and correcting the sampling period of the pressure, flow and temperature sampling sequences; Combine pump frequency change rate, valve position change direction, flow setting change and medium identification into operating condition labels; When adjacent operating condition labels switch, a transition buffer segment is retained, and the transition buffer segment is stored separately from the steady-state conveying segment, start-stop transition segment, valve switching segment and flow regulation segment. The corresponding time boundary and conveying stage mark are recorded for each operating condition segment.

[0008] Preferably, the calculation of pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual and temperature compensation residual according to the operating condition segment includes: selecting synchronous sampling points for pressure, flow, temperature, pump frequency and valve position in each operating condition segment, and establishing a local reference mapping between each state quantity; Pressure-flow correlation residuals are generated based on the synchronous relationship between pressure and flow rate changes; Pump frequency flow hysteresis residual is generated based on the chronological relationship between pump frequency adjustment time and flow response time; Generate valve position pressure direction residual based on the valve position movement direction and the upstream and downstream pressure change direction; After compensating the pressure sequence based on the temperature change, a temperature compensation residual is generated, and the four types of residuals are archived according to the operating condition labels.

[0009] Preferably, multi-scale change point identification is performed on the process data and the residuals, including: calculating candidate points for state variable abrupt changes and candidate points for residual abrupt changes in a short time window, a medium time window, and a long time window, respectively. Adjacent candidate points within the same working condition segment are merged into variable point clusters based on their temporal overlap relationship. For each variable cluster, record the time of the change, the direction of the change, the residual category, the duration, the amplitude level, and the corresponding operating condition label. When a state variable mutation candidate point and at least one type of residual mutation candidate point appear in the same variable point cluster, the mutation fingerprint is generated, and isolated candidate points that are not included in the variable point cluster are recorded as sampling points to be confirmed.

[0010] Preferably, the step of retaining a transition buffer segment when adjacent operating condition labels switch includes: determining the buffer start and end boundaries by the reverse point of the pressure slope, the start point of the flow response, and the stable point of temperature sampling before and after the pump valve control command is triggered; When multiple pump and valve control commands appear in the buffer section, control commands with a time interval less than a multiple of the preset sampling interval are merged into a composite operating condition label, and the order of each control command in the composite operating condition label is recorded. An independent reference mapping is set for the state variable sequence under the composite operating condition label, and the residual is not calculated together with the steady-state delivery segment. The corresponding sampling point sequence number is configured for the independent reference mapping.

[0011] Preferably, the calculation of pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual and temperature compensation residual according to the operating condition segment further includes: setting residual baselines according to the operating condition label and the composite operating condition label respectively; Before updating the residual baseline, discard sample segments containing unexplained variable points; Convert the residual values ​​into direction codes, deviation levels, and persistence counts; When at least two types of residuals exhibit directional coupling within the same time frame, a residual combination identifier is generated and written into the mutation fingerprint, while the corresponding sampling point number is saved in the archived record.

[0012] Preferably, the process data and the residual are subjected to multi-scale change point identification, including: using short time window candidate points as cluster centers, using medium time window candidate points to confirm the continuous boundary, and using long time window candidate points to confirm the steady-state offset interval. When the mutation directions corresponding to short time windows, medium time windows and long time windows are inconsistent, retain the combination with the smallest time distance between residual mutation candidate points and state variable mutation candidate points. Short time window candidate points that are not confirmed by medium and long time windows are marked as transient candidate points, and the transient candidate points are associated with the residual combination identifier and stored.

[0013] Preferably, the interpretability matching of the mutation fingerprint with the pump valve control command includes: searching for corresponding commands for pump frequency adjustment, valve position action, start / stop switching and flow setting change before and after the change point of the mutation fingerprint; Match the instruction direction and instruction sequence with the residual direction, mutation direction, and duration; The mutation fingerprint of a successfully matched combination whose residual combination identifier belongs to the permitted combination of the corresponding working condition label is marked as a working condition disturbance. Mutation fingerprints with no corresponding instruction or whose instruction direction conflicts with the residual direction are marked as abnormal mutation candidates, and the instruction number on which the matching is based is recorded.

[0014] Preferably, the step of generating early warning information for sudden changes in the status of hazardous chemical transport based on the matching results includes: writing the mutation fingerprints marked as candidates for abnormal mutations into the edge alarm queue according to the residual combination identifier, amplitude level and duration, and retaining the matching results corresponding to the mutation fingerprints; When the same residual combination identifier appears consecutively within the same working condition segment, the edge alarm queue status will be switched from observation status to warning status. When a new mutation fingerprint in the early warning state has the same residual direction as the previous mutation fingerprint and the magnitude level does not drop, the output includes early warning information including the transport stage marker, residual category, and change point time.

[0015] Preferably, after the edge alarm queue state is switched from observation state to warning state, the method further includes: performing backtracking verification on the working condition segment corresponding to the warning state, using the pump valve control command, the working condition label, the composite working condition label, the residual baseline and the permission combination as a verification index; When the backtracking verification finds that the mutation fingerprint falls within the independent reference mapping range of the composite working condition label, the edge alarm queue status is reset to the observation status and the verification record is retained. When the backtracking verification fails to hit the independent benchmark mapping range and the residual combination identifier continues to update, an acknowledgment alarm message is generated and the corresponding mutation fingerprint sequence is locked.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In the edge computing node, pressure, flow rate, temperature, pump frequency, valve position, and pump / valve control commands are time-aligned. Pressure-flow correlation residuals, pump frequency-flow hysteresis residuals, valve position-pressure direction residuals, and temperature compensation residuals are calculated for different operating conditions such as steady-state delivery, start-stop transition, valve switching, and flow regulation. This transforms alarm judgment from a single state quantity exceeding limit to a joint verification of state quantity changes and the delivery correlation. Multi-scale change point identification is used to determine the change point time, residual direction, duration, and amplitude level. Property constraint residuals are used to determine whether state changes conform to the process correlation under the current operating condition. Pump / valve control command interpretability matching is used to eliminate pressure rebound, flow rate climb, or short-term oscillations caused by normal control actions. These three types of processing form a closed criterion on the same time axis at the edge side. Only when state quantity abrupt change, residual abnormality, and control command mismatch are simultaneously satisfied will a hazardous chemical delivery state abrupt change warning information be generated. This addresses the difficulty in distinguishing between normal operating disturbances and genuine dangerous sudden changes in the background technology, reducing false alarms caused by pump and valve actions, changes in flow settings, or temperature compensation fluctuations. Furthermore, it generates early warnings based on residual combinations and the persistence of change points before imbalances such as leaks, blockages, and air resistance reach fixed alarm boundaries. This process is completed at the field edge nodes, eliminating the need to wait for full data transmission and centralized analysis at the central end. Alarm outputs can carry transport stage markers, residual categories, change point times, and matching results, providing clear data sources and judgment criteria for early warning information. For transport processes involving short-term impacts, continuous deviations, and steady-state migrations, short, medium, and long time windows participate in the same change fingerprint generation, ensuring that instantaneous fluctuations do not independently determine alarms, and continuous residuals do not trigger alarms independently of change point locations. The alarm conclusion remains consistent with the time, direction, and duration of the change.

[0017] 2. Further, through progressive processing using operating condition tags, composite operating condition tags, transition buffer sections, residual baselines, residual combination identifiers, and edge alarm queues, the early warning system for sudden changes in the state of hazardous chemical transportation gains more stable operating condition adaptability. When dividing operating condition segments, the trigger time of pump and valve control commands is used as the segment anchor point, and operating condition tags are formed by combining pump frequency change rate, valve position change direction, flow setting change, and medium identifier. This avoids mixing start-stop transitions, valve switching, and steady-state transportation into the same residual baseline. For continuously occurring control commands, composite operating condition tags and independent baseline mappings are set, reducing misjudgments during complex operation phases. Short, medium, and long time windows record transient candidate points, persistent boundaries, and steady-state offset intervals, respectively, enabling the mutation fingerprint to cover instantaneous impacts, persistent deviations, and stable state transitions. After residual direction encoding, deviation level, persistence count, and residual combination identifiers are jointly written into the edge alarm queue, the observed state, early warning state, and confirmed alarm state can be progressively updated according to the same mutation sequence. The backtracking verification uses pump and valve control commands, operating condition labels, composite operating condition labels, residual baselines, and permissible combinations as verification indexes. This allows events already in the warning state to still be reviewed based on independent benchmark mappings, preventing single abnormal samplings under complex operating conditions from directly generating confirmed alarms. When a mutation fingerprint falls within the independent benchmark mapping range corresponding to a composite operating condition label, the edge alarm queue can be reset to the observation state and the verification record retained. When a mutation fingerprint does not hit the independent benchmark mapping range and the residual combination identifier is continuously updated, the system can lock the corresponding mutation fingerprint sequence and generate confirmed alarm information, ensuring the alarm state remains consistent with the edge-side data processing chain. Through this progressive processing, each state switch in the alarm queue corresponds to the recorded operating condition boundary, residual category, and interpretability matching result. Subsequent reviews can trace back the corresponding data segment along the sampling point sequence number, preventing the same delivery event from being repeatedly interpreted or marked contradictorily in different processing stages. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a method for early warning of sudden changes in the status of hazardous chemical transportation based on edge computing. Figure 2 Flowchart for dividing working condition segments and generating residuals for physical property constraints; Figure 3 The flowchart shows the process for multi-scale change point identification, alarm queue update, and backtracking verification. Detailed Implementation

[0019] In one embodiment, refer to Appendix Figure 1An edge computing-based early warning method for sudden state changes in hazardous chemical transportation is deployed in edge computing nodes on the transportation control side. These nodes receive pressure, flow rate, temperature, pump frequency, valve position, and pump / valve control commands during the hazardous chemical transportation process. They then map process data from different sources onto the same time axis. Subsequently, based on the pump / valve control commands and the relationship between state variables, the system divides the process into segments such as steady-state transportation, start-stop transition, valve switching, flow regulation, and abnormal recovery. Within each segment, it calculates the pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position-pressure direction residual, and temperature compensation residual. Finally, it performs multi-scale change point identification on the process data and residuals under short, medium, and long time windows, obtaining information including the change point time, residual direction, duration, amplitude level, and operating condition standard. The edge computing node performs interpretability matching between the mutation fingerprint and the pump valve control command. When the mutation fingerprint cannot be interpreted by the corresponding control command, the corresponding operating condition label, and the corresponding permission combination, a warning message for the mutation of the hazardous chemical transportation status is generated. The working logic of this embodiment is to process the curve mutation, physical property constraint residual, and control command interpretation relationship in the same edge-side discrimination chain, so that the alarm trigger no longer depends only on the single point of pressure, flow or temperature exceeding the limit, but is jointly determined by whether the state variable mutation violates the correlation relationship under the current transportation operating condition. This embodiment is suitable for continuous transportation scenarios of hazardous chemicals that need to retain the real-time nature of on-site alarms and need to distinguish between normal operating condition disturbances and dangerous mutations. It can form a warning record that includes sampling points, operating condition boundaries, residual categories and matching conclusions.

[0020] In one embodiment, the pressure sequence, flow sequence, temperature sequence, pump frequency sequence, valve position sequence, and pump / valve control command sequence received by the edge computing node may come from different sampling links. During processing, a unified sampling scale is used as the alignment benchmark. This unified sampling scale is generated by the event clock within the edge computing node. The trigger time of the pump / valve control command is recorded as a segmented anchor point. Continuous state variables such as pressure, flow, and temperature are aligned using interpolation between adjacent sampling points. State variables with holding characteristics, such as pump frequency and valve position, are aligned using the most recent valid sampling point. Each aligned sampling record includes the unified sampling time, the value of each state variable, the corresponding control command number, the relative trigger time of the command, and the medium identifier. If a state variable is missing at the unified sampling scale, the edge computing node does not directly use the supplementary value for alarm purposes. Instead, it adds a missing marker to the corresponding sampling record and reduces the participation level of that sampling record during residual calculation. The alignment logic can be expressed by the following formula: ; in, The index of the original sampling point is used to mark the first... The chronological order of the original sampling times of continuous state variables; Indicates the first Continuous state variables at uniform sampling time Alignment value, and They represent the first Class state variables in Adjacent to each other The, the The original sampled values ​​corresponding to each original sampling time. , The fraction represents the rate of change per unit time. If the pressure is 100 at sampling time 2 and 108 at sampling time 4, the pressure alignment value at sampling time 3 is 104. The pump frequency and valve position adopt the hold value confirmed in the most recent control cycle at the same time. This embodiment uses a unified time axis to ensure that subsequent residuals, change points and command interpretations all have the same time base.

[0021] In one embodiment, refer to Appendix Figure 2 The segmentation of operating conditions uses the triggering time of pump and valve control commands as the segmentation anchor point. Operating condition labels are generated by combining pump frequency change rate, valve position change direction, flow setting change, and medium identification. Pump frequency change rate characterizes the direction of change in the conveying power input; valve position change direction characterizes the direction of change in the pipeline throttling relationship; flow setting change characterizes the change in the control target; and medium identification characterizes the differences in temperature compensation and pressure response between different hazardous chemical media. Edge computing nodes retain a transition buffer segment when adjacent operating condition labels switch. The transition buffer segment is defined by the start and end points of the state variable responses before and after the control command triggering and is not used in conjunction with the steady-state conveying segment for residual calculation. The steady-state conveying segment and the start / stop transition segment are also included. Valve switching segments, flow regulation segments, and anomaly recovery segments are stored separately. Each segment records the start time, end time, operating condition label, medium identifier, control command number, transition buffer marker, and sampling point sequence number. If multiple pump and valve control commands appear within a buffer segment, the edge computing node synthesizes these control commands into a composite operating condition label in chronological order, retaining the original number of each control command. The composite operating condition label does not replace the original control command but serves as an operating condition index for subsequent interpretability matching and backtracking verification. This embodiment uses operating condition segments and operating condition labels to prevent the residual benchmark from being mixed across different delivery stages, avoiding placing start-stop transitions and steady-state delivery in the same alarm criterion. The operating condition segments and edge recording fields are shown in Table 1 below.

[0022] Table 1. Fields for Operating Condition Segments and Edge Records

[0023] In one embodiment, the pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual, and temperature compensation residual are all calculated within the operating condition segment. Edge computing nodes first select time-aligned synchronous sampling points for pressure, flow, temperature, pump frequency, and valve position within the same operating condition segment, and then establish a local reference mapping. This local reference mapping is formed from sampling points within the current operating condition segment that are not marked as unexplained variable points. The pressure-flow correlation residual represents the deviation between pressure and flow changes; the pump frequency-flow hysteresis residual represents the time deviation of the flow response after pump frequency adjustment; the valve position pressure direction residual represents the directional deviation between the valve position action direction and the upstream and downstream pressure change directions; and the temperature compensation residual represents the remaining deviation after temperature change compensation of the pressure sequence. The residual calculations use the same notation system.

[0024] in, This indicates the pressure-flow correlation residual. This indicates the pump frequency flow hysteresis residual. This indicates the residual pressure direction at the valve position. This indicates the temperature compensation residual. , , , , These represent the changes in pressure, flow rate, pump frequency, valve position, and temperature between adjacent uniform sampling times, respectively. and These represent the changes in upstream pressure and the changes in downstream pressure, respectively. , , Operating condition label The corresponding local reference coefficients, Operating condition label The corresponding flow response hysteresis sampling offset, The sign represents the operation: positive change is represented by 1, negative change by -1, and no change by 0. For example, if the pressure change is 12 and the flow rate change is 3 within a certain steady-state segment... If it is 4, then If the change in flow rate is 1 for the same change in pressure, then... The value is 8. In this embodiment, four types of residuals are used to transform a single state quantity curve into a calculable quantity corresponding to the relationship with the hazardous chemical transportation process.

[0025] Furthermore, edge computing nodes maintain residual baselines for different operating condition labels. These residual baselines are not fixed values ​​but are formed by updating the most recently confirmed interpretable sampling segments under the same operating condition label. Before updating, sampling segments containing unexplained variable points and those with too many missing markers are removed. The remaining sampling segments enter the residual baseline candidate region according to their sampling point sequence numbers. The edge computing nodes extract the median trend, directional distribution, and persistence count distribution from the residual sequences in the candidate region to obtain the residual baseline record under the current operating condition label. This residual baseline record includes at least the operating condition label, composite operating condition label, media identifier, residual category, directional encoding, deviation level boundary, allowed combinations, and independent baseline mapping number. The residual value is not written to the mutation fingerprint before... The residuals are converted into directional codes, deviation levels, and persistence counts. The directional codes represent positive, negative, or no directional deviations from the baseline. The deviation levels represent the position of the residual amplitude relative to the baseline distribution. The persistence counts represent the continuation of the same type of deviation at consecutive sampling points. When at least two types of residuals are directionally coupled within the same time range, the edge computing node generates a residual combination identifier. For example, a combination identifier is formed when the pressure-flow correlation residual is positively deviated and the pump frequency-flow hysteresis residual is negatively deviated. Another combination identifier is formed when the valve position pressure directional residual and the temperature compensation residual are both abnormal. In this embodiment, the residual interpretation under different operating conditions is fixed in a traceable data record through the residual baseline and residual combination identifier.

[0026] In one embodiment, refer to Appendix Figure 3 Multi-scale change point identification is performed simultaneously within short, medium, and long time windows. The short time window is used to capture instantaneous jumps in pressure, flow, temperature, and residuals. The medium time window is used to confirm whether the jump continues as a continuous deviation. The long time window is used to confirm whether a steady-state shift has formed before and after the abrupt change. Edge computing nodes calculate the difference in mean values ​​before and after the window for each type of state variable and residual sequence, and use the absolute deviation within the window as a normalization benchmark to obtain the candidate change point intensity. The candidate change point intensity can be expressed by the following formula: ; in, Indicates time window At the unified sampling time The strength of candidate change points at the location, This represents the pressure, flow rate, temperature, or any residual sequence involved in the change point identification. This represents the mean value during the latter part of the time window. This represents the mean value at the beginning of the time window, and the median value represents the absolute deviation within the time window. This represents a small positive quantity used to prevent the denominator from being zero. For example, if the mean of a residual in the latter part of the window is 18, the mean in the first part of the window is 10, and the median absolute deviation is 2. If a negligible small value is taken, the candidate change point strength is close to 4. The edge computing node compares the candidate change point strength with the deviation level boundary in the residual baseline under the same working condition to form candidate points. In this embodiment, the state variables and residual sequences are processed by the same formula, so that the change point identification results can be directly associated with the residual combination identifier.

[0027] In this embodiment, short-time window candidate points are used as the centers of variable point clusters, medium-time window candidate points are used to confirm the continuity boundary, and long-time window candidate points are used to confirm the steady-state offset interval. Edge computing nodes merge adjacent candidate points within the same working condition segment into variable point clusters based on their temporal overlap. During merging, the occurrence time of the short-time window candidate point is used as the initial center, the start and end range of the medium-time window candidate point is used to correct the variable point cluster boundary, and the steady-state offset direction of the long-time window candidate point is used to record the state transition direction after the abrupt change. When the abrupt change directions corresponding to the short, medium, and long time windows are consistent, the variable point cluster directly retains that abrupt change direction. When the abrupt change directions corresponding to different time windows are inconsistent, the edge computing nodes retain the residual abrupt change candidate points and the state transition direction. The combination of candidate points with the smallest time distance for state variable mutations is selected, and other candidate points are recorded as sampling points to be confirmed. The variable cluster recording fields include the mutation time, mutation direction, residual category, duration, amplitude level, operating condition label, composite operating condition label, sampling point sequence number, and confirmation mark. When a candidate point for a state variable mutation and at least one candidate point for a residual mutation fall into the same variable cluster, the edge computing node generates a mutation fingerprint. If there are only candidate points for state variables and no candidate points for residual variables, a mutation fingerprint is not generated and isolated candidate points are recorded. If there are only candidate points for residual variables and no candidate points for state variables appear synchronously, the corresponding sampling segment is marked as a suspected data quality segment. This embodiment avoids direct entry into the alarm chain for a single sampling jump through variable clusters. The mutation fingerprint and interpretability matching fields are shown in Table 2 below.

[0028] Table 2 Mutation fingerprints and fields matching interpretability

[0029] In a preferred embodiment, the start and end boundaries of the transition buffer section are not directly determined by a fixed time length, but by the response characteristics of the state variables. Edge computing nodes search for the pressure slope reversal point, the flow response start point, and the temperature sampling stability point before and after the pump / valve control command is triggered. The pressure slope reversal point is determined by the continuous change in the sign of the pressure change; the flow response start point is determined by the flow change moving from the stable region to the continuous deviation region; and the temperature sampling stability point is determined by the temperature change returning to near the baseline of the same operating condition. The buffer start point is the closest response preparation point before the control command is triggered, and the buffer end point is the latest response stability point after the control command is triggered. If the buffer section... When multiple pump and valve control commands appear, the edge computing node determines whether the sampling interval between adjacent control commands is less than a preset sampling interval multiple. When the merging condition is met, a composite operating condition label is formed. The composite operating condition label retains the order of pump frequency adjustment, valve position action, start-stop switching, and flow setting change, and generates an independent reference mapping number. The sampling points of the independent reference mapping come from the historical segments that are confirmed to be interpretable under the composite operating condition label. They are not used to calculate the residuals together with the steady-state delivery segments. In this embodiment, the buffer boundary is determined by the response characteristics, so that the dense control action stage can be processed independently, avoiding the misclassification of multiple control commands into a single steady-state residual baseline.

[0030] Furthermore, when establishing the independent benchmark mapping, the sampling point sequence number is used to maintain the original timing. The edge computing node divides the pressure, flow, temperature, pump frequency, and valve position sequences under the composite operating condition label into several response segments according to the order of control commands. Each response segment retains the corresponding command number, response start point, response stability point, residual category, and interpretable marker. The residual baseline of the composite operating condition label is composed of these response segments. However, each response segment still retains its own local benchmark coefficient. If a mutation fingerprint occurs later, the edge computing node first indexes the independent benchmark mapping according to the composite operating condition label, then finds the corresponding response segment according to the sampling point sequence number, and compares the residual direction and duration of the mutation fingerprint with the response segment record. If the mutation fingerprint falls within the range of the independent benchmark mapping, the mutation fingerprint is written into the observation state and a verification record is retained. If the mutation fingerprint does not fall within the range of the independent benchmark mapping, it enters the abnormal mutation candidate processing. In this embodiment, the independent benchmark mapping is not a separate addition of a new alarm route, but an extension of the operating condition segment division and residual calculation, so that the residual judgment under the composite control action can still return to the same edge-side closed-loop criterion.

[0031] In one embodiment, interpretable matching uses abrupt fingerprints as input and pump / valve control commands, operating condition labels, composite operating condition labels, residual baselines, and permitted combinations as matching indices. Edge computing nodes retrieve control commands such as pump frequency adjustments, valve position actions, start / stop switching, and flow setting changes before and after the change point of the abrupt fingerprint. The retrieval range is determined by the response hysteresis sampling offset corresponding to the operating condition label. During matching, the command direction, command sequence, residual direction, abrupt direction, and duration are compared, and it is determined whether the residual combination identifier belongs to the permitted combination of the corresponding operating condition label. The discriminant of interpretable matching can be expressed by the following formula: ; in, Indicating mutation fingerprints With control commands The matching discriminant between them This indicates an indicator operation; the value is 1 if the condition within the parentheses is true and 0 if it is false. Indicates the direction of mutation in the mutation fingerprint. Indicates the expected direction corresponding to the control command. Indicates the moment of change. Indicates the timing of the control command trigger. Operating condition label The corresponding response time window represents the residual combination identifier in the mutation fingerprint. Indicates control commands In working condition label The set of allowed residual combinations Represents the residual class sequence in the mutation fingerprint. The residual response sequence corresponding to the control command is represented. If all four indicator operations are 1, the mutation fingerprint is marked as a condition disturbance. If there is no corresponding control command, the command direction conflicts with the residual direction, or the residual combination identifier does not belong to the permitted combination, the mutation fingerprint is marked as an abnormal mutation candidate. In this embodiment, the curve changes that can be explained by the control action are separated from the alarm chain by matching the discriminant quantity.

[0032] In this embodiment, the permitted combinations are jointly defined by the operating condition label and the composite operating condition label. The permitted combination range for the steady-state delivery segment is relatively narrow. The start-stop transition segment and valve switching segment allow for pressure rebound and flow rate rise in the corresponding direction. The flow regulation segment allows for short-term deviations in pump frequency flow hysteresis residuals. The permitted combinations for the composite operating condition segment are determined according to the instruction sequence and independent reference mapping records. The edge computing node does not directly hardcode the permitted combinations as fixed entries, but maintains the permitted combination record based on the residual combination identifier in the confirmed interpretable segment. The permitted combination record includes the operating condition label, control instruction type, instruction direction, residual combination identifier, and holding... The duration range and sampling point sequence number are specified. Before a new segment is added to the permitted combination record, it must meet the following conditions: the corresponding control instruction exists, the change point time falls within the response time window, the residual direction is consistent with the instruction direction, and no abnormal mutation candidates are found during backtracking verification. If a mutation fingerprint hits both the steady-state residual baseline and the composite operating condition independent benchmark mapping, the edge computing node will preferentially use the composite operating condition independent benchmark mapping for interpretation and retain the record of the steady-state residual baseline hit for subsequent verification. In this embodiment, through the collaborative recording of permitted combination and independent benchmark mapping, the interpretable fluctuations generated by normal control actions will not be confused with the state imbalance without instruction support as the same type of event.

[0033] Furthermore, the edge alarm queue is used to receive abnormal mutation candidates, observation status, early warning status, and confirmed alarm status. Queue entries include mutation fingerprint number, operating condition label, composite operating condition label, residual combination identifier, residual direction, amplitude level, duration, matching result, instruction number, and backtracking verification record. After an abnormal mutation candidate is written into the queue, the edge computing node compares the residual combination identifier and residual direction of the current queue entry with the previous queue entry. If the same residual combination identifier appears consecutively within the same operating condition segment and the direction remains consistent, the queue switches from observation status to early warning status. If the amplitude level of a newly added mutation fingerprint in the early warning status does not decrease and the residual direction is consistent with the previous mutation fingerprint, then early warning information including the transport stage marker, residual category, change point time, and matching basis is output. The continuous anomaly count of the edge alarm queue can be expressed by the following formula: ; in, Indicates the first The consecutive anomaly count after each queue entry is updated This represents the consecutive exception count of the previous queue entry, with 0 indicating the lower bound of the count. and These represent the residual combination identifiers of the current and previous mutation fingerprints, respectively. and These represent the residual directions of the current and previous mutation fingerprints, respectively. Indicates the current mutation fingerprint, Operating condition label The corresponding independent benchmark mapping range, if the previous consecutive anomaly count is 1, the current residual combination identifier and residual direction are consistent with the previous entry and do not hit the independent benchmark mapping range, then the current consecutive anomaly count is 2. If the current mutation fingerprint hits the independent benchmark mapping range, the count is deducted and kept not less than 0. In this embodiment, the early warning output depends on the continuous mutation sequence rather than a single sampling anomaly by using queue counting.

[0034] In a preferred embodiment, after the warning state is generated, it is not immediately fixed as a confirmed alarm. Instead, a backtracking verification is performed. The edge computing node uses the pump and valve control command, operating condition label, composite operating condition label, residual baseline, and permitted combination as verification indexes. It backtracks the relevant operating condition segments from the sampling point sequence number corresponding to the current mutation fingerprint. The backtracking content includes the control command trigger time, response buffer start point, response stability point, residual baseline version, residual combination identifier source, and permitted combination hit record. If the backtracking verification finds that the mutation fingerprint falls within the independent baseline mapping range of the composite operating condition label, the edge alarm queue state is reset to the observation state. The system writes the hit independent reference mapping number, response segment number, and reset time into the queue entry. If the backtracking verification fails to hit the independent reference mapping range, and the residual combination identifier of the subsequently added mutation fingerprint is continuously updated, the edge computing node generates a confirmation alarm message and locks the corresponding mutation fingerprint sequence. The locked content includes the starting change point, the ending change point, the participating residual category, the control command matching result, and the backtracking verification record. In this embodiment, backtracking verification enables events that have entered the warning state under complex working conditions to still be reviewed, avoiding the composite control action from being mistakenly fixed as a dangerous mutation, while retaining the unexplainable continuous state imbalance sequence.

[0035] In this embodiment, after the mutation fingerprint sequence is locked, the edge computing node does not modify the original output sampling record. Instead, it appends the locking result as an alarm record to the sampling point sequence number. The alarm record includes the transport stage marker, residual category, residual direction, change point time, duration, amplitude level, abnormal mutation candidate marker, warning status change record, confirmed alarm generation record, and backtracking verification result. Subsequently, when a similar mutation fingerprint appears again under the same operating condition label, the edge computing node retrieves the residual combination identifier and control command interpretation relationship based on the existing locked sequence, but does not directly replace the current alarm result with the historical alarm result. Instead of calculating, the current data is still reprocessed in the order of time alignment, working condition segment division, residual calculation, multi-scale change point identification, and interpretability matching. If the current mutation fingerprint has the same residual combination identifier and the same direction encoding as the historical locking sequence, but the current hits the composite working condition independent reference mapping, the current event is processed as an observation state. If the current event still cannot be interpreted by the control command and forms a continuous abnormal count in the queue, the confirmation alarm information is regenerated. This embodiment uses the method of locking sequence and recalculation to make the edge-side alarm record have a basis for review, while avoiding the direct overwriting of the current working condition by historical conclusions.

[0036] In a preferred embodiment, the residual baseline and permitted combination update adopt a delayed write method. After a transmission segment ends, the edge computing node reads the condition label, composite condition label, mutation fingerprint, matching result and queue status within the segment. Only segments marked as condition disturbances and not entering the abnormal mutation candidate can enter the update candidate area. Segments entering the candidate area must also meet the following requirements: continuous sampling point sequence, complete control command number, complete transition buffer boundary, and residual combination identifier matching the permitted combination. The edge computing node merges the local baseline coefficient of the candidate segment with the existing residual baseline. During merging, the original version is retained and a new version number is generated. The residual baseline currently in use during alarm judgment is not replaced in the same sampling period. It is switched to the new version after the current condition segment ends. If the new version causes too many abnormal mutation candidates that cannot be explained in subsequent backtracking verification, the edge computing node can roll back to the previous version and retain the rollback record. This embodiment uses delayed write and version retention to make the residual baseline updated with the condition record on the edge side, while avoiding the mutation fingerprint being judged being affected by the baseline change in the same period.

[0037] Specifically, edge computing nodes differentiate between isolated candidate points and suspected data quality fragments. Isolated candidate points refer to sampling points that have state variable mutation candidate points but have not formed residual mutation candidate points. Suspected data quality fragments refer to sampling fragments that have residual mutation candidate points but have not appeared synchronously with state variable mutation candidate points. Isolated candidate points are recorded in the pending confirmation sampling area and participate in variable cluster merging again in the subsequent medium or long time window. If no continuous boundary is formed subsequently, they will not enter the edge alarm queue. Suspected data quality fragments retain missing markers, sampling point sequence numbers, and corresponding residual categories, and are excluded when the residual baseline is updated. If a suspected data quality fragment subsequently overlaps with a state variable mutation candidate point in the same operating condition fragment in time, a new variable cluster is generated and enters mutation fingerprint processing. If no time overlap occurs, it is only retained as a data processing record. This embodiment separates the processing of short-term state variable jumps, isolated residual deviations, and real mutation sequences through the pending confirmation sampling area, avoiding a single abnormal sampling from directly changing the alarm status.

[0038] In one embodiment, the medium identifier is used to participate in the generation of operating condition labels and the calculation of temperature compensation residuals. The edge computing node does not introduce new hardware structures based on the medium name, nor does it change the existing acquisition objects. Instead, it uses the medium identifier as the index field of the residual baseline. When the medium identifiers are the same, they share the corresponding temperature compensation baseline. When the medium identifiers are different, they maintain temperature compensation coefficients and permitted combination records respectively. If the transport medium is switched, the edge computing node regards the switching command as the operating condition segment anchor point and forms a medium switching buffer segment. The temperature compensation residual in the medium switching buffer segment is not statistically analyzed together with the steady-state transport segment before and after the switch. The pressure-flow correlation residual and the pump frequency-flow hysteresis residual are still calculated according to the local reference mapping in the current segment. When the medium switching buffer segment is completed and enters stable transport, the residual baseline corresponding to the new medium identifier is allowed to participate in the early warning judgment. This embodiment restricts the mixing of residual baselines under different hazardous chemical media through the medium identifier, so that the calculation boundaries of temperature compensation and pressure response are consistent with the actual transport stage.

[0039] Furthermore, the content of the early warning information is directly composed of fields in the edge alarm queue. The early warning information includes at least the transport stage marker, residual category, residual combination identifier, change point time, duration, amplitude level, operating condition label, composite operating condition label, matching result, and instruction number. The confirmed alarm information adds the mutation fingerprint sequence locking result and backtracking verification record to the early warning information. The edge computing node uses the same data structure to store the early warning information and the confirmed alarm information, with the only difference being the queue status and locking mark. If the early warning status is reset to the observation status, the original early warning information is not deleted, but is marked as reset and written to the independent benchmark mapping hit record. If the confirmed alarm information is generated, the mutation fingerprint sequence participating in the confirmation is set as a read-only record. Subsequent residual baseline updates do not change the calculation basis of the locked record. This embodiment enables early warning, confirmation, reset, and verification to be backtracked along the same sampling point sequence number through a unified data structure, and the alarm information is consistent with the edge side calculation process.

[0040] In this embodiment, after the edge computing node completes a full processing flow, it forms a closed data chain consisting of time-aligned records, operating condition segment records, residual records, variable point cluster records, mutation fingerprint records, interpretability matching records, and edge alarm queue records. Time-aligned records ensure that data from different processes can be compared at the same sampling time. Operating condition segment records ensure that the residual baseline is not used across the transportation stage. Residual records convert the correlation between pressure, flow rate, temperature, pump frequency, and valve position into computable quantities. Variable point cluster records combine short-term jumps, continuous deviations, and steady-state transitions into a unified mutation fingerprint. Interpretable matching records compare the mutation fingerprint with pump and valve control commands in terms of direction, timing, and permissible combinations. Edge alarm queue records progressively process abnormal mutation candidates into observation states, warning states, or confirmed alarm states. The advantage of this embodiment is that the early warning of mutations in the hazardous chemical transportation state no longer relies on a single threshold or a single curve shape, but forms a complete discrimination chain from state quantity sampling to alarm information generation at the edge. Interpretable fluctuations generated by normal control actions can be separated, and mutation sequences that cannot be explained by control commands and continuously violate the transportation correlation can be recorded and alarmed.

Claims

1. A method for early warning of sudden changes in the status of hazardous chemical transportation based on edge computing, characterized in that, include: Data on the transportation process of hazardous chemicals is acquired at the edge computing node. The process data includes pressure, flow rate, temperature, pump frequency, valve position, and pump control commands. The process data is time-aligned and divided into operating condition segments; Calculate the pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual, and temperature compensation residual based on the aforementioned operating condition segment. Multi-scale change point identification is performed on the process data and the residuals to generate a mutation fingerprint including the change point time, residual direction, duration and magnitude. The mutation fingerprint is matched with the pump and valve control commands for interpretability, and a warning information on the mutation of hazardous chemical transportation status is generated based on the matching result.

2. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 1, characterized in that, The process data is time-aligned and divided into operating condition segments, including: using the trigger time of the pump valve control command as the segment anchor point, and correcting the sampling period of the pressure, flow and temperature sampling sequences; Combine pump frequency change rate, valve position change direction, flow setting change and medium identification into operating condition labels; When adjacent operating condition labels switch, a transition buffer segment is retained, and the transition buffer segment is stored separately from the steady-state conveying segment, start-stop transition segment, valve switching segment and flow regulation segment. The corresponding time boundary and conveying stage mark are recorded for each operating condition segment.

3. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 2, characterized in that, The calculation of pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual and temperature compensation residual according to the operating condition segment includes: selecting synchronous sampling points for pressure, flow, temperature, pump frequency and valve position in each operating condition segment, and establishing local reference mapping between each state quantity. Pressure-flow correlation residuals are generated based on the synchronous relationship between pressure and flow rate changes; Pump frequency flow hysteresis residual is generated based on the chronological relationship between pump frequency adjustment time and flow response time; Generate valve position pressure direction residual based on the valve position movement direction and the upstream and downstream pressure change direction; After compensating the pressure sequence based on the temperature change, a temperature compensation residual is generated, and the four types of residuals are archived according to the operating condition labels.

4. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 3, characterized in that, Multi-scale change point identification is performed on the process data and the residuals, including: calculating candidate points for state variable abrupt changes and candidate points for residual abrupt changes in short time windows, medium time windows and long time windows respectively; Adjacent candidate points within the same working condition segment are merged into variable point clusters based on their temporal overlap relationship. For each variable cluster, record the time of the change, the direction of the change, the residual category, the duration, the amplitude level, and the corresponding operating condition label. When a state variable mutation candidate point and at least one type of residual mutation candidate point appear in the same variable point cluster, the mutation fingerprint is generated, and isolated candidate points that are not included in the variable point cluster are recorded as sampling points to be confirmed.

5. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 2, characterized in that, The method of retaining a transition buffer segment when adjacent operating condition labels switch includes: determining the buffer start and end boundaries by the reverse point of the pressure slope before and after the pump valve control command is triggered, the start point of the flow response and the stable point of temperature sampling. When multiple pump and valve control commands appear in the buffer section, control commands with a time interval less than a multiple of the preset sampling interval are merged into a composite operating condition label, and the order of each control command in the composite operating condition label is recorded. An independent reference mapping is set for the state variable sequence under the composite operating condition label, and the residual is not calculated together with the steady-state delivery segment. The corresponding sampling point sequence number is configured for the independent reference mapping.

6. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 5, characterized in that, The calculation of pressure-flow correlation residual, pump frequency-flow hysteresis residual, valve position pressure direction residual and temperature compensation residual according to the operating condition segment also includes: setting residual baselines according to the operating condition label and the composite operating condition label respectively. Before updating the residual baseline, discard sample segments containing unexplained variable points; Convert the residual values ​​into direction codes, deviation levels, and persistence counts; When at least two types of residuals exhibit directional coupling within the same time frame, a residual combination identifier is generated and written into the mutation fingerprint, while the corresponding sampling point number is saved in the archived record.

7. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 6, characterized in that, Multi-scale change point identification is performed on the process data and the residual, including: using short time window candidate points as cluster centers, using medium time window candidate points to confirm the persistence boundary, and using long time window candidate points to confirm the steady-state offset interval. When the mutation directions corresponding to short time windows, medium time windows and long time windows are inconsistent, retain the combination with the smallest time distance between residual mutation candidate points and state variable mutation candidate points. Short time window candidate points that are not confirmed by medium and long time windows are marked as transient candidate points, and the transient candidate points are associated with the residual combination identifier and stored.

8. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 7, characterized in that, The interpretability matching of the mutation fingerprint with the pump and valve control commands includes: searching for corresponding commands for pump frequency adjustment, valve position action, start / stop switching and flow setting change before and after the change point of the mutation fingerprint; Match the instruction direction and instruction sequence with the residual direction, mutation direction, and duration; The mutation fingerprint of a successfully matched combination whose residual combination identifier belongs to the permitted combination of the corresponding working condition label is marked as a working condition disturbance. Mutation fingerprints with no corresponding instruction or whose instruction direction conflicts with the residual direction are marked as abnormal mutation candidates, and the instruction number on which the matching is based is recorded.

9. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 8, characterized in that, The step of generating early warning information for sudden changes in the status of hazardous chemical transport based on the matching results includes: writing the mutation fingerprints marked as candidates for abnormal mutations into the edge alarm queue according to the residual combination identifier, amplitude level and duration, and retaining the matching results corresponding to the mutation fingerprints; When the same residual combination identifier appears consecutively within the same working condition segment, the edge alarm queue status will be switched from observation status to warning status. When a new mutation fingerprint in the early warning state has the same residual direction as the previous mutation fingerprint and the magnitude level does not drop, the output includes early warning information including the transport stage marker, residual category, and change point time.

10. The method for early warning of sudden changes in the state of hazardous chemical transportation based on edge computing according to claim 9, characterized in that, After the edge alarm queue status is switched from observation status to warning status, it also includes: performing backtracking verification on the working condition segment corresponding to the warning status, using the pump valve control command, the working condition label, the composite working condition label, the residual baseline and the permission combination as a verification index; When the backtracking verification finds that the mutation fingerprint falls within the independent reference mapping range of the composite working condition label, the edge alarm queue status is reset to the observation status and the verification record is retained. When the backtracking verification fails to hit the independent benchmark mapping range and the residual combination identifier continues to update, an acknowledgment alarm message is generated and the corresponding mutation fingerprint sequence is locked.

Images