AI-based heating system fault early warning and self-diagnosis method

Abstract

The application discloses an AI-based heating system fault early warning and self-diagnosis method, which comprises the following steps: collecting pipe network topology, control instructions, execution feedback and operation monitoring data, extracting and pairing control events and execution events, and respectively constructing target working condition paths and actual working condition paths. Through an expected response model, corresponding response sequences are generated, target residuals and actual residuals are formed after alignment with operation monitoring sequences, and fault evidence sequences are formed in combination with physical constraints to determine fault categories and associated objects. According to fault evidence control sample access and model updating, the source relationship between samples and model versions is established; when the access sample is verified as a fault sample, the affected model version is revoked and the previous model version is restored. The application can distinguish working condition switching and fault anomalies, and reduce the risk of gradually polluting the model with fault samples.

Description

Technical Field

[0001] This application relates to the field of smart heating technology, and in particular to an AI-based method for early warning and self-diagnosis of heating system faults. Background Technology

[0002] Urban centralized heating systems include heat sources, pipe networks, heat exchange stations, circulating pumps, regulating valves, heat exchangers, and temperature, pressure, and flow detection devices. During operation, the heat source combination, pump frequency, valve opening, and heat exchange parameters need to be dynamically adjusted based on outdoor temperature, user load, heat source capacity, and hydraulic conditions. Existing fault early warning methods mostly employ fixed thresholds, historical normality models, neural networks, or digital twins to identify leaks, blockages, equipment efficiency degradation, and sensor anomalies.

[0003] However, normal operations such as heat source switching, pump unit addition / removal, valve adjustment, and branch switching can also cause fluctuations in pressure, flow, and temperature. Due to time differences between control commands, actuator actions, and measurement point responses, and the delay in the transport of the heat medium in the pipeline network, the transient response resulting from normal switching may resemble the early signs of valve jamming, pump performance degradation, pipeline leakage, or sensor drift. Existing methods easily misjudge normal switching as a fault, or suppress actual faults superimposed during switching as normal fluctuations. Therefore, it is necessary to address how to combine control commands, execution feedback, and thermal delays to isolate the expected response of operating condition switching and identify superimposed faults.

[0004] Furthermore, to adapt to changes in load and environment, existing AI models often utilize rolling data for online updates. Early examples such as slow leaks, heat exchanger scaling, decreased pump efficiency, or sensor drift may be mistakenly written into the model as new normal operating conditions, causing the normal baseline to gradually shift towards a fault state. This leads to a decrease in the anomaly score, or even the alarm disappearing while the fault is still developing. Therefore, it is necessary to address how to isolate suspected anomaly samples and control model updates based on command feedback consistency, mass balance, heat balance, and subsequent verification results to prevent progressive faults from being learned as normal states. Summary of the Invention

[0005] To address the aforementioned problems, embodiments of the present invention provide an AI-based method for early warning and self-diagnosis of heating system faults, the method comprising: Collect pipeline topology data, control command sequences, execution feedback sequences, and operation monitoring sequences; Control events and execution events are extracted from the control command sequence and execution feedback sequence respectively and paired to form an event pairing state; based on the pipeline topology data, a target operating condition path is formed with the target state of the control event and an actual operating condition path is formed with the actual state of the execution event. The target operating condition path and the actual operating condition path are input into the expected response model to generate the target expected response sequence and the actual expected response sequence. The response arrival time scale is aligned with the operation monitoring sequence to form the target residual sequence and the actual residual sequence. The operation monitoring sequence generates a physical constraint sequence. The event pairing status, the target residual sequence, the actual residual sequence and the physical constraint sequence are combined into a fault evidence sequence. The fault evidence sequence is input into the fault diagnosis model to generate a fault warning record. The fault warning record includes the fault category, associated objects and fault score. The samples from the start of the control event to the verification time point are placed into the pending sample set. The admission status is determined based on the fault evidence sequence. The fault diagnosis model is updated with the admission samples and the model version is registered. Establish the update source relationship between the admission sample and the model version; if the admission sample is verified as a faulty sample, determine and revoke the affected model version along the update source relationship, and restore the previous model version of the update source of the fault-free sample.

[0006] Furthermore, the control event includes a device identifier, instruction type, target state, and instruction timestamp; the execution event includes a device identifier, feedback type, actual state, and feedback timestamp. When the device identifiers of the control event and the execution event are consistent, the instruction type corresponds to the feedback type, and the instruction timestamp is earlier than the feedback timestamp, the event pairing status is marked as paired. When the pairing conditions are not met, the event pairing status is marked as unpaired, and the actual state of the corresponding device before the control event is extracted from the execution feedback sequence, and the extracted actual state is written into the actual operating condition path.

[0007] Furthermore, both the target operating condition path and the actual operating condition path include a control object node, an upstream monitoring node, a downstream monitoring node, and pipe segments connecting adjacent nodes. The control object node, upstream monitoring node, and downstream monitoring node are arranged according to the direction of the heat medium flow. The target state is written into the control object node of the target operating condition path, and the actual state is written into the control object node of the actual operating condition path. Based on the pipe segment volume and the instantaneous flow rate in the operating monitoring sequence, a timescale corresponding to the temperature data is generated in the target response arrival timescale. Based on the response order of the pressure data and instantaneous flow rate in the verified normal control event, a timescale corresponding to the pressure data and instantaneous flow rate is generated in the target response arrival timescale. When the actual state changes, an actual response arrival timescale is generated along the actual operating condition path according to the same data type correspondence. When the actual state does not change, a no-response mark arrival timescale is generated.

[0008] Furthermore, the expected response model includes a state encoding layer, a topology propagation layer, and a response decoding layer. The state encoding layer combines the supply water temperature, return water temperature, supply water pressure, return water pressure, and instantaneous flow rate before the control event with the target state to form a target state code, and combines the supply water temperature, return water temperature, supply water pressure, return water pressure, and instantaneous flow rate with the actual state to form an actual state code. The topology propagation layer propagates the target state code according to the target operating condition path and the target response arrival time scale, and propagates the actual state code according to the actual operating condition path and the actual response arrival time scale. Among them, the temperature state is propagated according to the time scale corresponding to the temperature data, and the pressure state and flow rate state are propagated according to the time scale corresponding to the pressure data and instantaneous flow rate. The response decoding layer outputs the expected temperature value, expected pressure value, and expected flow rate value corresponding to the monitoring node, forming the target expected response sequence and the actual expected response sequence to determine the initial state of the node before the event occurs.

[0009] Furthermore, the target residual sequence is formed by subtracting the operational monitoring sequence from the target expected response sequence according to the target response arrival time markers corresponding to the data types; the actual residual sequence is formed by subtracting the operational monitoring sequence from the actual expected response sequence according to the actual response arrival time markers corresponding to the data types; the target residual sequence and the actual residual sequence are subtracted for the same monitoring node and the same acquisition time marker to form the instruction execution deviation component; the actual residual sequence is arranged according to the actual response arrival time marker to form the residual propagation component; constraint deviation terms are extracted from the physical constraint sequence to form the physical deviation component; the instruction execution deviation component, residual propagation component, and physical deviation component are written into the fault evidence sequence item according to the connection relationship between the controlled object node, upstream monitoring node, and downstream monitoring node to form the physical deviation component; the instruction execution deviation component, residual propagation component, and physical deviation component are written into the fault evidence sequence according to the connection relationship between the controlled object node, upstream monitoring node, and downstream monitoring node.

[0010] Furthermore, using pipeline branch nodes as mass balance nodes, mass balance terms are formed based on the instantaneous flow rates entering and leaving the pipeline branch nodes; using heat exchangers as heat power balance nodes, instantaneous primary heat power is formed based on the primary side supply water temperature, primary side return water temperature, primary side instantaneous flow rate, and the density and specific heat capacity of the primary side heat medium; instantaneous secondary heat power is formed based on the secondary side supply water temperature, secondary side return water temperature, secondary side instantaneous flow rate, and the density and specific heat capacity of the secondary side heat medium; and a heat power balance term is formed based on the difference between the primary and secondary side instantaneous heat power; using the pipe section where the regulating valve or circulating pump is located as the pressure and flow constraint pipe section, pressure and flow relationship terms are formed based on the pressure difference across the regulating valve, the actual opening degree of the regulating valve, and the instantaneous flow rate, or based on the difference between the inlet and outlet pressures of the circulating pump, the actual frequency of the circulating pump, and the instantaneous flow rate; and arranging the mass balance terms, heat power balance terms, and pressure and flow relationship terms according to the response arrival time scale to form a physical constraint sequence.

[0011] Furthermore, a baseline segment prior to the control event is extracted from the operational monitoring sequence, and a verification segment is extracted between the termination of the actual expected response sequence and the verification time marker. The baseline segment and the actual state before the control event are input into the expected response model to generate a baseline expected sequence. The operational monitoring values ​​in the baseline segment are aligned with the baseline expected sequence and the difference is calculated to form a baseline residual sequence. Based on the baseline residual sequence, residual baseline relationships and physical constraint baseline relationships are formed. The event pairing status is set as a pairing access condition, the target residual sequence and the actual residual sequence meeting the residual baseline relationship within the verification segment are set as residual access conditions, and the physical constraint sequence meeting the physical constraint baseline relationship within the verification segment are set as constraint access conditions. When the pairing access conditions, residual access conditions, and constraint access conditions are all met, the corresponding sample is marked as an access sample; otherwise, the corresponding sample is retained in the pending sample set.

[0012] Furthermore, the fault diagnosis model includes a condition adaptation branch, a fault memory branch, and an evidence fusion layer. A fault memory sample set is formed using verified progressive fault samples, and the fault memory branch is trained based on this set. The condition adaptation branch is updated using admission samples, while maintaining the model parameters of the fault memory branch and the evidence fusion layer unchanged during the update process, thus forming a candidate model. Verified normal samples and verified progressive fault samples of the same control event type are selected to form a verification sample pair. The verification sample pair is input into the candidate model, and when the fault score output by the candidate model for the verified progressive fault sample is higher than the fault score output for the verified normal sample, the candidate model is registered as a model version.

[0013] Further, the update source relationship includes sample identifier, training batch identifier, parent model version, and child model version; based on the sample identifier of the faulty sample, query the training batch identifier containing the faulty sample, and determine the corresponding child model version and subsequent model versions with the child model version as the parent model version as the affected model version; trace back along the inheritance direction of the parent model version, and determine the most recent model version in the update source relationship that does not contain the faulty sample as the preceding model version; revoke the affected model version, transfer the faulty sample to the fault memory sample set, and reprocess the corresponding fault evidence sequence using the preceding model version.

[0014] Furthermore, the AI-based method for early warning and self-diagnosis of heating system faults also includes: forming a topological contradiction record based on the observation response timescale and the actual response arrival timescale of each monitoring node to identify suspected topological pipe segments; assigning connected or blocked states to the suspected topological pipe segments respectively, reconstructing the actual operating path and verifying the physical constraint sequence to determine the effective topological state, thereby updating the pipeline topology data and regenerating the fault evidence sequence.

[0015] The technical effects and advantages of the AI-based fault early warning and self-diagnosis method for heating systems provided by this invention are as follows: This invention constructs a dual-condition path (target and actual) and combines physical constraints and a sample source tracing and update mechanism to achieve accurate early warning, self-diagnosis, and safe model iteration for heating faults. Based on control commands, execution feedback, and response arrival time scales, this invention generates the target expected response and the actual expected response, distinguishing between transient fluctuations caused by normal operating condition switching and abnormal responses caused by equipment and pipeline faults. It jointly analyzes the target residual, actual residual, and relationships with mass balance, heat balance, and pressure-flow, providing fault judgment with a basis in pipeline topology and thermal operation, and determining the fault category and associated objects. By controlling model updates through a pending sample set and admission conditions, it avoids progressively faulty samples being mistakenly included in the normal sample set, maintaining the fault diagnosis model's ability to identify slow anomalies. It establishes an update source relationship between admission samples and model versions, revoking the affected model version when a sample is verified as a faulty sample, thus achieving traceability, isolation, and recovery of the model update process. Attached Figure Description

[0016] Figure 1 This is a flowchart of the AI-based heating system fault early warning and self-diagnosis method in Example 1; Figure 2 This is a flowchart of the event pairing and working condition path generation method in Example 1; Figure 3 This is a flowchart of the topology correction method in Example 2. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0018] Please see Figure 1 As shown, embodiments of the present invention provide an AI-based method for early warning and self-diagnosis of heating system faults, the method comprising: S1. Collect pipeline topology data, control command sequences, execution feedback sequences, and operation monitoring sequences; S2. Extract control events and execution events from the control command sequence and execution feedback sequence respectively and pair them to form an event pairing state; based on the pipeline topology data, form a target operating condition path with the target state of the control events and form an actual operating condition path with the actual state of the execution events; S3. Input the target operating condition path and the actual operating condition path into the expected response model to generate the target expected response sequence and the actual expected response sequence; align the response arrival time scale with the operation monitoring sequence to form the target residual sequence and the actual residual sequence; generate a physical constraint sequence from the operation monitoring sequence; combine the event pairing status, the target residual sequence, the actual residual sequence, and the physical constraint sequence into a fault evidence sequence; input the fault evidence sequence into the fault diagnosis model to generate a fault warning record; the fault warning record includes a fault category, associated objects, and a fault score; the fault category is used to indicate abnormal control valve execution, abnormal circulation pump operation, pipe section leakage, abnormal heat exchanger performance, or abnormal monitoring device; the associated objects are represented by the equipment node identifier, monitoring node identifier, or pipe section identifier in the pipeline topology data.

[0019] S4. Place the samples from the start of the control event to the verification time point into the pending sample set, determine the admission status based on the fault evidence sequence, update the fault diagnosis model with the admission samples and register the model version. S5. Establish the update source relationship between the admission sample and the model version; if the admission sample is verified as a faulty sample, determine and revoke the affected model version along the update source relationship, and restore the previous model version of the update source of the fault-free sample.

[0020] The samples in this embodiment include at least control events, event pairing status, target operating condition path, actual operating condition path, target residual sequence, actual residual sequence, physical constraint sequence, fault evidence sequence, sample start time stamp, verification time stamp, and verification status, and establish the association between the data through sample identifiers.

[0021] In this embodiment, the control command sequence is formed by arranging the control records issued by the heating control platform to the regulating valve, circulating pump and other controlled equipment in chronological order; the execution feedback sequence is formed by arranging the equipment status records sent by the field controller or equipment feedback component in chronological order. The control command sequence and the execution feedback sequence adopt the same equipment coding rules and clock reference so that the command records and feedback records of the same controlled equipment can correspond.

[0022] Control events are extracted one by one from the control command sequence. Each control event includes a device identifier, command type, target state, and command timestamp. The device identifier indicates the physical device to which the control command is applied and corresponds to the device node in the pipeline topology data. The command type indicates the type of control action, including regulating valve opening, circulating pump start / stop, or circulating pump frequency adjustment. The target state indicates the opening degree, start / stop state, or operating frequency that the control command requires the device to achieve. The command timestamp indicates the time when the heating control platform issued the control command.

[0023] Execution events are extracted from the execution feedback sequence. Each feedback record in the execution feedback sequence includes a device identifier, feedback type, actual status, and feedback timestamp. When the device controller sends device action confirmation information, or when the device status changes relative to the previous valid feedback record, the corresponding feedback record is identified as an execution event. The feedback type is used to indicate the data type to which the actual status belongs. The actual status is used to indicate the actual opening degree, actual start / stop status, or actual operating frequency of the device feedback. The feedback timestamp is used to indicate the time when the actual status is confirmed or collected by the field controller.

[0024] Before performing event pairing, a correspondence is established between command types and feedback types. The control valve opening adjustment corresponds to the actual control valve opening feedback, the circulating pump start / stop command corresponds to the circulating pump operating status feedback, and the circulating pump frequency adjustment command corresponds to the circulating pump actual frequency feedback. The correspondence serves as the data basis for event pairing, and the similarity between the command name or feedback name is not used for judgment.

[0025] For any control event, firstly, the execution events of the same device are determined from the execution feedback sequence based on the device identifier. Then, the retrieval is performed in order of feedback timestamp from first to last. Execution events whose feedback type corresponds to the instruction type of the control event and whose feedback timestamp is later than the instruction timestamp are identified as candidate execution events. To prevent subsequent control instruction feedback from being incorrectly paired with earlier control events, the feedback timestamp of the candidate execution event is located between the instruction timestamp of the control event and the instruction timestamp of the next control event of the same device. When candidate execution events exist, the candidate execution event with the earliest time sequence is paired with the control event, and the event pairing status is marked as paired.

[0026] After pairing is completed, the corresponding control object node is queried in the pipeline topology data according to the device identifier. The target state of the control event is written into the control object node in the target operating condition path, and the actual state of the execution event is written into the corresponding control object node in the actual operating condition path. The target operating condition path consists of the device state required by the control command, and the actual operating condition path consists of the device state confirmed by the execution feedback. The pairing relationship between the control event and the execution event is saved along with the corresponding control object node for subsequent use in generating the target expected response sequence and the actual expected response sequence.

[0027] When there is no candidate execution event between the instruction time stamp of a control event and the instruction time stamp of the next control event of the same device, the event pairing status is marked as unpaired. For control events in the unpaired state, the most recent valid feedback record with the same device identifier and corresponding feedback type before the control event is found in the execution feedback sequence from back to front according to the feedback time stamp, and the device state recorded in the valid feedback record is determined as the actual state before the control event occurred.

[0028] Based on the equipment identifier, the corresponding control object node is located in the pipeline topology data. The actual state before the control event occurs is written into the control object node in the actual operating path, instead of replacing the actual state with the target state of the control event. As a result, the equipment that has not received action feedback maintains the equipment state before the control event occurs in the actual operating path. The corresponding control event, actual state and unpaired state establish an association record and are jointly input into the subsequent fault evidence sequence generation process.

[0029] For example: The heating control platform issues an opening adjustment command to a regulating valve, but before the next control command for the regulating valve is generated, no corresponding actual opening change record appears in the execution feedback sequence. In this case, the control event is marked as unpaired. The system extracts the most recent valid actual opening from the feedback record before the control event and writes the actual opening into the actual operating condition path. At the same time, the target opening in the control command is written into the target operating condition path. The target operating condition path and the actual operating condition path retain the command requirement status and feedback confirmation status, respectively. Subsequently, the corresponding expected response sequence is generated according to the two operating condition paths.

[0030] like Figure 2As shown, the control command sequence is used to extract control events, and the execution feedback sequence is used to extract execution events. Control events and execution events are paired to form an event pairing state. In the paired state, the actual state corresponding to the execution event is written into the actual operating condition path. In the unpaired state, the actual state before the control event is extracted and written into the actual operating condition path. At the same time, the controlled object node, upstream monitoring node and downstream monitoring node are determined according to the pipeline topology data, and the target operating condition path and the actual operating condition path are constructed accordingly. Then, the target response arrival time scale and the actual response arrival time scale are calculated respectively.

[0031] In this embodiment, the pipeline topology data includes node records and pipe segment records. Node records are used to record node identifiers, node types, and connected pipe segments. Pipe segment records are used to record the start point, end point, and volume of the pipe segment. The controlled object node is the regulating valve node or circulating pump node corresponding to the control event. The upstream monitoring node is the monitoring node located before the controlled object node along the direction of heat medium flow and has valid operational monitoring data. The downstream monitoring node is the monitoring node located after the controlled object node along the direction of heat medium flow and has valid operational monitoring data.

[0032] The direction of heat medium flow is determined based on the designed transport direction of the pipe section and the instantaneous flow direction in the operation monitoring sequence. Starting from the controlled object node, the adjacent effective monitoring node is searched in the opposite direction of the heat medium flow to determine the upstream monitoring node; the adjacent effective monitoring node is searched along the heat medium flow to determine the downstream monitoring node, and the pipe section connecting the upstream monitoring node, the controlled object node and the downstream monitoring node is written into the corresponding operating condition path.

[0033] The target operating condition path and the actual operating condition path are constructed using the same node identifier and pipe segment identifier. The target state of the control event is written into the control object node in the target operating condition path, and the actual state of the execution event is written into the control object node in the actual operating condition path. When the event pairing state is unpaired, the actual state extracted before the control event occurs is written into the actual operating condition path. When the device corresponding to the control object node is in a blocked state, the downstream extension of the corresponding path is terminated; when it is in a connected state, the connection relationship between the control object node and the downstream pipe segment is preserved. When the actual operating condition path contains a no-response marker, the actual expected response sequence is formed based on the initial state of the node before the control event occurs. When the actual operating condition path does not contain a no-response marker, the actual state code is propagated according to the actual operating condition path and the actual response arrival time stamp.

[0034] The response arrival time stamp is used to record the estimated time when the state change of the controlled object node propagates to the monitoring node, and the corresponding time stamp is saved according to the data type. For temperature data, the transport volume is accumulated according to the acquisition order of the operation monitoring sequence based on the pipe section volume and the corresponding instantaneous flow rate. When the accumulated transport volume reaches the current pipe section volume, the corresponding time stamp is determined as the time stamp of the temperature data at the current monitoring node.

[0035] For pressure data and instantaneous flow, the collection time stamps at which the pressure data and instantaneous flow of each monitoring node begin to deviate from the baseline operating state are extracted from the verified normal control events corresponding to the instruction type, control object node type, and operating condition path. The response order is formed according to the arrangement relationship of the monitoring nodes in the operating condition path, and the collection time stamps are used as the time stamps of the corresponding data types.

[0036] When generating response arrival time stamps along the target operating condition path, the command time stamp is used as the calculation starting point; when generating response arrival time stamps along the actual operating condition path, the feedback time stamp is used as the calculation starting point; the time stamps corresponding to each data type generated by the target operating condition path together constitute the target response arrival time stamp, and the time stamps corresponding to each data type generated by the actual operating condition path together constitute the actual response arrival time stamp.

[0037] When the controlled object node connects to a branch pipe segment, the pipe segment connected to the controlled object node is extended along the current heat medium flow direction, and each branch pipe segment terminates at the downstream monitoring node first queried along the branch. The path branch is formed by the controlled object node, the corresponding downstream monitoring node and the pipe segment between them. The operation monitoring data of the upstream monitoring node is used to form the initial operating state before the control event occurs, and is written into the target operating condition path and the actual operating condition path respectively.

[0038] When the event pairing status is paired and the actual status changes relative to the actual status before the control event occurs, the feedback time stamp is used as the starting time stamp for the actual operating path, and the actual response arrival time stamp is generated. When the event pairing status is unpaired, or the actual status has not changed, a no-response flag is written into the actual operating path, and the initial operating status before the control event occurs is used as the status basis for the actual operating path.

[0039] When calculating the response arrival time, the transport volume of the current pipe segment between adjacent acquisition time points is accumulated according to the acquisition sequence of the operation monitoring sequence. When the accumulated transport volume reaches the current pipe segment volume, the transport termination time point of the current pipe segment is determined. If the instantaneous flow is missing, zero, or flows in the opposite direction to the current path, the cumulative calculation of the current path is paused, and the calculation continues after obtaining a valid instantaneous flow consistent with the current path flow direction.

[0040] In this embodiment, the expected response model is trained using historical operation records verified as normal control processes. Each historical operation record includes the operation monitoring sequence before the control event, the target state of the control event, the actual state of the execution event, the target operating condition path, the actual operating condition path, the response arrival time scale, and the operation monitoring data of each monitoring node after the control event. The operation monitoring data after the control event serves as the basis for the expected output during model training.

[0041] The status coding layer is used to convert the equipment status and the pipeline operation status before the occurrence of the control event into status codes that can be transmitted along the operating path. Based on the command time stamp of the control event, the most recent valid supply water temperature, return water temperature, supply water pressure, return water pressure and instantaneous flow rate before the time stamp are extracted from the operation monitoring sequence and written to the corresponding node according to the monitoring node identifier. The most recent valid data refers to data with normal acquisition status, data type matching the monitoring node and acquisition time stamp not later than the command time stamp.

[0042] The state coding layer arranges the supply water temperature, return water temperature, supply water pressure, return water pressure, and instantaneous flow rate of each monitoring node in a fixed data order to form the node initial state. For the target operating condition path, the target state of the control event is combined with the node initial state of the controlled object node to form the target state code. For the actual operating condition path, the actual state of the execution event is combined with the node initial state of the controlled object node to form the actual state code. The target state code and the actual state code are generated using the same set of model parameters of the state coding layer, so that the difference between the two state codes comes from the target state, the actual state, and the corresponding operating condition path, rather than from the difference in coding rules.

[0043] The topology propagation layer sequentially reads the control object nodes, pipe segments, and monitoring nodes in the operating path according to the direction of the heat medium flow. When propagating the target state code, the time position of the target state code arriving at each monitoring node is determined by the target response arrival time stamp. When propagating the actual state code, the time position of the actual state code arriving at each monitoring node is determined by the actual response arrival time stamp. Before the response arrival time stamp, the initial state of the corresponding monitoring node is retained. From the response arrival time stamp, the state code transmitted by the previous node, the initial state of the current monitoring node, and the arrival time stamp difference between adjacent nodes are input into the propagation unit.

[0044] In one implementation, the propagation unit in the topology propagation layer employs a gated temporal propagation unit. The propagation unit connects the propagation state code of the previous node, the initial state of the current monitoring node, and the arrival time difference between adjacent nodes to form the propagation input. Based on the propagation input, the state retention coefficient and the state reception coefficient are calculated respectively. The state retention coefficient is used to determine the proportion of the initial state of the current monitoring node in the current propagation state code, and the state reception coefficient is used to determine the proportion of the propagation state code of the previous node in the current propagation state code. The product of the initial state of the node and the state retention coefficient, and the product of the propagation state code of the previous node and the state reception coefficient are added to form the propagation state code of the current monitoring node. The state retention coefficient and the state reception coefficient are calculated by the trainable weight matrix, bias term, and activation function of the propagation unit, and both use the same input data. The propagation state code of the current monitoring node is used as the propagation state code of the previous node of the next monitoring node, until the propagation processing of each monitoring node in the corresponding working condition path is completed.

[0045] The response decoding layer includes a temperature decoding unit, a pressure decoding unit, and a flow decoding unit. The temperature decoding unit encodes and outputs the expected value of the supply water temperature or the expected value of the return water temperature corresponding to the configuration of the monitoring node according to the propagation status of the monitoring node. The pressure decoding unit outputs the expected value of the supply water pressure or the expected value of the return water pressure. The flow decoding unit outputs the expected value of the flow rate. Monitoring nodes without corresponding detection devices do not generate expected values ​​of this data type.

[0046] The expected values ​​of temperature, pressure, and flow rate of each monitoring node are arranged according to the monitoring node identifier and the response arrival time. The data obtained by the target state code propagation forms the target expected response sequence, and the data obtained by the actual state code propagation forms the actual expected response sequence. The target expected response sequence and the actual expected response sequence use the same data fields and arrangement rules so that they can be aligned with the running monitoring sequence in the future.

[0047] When training the expected response model, the target state or actual state in the historical operation record, the operation monitoring data before the control event, the corresponding working condition path and the response arrival time stamp are used as model inputs. The actual temperature, pressure and flow rate collected after the control event are used as training labels. The deviations between the expected temperature value, expected pressure value and expected flow rate value and the corresponding training labels are calculated respectively. The model parameters of the state encoding layer, topology propagation layer and response decoding layer are adjusted according to the deviations. Data items that lack valid training labels are not included in the adjustment of the corresponding model parameters.

[0048] When training the expected response model, temperature prediction loss, pressure prediction loss, and flow prediction loss are respectively formed based on the differences between the expected temperature value and the temperature training label, the expected pressure value and the pressure training label, and the expected flow rate and the flow rate training label. The prediction losses with valid training labels are then weighted and summed to form the model training loss. The trainable weight matrices and bias terms in the state encoding layer, topology propagation layer, and response decoding layer are updated in reverse based on the model training loss. Training parameters include at least the input data dimension, propagation state encoding dimension, batch size, parameter update step size, and stopping condition, and are saved with the model version. Training stops when the model training loss corresponding to the validation data no longer decreases, or when the preset stopping condition is reached.

[0049] The state encoding layer employs a feature mapping unit to map the initial state of nodes to the device state as a state vector; the topology propagation layer uses a time-series propagation unit that shares parameters based on network adjacency relationships, updating the state vector node by node according to the response arrival time; the response decoding layer includes temperature, pressure, and flow output units. During training, prediction bias is calculated only for data items with valid training labels, and the model parameters are updated based on the prediction bias.

[0050] In this embodiment, the target residual sequence and the actual residual sequence are established according to the monitoring node identifier, data type, and acquisition time scale. The data types include supply water temperature, return water temperature, supply water pressure, return water pressure, and instantaneous flow rate. The supply water temperature and return water temperature are aligned with the time scale corresponding to the temperature data in the response arrival time scale. The supply water pressure, return water pressure, and instantaneous flow rate are aligned with the time scale corresponding to the pressure data and instantaneous flow rate in the response arrival time scale. For any monitoring node, the expected value in the target expected response sequence from the target response arrival time scale is matched with the monitoring value of the same monitoring node and the same data type in the running monitoring sequence. The expected value in the actual expected response sequence from the actual response arrival time scale is matched with the monitoring value in the same way. When the time scale of the expected value is between adjacent acquisition time scales, the corresponding monitoring value is determined according to the adjacent monitoring value and its acquisition time scale.

[0051] Using the difference between the monitored value and the corresponding expected value as the unified direction of the difference calculation, the records obtained by subtracting the monitored sequence from the target expected response sequence are arranged according to the monitoring node identifier, data type, and collection time stamp to form the target residual sequence; the records obtained by subtracting the monitored sequence from the actual expected response sequence are arranged according to the same fields to form the actual residual sequence. Data types that are not equipped with corresponding detection devices or lack valid monitoring values ​​are not included in the difference calculation.

[0052] Based on the monitoring node identifier, data type, and acquisition time stamp, the target residual sequence is matched with the actual residual sequence item by item, and the actual residual is subtracted from the target residual to form the instruction execution deviation component. Each instruction execution deviation component records the corresponding control event, control object node, monitoring node, and data type.

[0053] According to the node arrangement order in the actual working condition path, the residual records corresponding to each monitoring node in the actual residual sequence are arranged according to the actual response arrival time, forming residual propagation components. The residual propagation components record at least the monitoring node identifier, the preceding node identifier, the actual response arrival time, the data type, and the residual value. When there are path branches, residual propagation components are formed according to each path branch.

[0054] The mass balance term, heat power balance term, and pressure-flow relationship term in the physical constraint sequence record the calculated values ​​and constraint deviations of the corresponding constraint relationships, respectively. The constraint deviations are extracted as constraint deviation terms and arranged according to the constraint nodes and acquisition time scales to form physical deviation components.

[0055] Finally, starting from the controlled object node, the upstream and downstream monitoring nodes connected to it are queried according to the pipeline topology data. The instruction execution deviation component, residual propagation component and physical deviation component corresponding to the same control event are bound to the corresponding node and pipe segment. They are written into the fault evidence sequence in the order of control event, path branch, monitoring node and acquisition time stamp. Each record in the fault evidence sequence retains its source sequence and source data identifier.

[0056] In this embodiment, the physical constraint sequence consists of a mass balance term, a heat power balance term, and a pressure-flow relationship term. Before generating each constraint term, the supply water temperature, return water temperature, pressure, and instantaneous flow rate are extracted from the operation monitoring sequence according to the monitoring node identifier and the acquisition time scale. The data records participating in the same constraint calculation are determined according to the response arrival time scale.

[0057] A pipeline branch node refers to a node in the pipeline topology data that connects an incoming pipe segment and at least one outgoing pipe segment. Based on the direction of the heat medium flow, the inflow or outflow attributes of each connected pipe segment relative to the pipeline branch node are determined. The instantaneous flow rates entering the pipeline branch node are summarized, and the instantaneous flow rates leaving the pipeline branch node are summarized. The difference between the two is used to form a mass balance term. The mass balance term records the pipeline branch node, the pipe segments involved in the calculation, the acquisition time scale, and the flow rate difference.

[0058] The heat exchanger serves as a heat power balance node. The instantaneous heat power on the primary side is calculated based on the difference between the primary side supply and return water temperatures, the instantaneous flow rate on the primary side, and the density and specific heat capacity of the primary side heat transfer medium. Similarly, the instantaneous heat power on the secondary side is calculated based on the difference between the secondary side supply and return water temperatures, the instantaneous flow rate on the secondary side, and the density and specific heat capacity of the secondary side heat transfer medium. The density and specific heat capacity of the heat transfer medium are retrieved from preset property data based on the type of heat transfer medium and corresponding temperature data. A heat power balance term is formed by the difference between the instantaneous heat power on the primary side and the instantaneous heat power on the secondary side, and the heat exchanger identifier, data acquisition time stamp, and data identifiers involved in the calculation are written into this heat power balance term.

[0059] The pressure-flow constrained pipe section is defined as a section equipped with a regulating valve or circulating pump, with pressure detection devices installed on both sides of the corresponding equipment and flow detection devices installed in the corresponding pipeline. For the pipe section where the regulating valve is located, the pressure difference between the upstream and downstream pressures of the regulating valve is used to form the pressure differential across the regulating valve, and the pressure differential across the regulating valve, the actual opening degree of the regulating valve, and the instantaneous flow rate are used to form the valve operation record. For the pipe section where the circulating pump is located, the pressure differential between the outlet and inlet pressures of the circulating pump is used to form the inlet and outlet pressure differential of the circulating pump, and the actual frequency of the circulating pump, and the instantaneous flow rate are used to form the pump operation record. Based on the verified normal operation records of the same equipment, a benchmark correspondence is established between the pressure differential, instantaneous flow rate, and actual opening degree or actual frequency, and the deviation of the current operation record from the benchmark correspondence is used to form the pressure-flow relationship item.

[0060] According to the response arrival time scale corresponding to each constraint, the mass balance term, heat power balance term, and pressure-flow relationship term are arranged in sequence and associated with the corresponding control event, control object node, monitoring node, and pipe segment to form a physical constraint sequence. If the data involved in the calculation is missing, the acquisition status is abnormal, or the time scale cannot be matched, the corresponding constraint term is marked as an invalid constraint term and is not replaced by data from other time periods.

[0061] To ensure that the event pairing status, target residual sequence, actual residual sequence, and physical constraint sequence can be input into the fault diagnosis model in a unified data format, this implementation calculates a comprehensive fault evidence score S based on the fault evidence sequence. f (t), the comprehensive fault evidence score is calculated based on the monitoring node, data collection timescale, and physical constraints, including: ; In the formula, t This is the current data collection time stamp. ω 1 , ω 2 , ω 3 and ω 4These are the weighting coefficients for the instruction execution deviation component, residual propagation component, physical deviation component, and event pairing state, respectively. i To monitor the node index, N This represents the total number of monitoring nodes participating in the calculation. j For the first i The index of the predecessor node with upstream connection relationship for each monitoring node. U(i) For the first i The set of upstream predecessor nodes corresponding to each monitoring node, ρ i For the first i The data reliability coefficient or node weight of each monitoring node For the first i Each monitoring node at time t The self-standard residual value, For the first i Each monitoring node at time t The actual residual value, a ji Preorder node j With monitoring nodes i The path correlation coefficient between them, Δ t ji Preorder node j The residual propagation to the monitoring node i The propagation time difference M This represents the total number of physical constraint terms involved in the calculation. m For the physical constraint term index, k m For the first m The constraint weight coefficients of each physical constraint term. c m (t) For the first m A physical constraint term at time... t The constraint deviation value, ξ(t) This is the event pairing status value. The value is 1 when the event pairing status value is in the paired state and 0 when the event pairing status value is in the unpaired state.

[0062] Specifically, the target residual value and the actual residual value are extracted from the target residual sequence and the actual residual sequence, respectively, according to the monitoring node identifier, data type, and acquisition time scale; the path correlation coefficient is determined based on the connection relationship between the preceding node and the monitoring node in the pipeline topology data; the propagation time difference is determined based on the time scale difference between the actual response arrival time scales; the constraint deviation value is extracted from the mass balance term, heat power balance term, or pressure-flow relationship term. For invalid constraint terms, their corresponding constraint deviation values ​​are not calculated.

[0063] After obtaining the comprehensive fault evidence score in the above manner, the comprehensive fault evidence score is associated with the corresponding control event, control object node, fault evidence sequence and acquisition time stamp, and used as one of the input data of the fault diagnosis model. Then, taking the command time stamp of the control event as the boundary, continuous and valid data records are extracted from the operation monitoring sequence to form a baseline segment.

[0064] In this embodiment, the command time stamp of the control event is used as the boundary to extract continuous and valid data records from the operation monitoring sequence to form a reference segment. The reference segment does not contain control events that have not yet ended, and the operating status of the corresponding equipment has not changed. The reference segment is used to characterize the pipeline network operating status before the occurrence of the control event.

[0065] The termination of the actual expected response sequence means that the actual expected value corresponding to the end monitoring node in the actual operating condition path has been generated, or the no-response marker in the actual operating condition path has been recorded. From the termination of the actual expected response sequence, data continues to be extracted from the operation monitoring sequence until the on-site inspection record, maintenance work order, alarm review record or operation personnel confirmation record is obtained. The formation time mark of the corresponding record is determined as the verification time mark. The data between the termination time mark and the verification time mark constitutes the verification segment.

[0066] According to the monitoring node identifier and data type, the residual records corresponding to the target residual sequence and the actual residual sequence are extracted from the benchmark segment. The residual values ​​of the same monitoring node and the same data type are sorted to form a residual benchmark relationship to characterize the normal fluctuation range and direction of change. In the same way, the mass balance term, heat power balance term and pressure flow relationship term are extracted from the benchmark segment to form the physical constraint benchmark relationship of the corresponding constraint term under normal operating conditions.

[0067] When determining the admission status of a sample to be judged, the event pairing status of the corresponding control event is read first. If the event pairing status is "paired", the pairing admission condition is met; if the event pairing status is "unpaired", the pairing admission condition is not met.

[0068] Subsequently, according to the monitoring node identifier and data type, the target residual sequence and actual residual sequence in the verification segment are compared with the residual benchmark item by item. When the residual record in the verification segment returns to the corresponding normal fluctuation range and its change direction does not continuously deviate from the benchmark segment, the residual admission condition is met. For residual records that lack effective monitoring values, they are not used as the basis for the establishment of the residual admission condition.

[0069] Then, the mass balance term, heat power balance term, and pressure-flow relationship term in the verification section are compared with the physical constraint benchmark relationship. When each effective constraint term is restored to the corresponding benchmark relationship and there are no constraint terms that continuously deviate, the constraint admission condition is met.

[0070] When the matching access condition, residual access condition, and constraint access condition are all met, the pending sample is marked as an access sample, and the corresponding control event, verification segment, verification record, and access time scale are recorded; when any access condition is not met or the data required to form an access conclusion is missing, the sample is kept in the pending sample set and is not used for fault diagnosis model updates.

[0071] In this embodiment, the fault diagnosis model includes a working condition adaptation branch, a fault memory branch, and an evidence fusion layer. The working condition adaptation branch arranges the event pairing status, instruction execution deviation component, and residual propagation component into a working condition feature sequence according to the control event, monitoring node, and acquisition time scale, and updates the working condition state vector time-by-time through a time-series feature extraction unit. The fault memory branch arranges the residual propagation component and physical deviation component into a fault feature sequence according to the associated object and response arrival time scale, and generates a fault state vector through another time-series feature extraction unit.

[0072] The evidence fusion layer connects the operating state vector and the fault state vector corresponding to the same control event to form a fused feature vector. The classification unit generates the category output value corresponding to each fault category based on the fused feature vector. The object matching unit matches the fused feature vector with the object features corresponding to the control object node, monitoring node and pipe segment in the fault evidence sequence to generate the object output value of each candidate object. The category output value is normalized to form a fault score. The fault category corresponding to the fault score and the object identifier corresponding to the object output value are respectively determined as the fault category and the associated object.

[0073] Progressive fault samples refer to samples that are confirmed by on-site inspection records, maintenance work orders, or alarm verification records, and that show a gradual continuation of residuals or physical constraint deviations in the continuous operation monitoring sequence. The fault evidence sequence of progressive fault samples is used as training input, and the fault categories and associated objects in the verification records are used as training labels to form a fault memory sample set. When training the fault memory branch, the model parameters of the fault memory branch are adjusted according to the difference between its output and the training labels, and the correspondence between the sample identifier and the verification record is saved.

[0074] The determined admission samples are used to update the operating condition adaptation branch. During the update, the fault evidence sequence corresponding to the admission sample is used as input, and the normal state and the corresponding control event type are used as training labels. Only the model parameters of the operating condition adaptation branch are adjusted. The fault memory branch and the evidence fusion layer use the model parameters of the currently registered model. The updated operating condition adaptation branch, together with the unupdated fault memory branch and evidence fusion layer, constitute the candidate model.

[0075] When updating the operating condition adaptation branch, an operating condition classification loss is formed based on the difference between the normal state training label corresponding to the admission sample and the output of the classification unit, and an operating condition recognition loss is formed based on the difference between the control event type training label and the operating condition state vector. The operating condition classification loss and the operating condition recognition loss are combined to form the operating condition adaptation loss. Based on the operating condition adaptation loss, only the trainable weight matrix and bias terms in the operating condition adaptation branch are updated, while the model parameters of the fault memory branch and the evidence fusion layer remain unchanged. Each update saves the training batch identifier, the sample identifiers participating in the training, the model parameters before the update, the model parameters after the update, and the state in which the stopping condition is met.

[0076] The verification sample pair consists of verified normal samples and verified progressive fault samples with the same control event type, the same control object node type, and corresponding fault categories. The verification sample pairs are input into the candidate model respectively, and the fault scores corresponding to the same fault category are read. When the fault score of the verified progressive fault sample is higher than the fault score of the verified normal sample, the parameter status of each branch and evidence fusion layer in the candidate model is saved, and the corresponding model version is registered. If the order relationship is not satisfied, the registration of the candidate model is cancelled, and the currently registered model is used.

[0077] In this embodiment, when a candidate model is registered as a model version, an update source relationship is generated synchronously. The update source relationship is saved one by one according to the model version. Each record includes a sample identifier, a training batch identifier, a parent model version, and a child model version. The sample identifier is used to uniquely correspond to the admission sample used in this update. The training batch identifier is used to correspond to the set of admission samples that actually participate in training during a model update. The parent model version is the registered model version used before this update. The child model version is the model version registered after the update is completed based on the parent model version. Each sample identifier in a training batch is associated with the training batch identifier.

[0078] When an admission sample is subsequently verified as a faulty sample, the sample identifier of the faulty sample is read, and the training batch identifier containing the sample identifier is queried in the update source relationship. The sub-model version corresponding to the training batch identifier is determined as the initial affected model version, and then the subsequent update source relationship is queried again using the sub-model version as the parent model version. The sub-model version obtained in the query continues to be used as the parent model version for the next query, until there are no subsequent sub-model versions. The model versions obtained in the query process are collectively determined as the affected model versions.

[0079] When determining the preceding model version, the parent model version is read from the initial affected model version, and the query proceeds step by step forward along the parent model version. For each queried model version, the training batch identifier that generated the model version and its associated sample identifier are read. The first model version that is not associated with the fault sample and is not identified as an affected model version is identified as the preceding model version.

[0080] When an affected model version is revoked, its running status is changed to "stop using," its calling relationship with the current fault diagnosis model is terminated, and the corresponding model parameters, training batch identifier, and update source relationship are retained. Subsequently, the fault sample is removed from the admission sample record, and together with the verified fault category, associated object, and original fault evidence sequence, it is written into the fault memory sample set. The model parameters saved by the previous model version are called, the fault evidence sequence is re-inputted, the fault category, associated object, and fault score are generated, and the regenerated result is associated with the corresponding fault sample.

[0081] Example 2: like Figure 3 As shown, this embodiment further improves upon the design of Embodiment 1. The difference lies in the fact that, in the actual operation of Embodiment 1, it was found that when the heating network is under maintenance and isolated, bypass sections are put into operation, manual valves are operated on-site, or check valves are opened or closed due to pressure differential changes, the actual connectivity status of the corresponding pipe sections may not be synchronously written into the control command sequence, execution feedback sequence, and network topology data. This results in the target operating condition path, the actual operating condition path, and the response arrival timescale being generated according to a topology structure inconsistent with the actual hydraulic connectivity, failing to distinguish between residual propagation anomalies caused by equipment failures and residual propagation anomalies caused by changes in the actual network connectivity status. Therefore, the AI-based heating system fault early warning and self-diagnosis method also includes an operational topology correction process.

[0082] The topology correction process starts with a control event and extracts the observation response timestamps of each monitoring node from the operational monitoring sequence. Specifically, the normal fluctuation range of the corresponding monitoring node and the corresponding data type is determined based on the baseline segment before the control event. Starting from the command timestamp of the control event, the monitoring values ​​are read sequentially according to the acquisition timestamp. When the current monitoring value and the next valid monitoring value are on the same side of the normal fluctuation range, and the direction of their change is consistent with the direction of the equipment state change corresponding to the control event, the acquisition timestamp of the current monitoring value is determined as the observation response timestamp of that monitoring node. Monitoring nodes that do not form the above continuous change are recorded as having no observed response.

[0083] The observation response timestamps of each monitoring node are arranged according to the direction of heat medium flow to form the observation response order; the actual predicted response order is formed according to the actual response arrival timestamps. The observation response order and the actual predicted response order are compared node by node. When a monitoring node recorded as unreachable in the actual operating path forms an observation response timestamp, or when an upstream monitoring node in the actual operating path does not observe a response while its downstream monitoring node forms an observation response timestamp, a topological contradiction record is generated.

[0084] The topology conflict record includes the control event, the last monitoring node with a consistent response order, the first monitoring node with a inconsistent response order, and the pipe segment between the two. The pipe segment between the two, as well as the bypass pipe segment and branch pipe segment directly connected to the monitoring node with the first inconsistent response order, are identified as suspected topology pipe segments.

[0085] For each suspected topology segment, it is set to either a connected state or a blocked state to form a candidate topology. Based on the candidate topology, the actual operating path and the actual response arrival time scale are regenerated, and the corresponding monitoring nodes are arranged according to the regenerated actual operating path. A candidate topology is determined as a valid candidate topology when it meets the following conditions: the monitoring nodes that form the observation response time scale are reachable in the regenerated actual operating path; the order of each observation response time scale is consistent with the order of the regenerated actual response arrival time scale; and the recalculated mass balance term and pressure-flow relationship term conform to the physical constraint benchmark relationship formed by the benchmark segment.

[0086] When only one candidate topology state meets the above conditions, the candidate topology state is determined as the current valid topology state, and a valid topology version is generated. When there are two or more valid candidate topology states or no valid candidate topology state exists, a topology verification record is generated, and the corresponding control event sample is retained in the pending sample set until the on-site inspection record or maintenance work order confirms the connectivity status of the suspected topology segment.

[0087] After generating a valid topology version, the pipeline topology data is updated with the current valid topology status. The target operating condition path, actual operating condition path, target response arrival time and actual response arrival time are regenerated for the corresponding control event. The target residual sequence, actual residual sequence, physical constraint sequence and fault evidence sequence are re-formed. The fault evidence sequence generated based on the pipeline topology data before the update is marked as a failure record and is not used for fault diagnosis model update.

[0088] The newly formed fault evidence sequence continues to determine the admission status according to the matching admission conditions, residual admission conditions and constraint admission conditions described in Example 1. Thus, when the actual connectivity status of the pipeline changes, the pipeline topology data used to generate the expected response is first corrected, and then the fault category and associated objects are determined, as well as the admission samples and model versions are registered.

[0089] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0090] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. An AI-based method for fault early warning and self-diagnosis of heating systems, characterized in that the method... include: Collect pipeline topology data, control command sequences, execution feedback sequences, and operation monitoring sequences; Control events and execution events are extracted from the control instruction sequence and execution feedback sequence respectively and paired to form an event pairing state; Based on the pipeline topology data, the target operating condition path is formed by the target state of the control event, and the actual operating condition path is formed by the actual state of the execution event. The target operating condition path and the actual operating condition path are input into the expected response model to generate the target expected response sequence and the actual expected response sequence. The response arrival time scale is aligned with the operation monitoring sequence to form the target residual sequence and the actual residual sequence. The operation monitoring sequence generates a physical constraint sequence. The event pairing status, the target residual sequence, the actual residual sequence and the physical constraint sequence are combined into a fault evidence sequence. The fault evidence sequence is input into the fault diagnosis model to generate a fault warning record. The fault warning record includes the fault category, associated objects and fault score. The samples from the start of the control event to the verification time point are placed into the pending sample set. The admission status is determined based on the fault evidence sequence. The fault diagnosis model is updated with the admission samples and the model version is registered. Establish the update source relationship between the admission sample and the model version; if the admission sample is verified as a faulty sample, determine and revoke the affected model version along the update source relationship, and restore the previous model version of the update source of the fault-free sample.

2. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 1, characterized in that, The control event includes a device identifier, instruction type, target state, and instruction timestamp. The execution event includes a device identifier, feedback type, actual state, and feedback timestamp. When the device identifiers of the control event and the execution event are consistent, the instruction type corresponds to the feedback type, and the instruction timestamp is earlier than the feedback timestamp, the event pairing status is marked as paired. When the pairing conditions are not met, the event pairing status is marked as unpaired, and the actual state of the corresponding device before the control event is extracted from the execution feedback sequence and written into the actual operating condition path.

3. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 2, characterized in that, Both the target operating condition path and the actual operating condition path include control object nodes, upstream monitoring nodes, downstream monitoring nodes, and pipe segments connecting adjacent nodes; the control object nodes, upstream monitoring nodes, and downstream monitoring nodes are arranged according to the direction of heat medium flow; the target state is written into the control object nodes of the target operating condition path, and the actual state is written into the control object nodes of the actual operating condition path; based on the pipe segment volume and the instantaneous flow rate in the operation monitoring sequence, a timescale corresponding to the temperature data in the target response arrival timescale is generated; Based on the response order of pressure data and instantaneous flow rate in the verified normal control events, a timescale corresponding to the pressure data and instantaneous flow rate is generated in the target response arrival timescale; When the actual state changes, an actual response arrival time stamp is generated along the actual operating condition path according to the same data type correspondence. When the actual state does not change, a no-response arrival time stamp is generated.

4. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 3, characterized in that, The expected response model includes a state encoding layer, a topology propagation layer, and a response decoding layer. The state encoding layer combines the water supply temperature, return water temperature, water supply pressure, return water pressure, and instantaneous flow rate before the control event with the target state to form a target state encoding, and combines the water supply temperature, return water temperature, water supply pressure, return water pressure, and instantaneous flow rate with the actual state to form an actual state encoding. The topology propagation layer propagates the target state code according to the target operating condition path and the target response arrival time stamp, and propagates the actual state code according to the actual operating condition path and the actual response arrival time stamp. Among them, the temperature state is propagated according to the time stamp corresponding to the temperature data, and the pressure state and flow state are propagated according to the time stamp corresponding to the pressure data and the instantaneous flow. The response decoding layer outputs the expected temperature value, expected pressure value, and expected flow value corresponding to the monitoring node, forming the target expected response sequence and the actual expected response sequence to determine the initial state of the node before the event occurs. The response decoding layer outputs the expected temperature, expected pressure, and expected flow rates corresponding to the monitoring nodes, forming the target expected response sequence and the actual expected response sequence.

5. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 4, characterized in that, The target residual sequence is formed by subtracting the running monitoring sequence from the target expected response sequence according to the target response arrival time markers corresponding to the data types. The actual residual sequence is formed by subtracting the running monitoring sequence from the actual expected response sequence according to the actual response arrival time markers corresponding to the data types. The command execution deviation component is formed by subtracting the target residual sequence from the actual residual sequence corresponding to the same monitoring node and the same acquisition time marker. The actual residual sequence is arranged according to the actual response arrival time markers to form the residual propagation component. The constraint deviation term is extracted from the physical constraint sequence to form the physical deviation component. The command execution deviation component, residual propagation component, and physical deviation component are written into the fault evidence sequence item according to the connection relationship between the controlled object node, upstream monitoring node, and downstream monitoring node to form the physical deviation component. The command execution deviation component, residual propagation component, and physical deviation component are written into the fault evidence sequence according to the connection relationship between the controlled object node, upstream monitoring node, and downstream monitoring node.

6. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 5, characterized in that, Using pipeline branch nodes as mass balance nodes, a mass balance term is formed based on the instantaneous flow rate entering and leaving the pipeline branch nodes. Using heat exchangers as heat power balance nodes, the instantaneous heat power on the primary side is formed based on the primary side supply water temperature, primary side return water temperature, primary side instantaneous flow rate, and the density and specific heat capacity of the primary side heat medium. The instantaneous heat power on the secondary side is formed based on the secondary side supply water temperature, secondary side return water temperature, secondary side instantaneous flow rate, and the density and specific heat capacity of the secondary side heat medium. A heat power balance term is formed based on the difference between the instantaneous heat power on the primary side and the instantaneous heat power on the secondary side. Using the pipe section where the regulating valve or the pipe section where the circulating pump is located as the pressure and flow constraint pipe section, a pressure and flow relationship term is formed based on the pressure difference across the regulating valve, the actual opening degree of the regulating valve, and the instantaneous flow rate, or based on the difference between the inlet and outlet pressures of the circulating pump, the actual frequency of the circulating pump, and the instantaneous flow rate. The mass balance terms, heat power balance terms, and pressure-flow relationship terms are arranged according to the response arrival time scale to form a physical constraint sequence.

7. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 6, characterized in that, Extract the baseline segment before the control event from the operational monitoring sequence, and extract the verification segment between the end of the actual expected response sequence and the verification time marker; input the baseline segment and the actual state before the control event into the expected response model to generate the baseline expected sequence; align the operational monitoring values ​​in the baseline segment with the baseline expected sequence and calculate the difference to form the baseline residual sequence; and form the residual baseline relationship and physical constraint baseline relationship based on the baseline residual sequence. Set the event pairing status to paired status as the pairing access condition, set the target residual sequence and the actual residual sequence to meet the residual benchmark relationship within the verification section as the residual access condition, and set the physical constraint sequence to meet the physical constraint benchmark relationship within the verification section as the constraint access condition. When the matching admission condition, residual admission condition, and constraint admission condition are all met, the corresponding sample is marked as an admission sample; otherwise, the corresponding sample is retained in the pending sample set.

8. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 7, characterized in that, The fault diagnosis model includes a working condition adaptation branch, a fault memory branch, and an evidence fusion layer; a fault memory sample set is formed using verified progressive fault samples, and the fault memory branch is trained based on the fault memory sample set. The operating condition adaptation branch is updated using admission samples. When updating the operating condition adaptation branch, the model parameters of the fault memory branch and the evidence fusion layer remain unchanged to form a candidate model. Verified normal samples and verified progressive fault samples with the same control event type are selected to form a verification sample pair. The verification sample pair is input into the candidate model. When the fault score output by the candidate model for the verified progressive fault sample is higher than the fault score output for the verified normal sample, the candidate model is registered as the model version. The candidate model outputs the normal category for the verified normal sample, and the associated object output for the verified progressive fault sample is consistent with the verification record.

9. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 8, characterized in that, The update source relationship includes sample identifier, training batch identifier, parent model version, and child model version. Based on the sample identifier of the faulty sample, the training batch identifier containing the faulty sample is queried, and the corresponding child model version and subsequent model versions with the child model version as the parent model version are identified as the affected model versions. Tracing back along the inheritance direction of the parent model version, the most recent model version that does not contain the faulty sample in the update source relationship is identified as the preceding model version. The affected model version is revoked, the faulty sample is transferred to the fault memory sample set, and the corresponding fault evidence sequence is reprocessed using the preceding model version.

10. The AI-based fault early warning and self-diagnosis method for heating systems according to claim 7, characterized in that, Also includes: Based on the observation response timescale and the actual response arrival timescale of each monitoring node, a topological contradiction record is formed to identify suspected topological pipe segments. Suspected topological pipe segments are assigned a connected state or a blocked state, the actual operating path is reconstructed and the physical constraint sequence is verified to determine the effective topological state. Based on this, the pipeline topology data is updated and a fault evidence sequence is regenerated.

Images