Integrated multi-sensor fusion heat exchange station edge computing control system

Abstract

The application discloses a heat exchange station edge computing control system integrated with multi-sensor fusion and belongs to the technical field of automatic control of central heating. In view of the problem that the existing heat exchange station control scheme is sensitive to sensor faults and relies on the cloud, leading to too large control time delay, an independent Kalman filter is established for each same sensor group in the edge computing node, and a health degree evaluation unit, an adaptive fusion unit, a model predictive control unit and an abnormal diagnosis and redundancy switching unit are arranged. The evaluation unit synthesizes a composite health degree based on the residual error, noise, communication and physical constraint four factors with the same caliber as the Kalman innovation. The fusion unit constructs the measurement noise covariance by using the reciprocal of the health degree. The control unit solves the constrained quadratic programming on the edge side in combination with the load prediction. The diagnosis unit triggers instantaneous isolation through the accumulation and detection of the innovation scalar component, and realizes control reference tracking alignment through valve position and frequency feedback.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for centralized heating, and in particular to an edge computing control system for heat exchange stations that integrates multi-sensor fusion. Background Technology

[0002] In a centralized heating system, the heat exchange station, as an energy conversion node between the primary and secondary pipe networks, plays a crucial role in converting and distributing the heat from the high-temperature heat source medium according to the actual needs of the users. Its operational quality directly affects the heating quality, heat energy utilization efficiency, and operational safety of the entire heating network.

[0003] In existing engineering practices, heat exchange stations typically employ a local control architecture combining a field-programmable logic controller (FPGA) with classic PID control, or a two-tier architecture with an FPGA and centralized cloud control. The former architecture generally uses single-point instrument readings as the control input. However, heat exchange stations commonly experience harsh conditions such as high temperatures, humidity, vibration, and scale buildup. Temperature, pressure, and flow sensors are prone to zero-point drift, sudden range changes, mechanical jamming, and even signal loss during long-term operation. If the controller uses the readings of a faulty sensor as the basis for adjustment, the regulating valves and variable frequency pumps will operate according to the erroneous data, causing the secondary network supply and return water temperatures to deviate from the setpoint, frequent valve oscillations, and even thermal imbalance accidents. While there have been attempts in engineering to use arithmetic averaging or median filtering for simple redundancy processing, these solutions only offer a post-hoc compromise between measured values ​​and cannot effectively identify the true reliability of a single sensor online. Essentially, they do not address the fundamental aspect of data reliability assessment.

[0004] While the latter architecture leverages cloud computing power to improve the overall decision-making level, the round-trip communication latency between the field and the cloud is typically hundreds of milliseconds or even seconds. Combined with the cloud's batch processing cycle, this results in a delayed response to second-level disturbances such as secondary valve switching, user window opening, and sudden changes in outdoor temperature. Once the communication link is interrupted, the control system often degenerates into a simple local adjustment mode, failing to maintain the required control quality. More importantly, these two problems are coupled in practical engineering—unreliable sensor readings sent to the cloud further contaminate the cloud-side model's training and decision-making foundation, while cloud latency makes it difficult for the local system to promptly correct abnormal data.

[0005] Current technologies lack an integrated solution that can simultaneously address online reliable assessment and real-time optimized control of multiple sensors at the heat exchange station site. Maintaining the continuity of state estimation and ensuring millisecond-level real-time response capability of the control loop under field conditions where sensors may drift or fail at any time has become a critical technical problem that needs to be solved to further improve the automation level of heat exchange stations. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing heat exchange station control schemes, which are sensitive to sensor failures and have excessive cloud control latency. It provides an edge computing control system for heat exchange stations that integrates multi-sensor fusion, which can perform online assessment of multi-sensor reliability, adaptive fusion, real-time optimized control and fault isolation switching in an integrated manner at the edge node, thereby improving heating quality and reducing energy consumption.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] An edge computing control system for a heat exchange station integrating multi-sensor fusion includes a data acquisition module, edge computing nodes, an actuator control module, a communication module, and a cloud collaboration module. The data acquisition module includes temperature sensor groups redundantly arranged in the primary side supply water pipe, primary side return water pipe, secondary side supply water pipe, and secondary side return water pipe; pressure sensor groups and instantaneous flow meters arranged on the primary and secondary sides; and an outdoor environment sensor group. The actuator control module includes a primary side electric regulating valve with valve position feedback and a secondary side circulating pump frequency converter with frequency feedback. The edge computing node independently establishes a discrete Kalman filter for each similar sensor group, and its state vector is related to the data measured by that sensor group. The physical quantities are consistent; the edge computing node includes a multi-source data preprocessing unit, a sensor health assessment unit, an adaptive Kalman fusion unit, a load prediction unit, a model prediction control unit, and an anomaly diagnosis and redundancy switching unit; the sensor health assessment unit is configured to calculate the residual factor for each sensor in the same sensor group, based on the residual between the value obtained by projecting the prior state estimate of the current period of the Kalman filter corresponding to the sensor through the corresponding observation row vector and the current measurement value of the sensor; calculate the noise statistics factor based on the deviation of the standard deviation of the first difference within the sliding window from the factory noise benchmark; calculate the communication integrity factor based on the proportion of valid messages within the sliding window; and calculate the communication integrity factor based on the thermodynamic prior relationship of the heat transfer process. The physical constraint compliance factor is calculated, and the four factors are synthesized according to a preset index to obtain the composite health of the sensor. The adaptive Kalman fusion unit is configured to construct the measurement noise covariance matrix of the Kalman filter online with the reciprocal of the composite health as the ratio, so that the variance term corresponding to the sensor with a low health is amplified, and the optimal state estimate and innovation vector for the current period are output accordingly. The model prediction control unit is configured to use the optimal state estimate as the initial value, the future multi-step load prediction sequence output by the load prediction unit as the feedforward perturbation, and based on the discrete state-space model of the heat exchanger lumped parameters, with the opening degree of the primary side regulating valve and the frequency of the secondary side circulating pump as decision variables, to solve the constrained quadratic rolling optimization. The problem is that only the first step instruction of the optimal control sequence is applied to the field actuator via the actuator control module; the anomaly diagnosis and redundancy switching unit is configured to perform bilateral accumulation and detection on each scalar component of the innovation vector. When it is determined that a sensor has a sudden failure, the composite health of the sensor is clamped to a preset floor value in the current sampling period. At the same time, a state covariance matrix reset instruction is sent to the adaptive Kalman fusion unit, and a control reference tracking alignment instruction is sent to the model prediction control unit. The control reference tracking alignment instruction causes the model prediction control unit to reset the previous step control quantity, which is used as the incremental reference, to the actual value corresponding to the valve position feedback and the frequency feedback.After determining that the fault has been cleared, the system sequentially enters the pre-recovery observation phase and the linear ramp-up phase. If any check factor exceeds its limit again during any phase, the system immediately reverts to the preset floor value.

[0009] As a preferred embodiment of the present invention, the temperature sensor groups on the primary water supply pipe, the primary water return pipe, the secondary water supply pipe, and the secondary water return pipe are all composed of three redundantly installed Pt1000 platinum resistance thermometers; the primary pressure sensor group and the secondary pressure sensor group are both composed of two 4~20mA pressure transmitters; the outdoor environment sensor group includes an outdoor temperature sensor, a solar radiation sensor, and a wind speed sensor; the data acquisition module also includes an indoor representative point temperature sensor group for monitoring the temperature at the user end, and the indoor representative point temperature sensor group uploads the measurement data to the edge computing node wirelessly.

[0010] As a preferred embodiment of the present invention, the sensor health assessment unit synthesizes the composite health score according to the following formula:

[0011]

[0012] in For the first The composite health of the sensors alone , , , These are the residual factor, noise statistics factor, communication integrity factor, and physical constraint compliance factor, respectively. , , , satisfy When the composite health score is lower than the preset health score threshold, the corresponding sensor is marked as suspicious.

[0013] As a preferred embodiment of the present invention, the residual factor is calculated using the following formula: ,in For the new information vector The scalar component This is the prior state estimate of the current period of the Kalman filter corresponding to the sensor. This is the observation row vector corresponding to the sensor; the noise statistics factor is calculated as follows: Calculation; the communication integrity factor is calculated according to Calculation; the physical constraint compliance factor is set to a preset low value when the sensor measurement violates the thermodynamic prior constraint, otherwise it is set to 1; where This is the current measurement value of the sensor. To allow for residual scaling, The standard deviation of the first difference within the sliding window. This is the factory noise standard for this sensor model. For noise scale, The proportion of valid messages within the sliding window. This is the penalty index.

[0014] As a preferred embodiment of the present invention, the adaptive Kalman fusion unit independently establishes state transition equations for each similar sensor group. With measurement equation , where the state vector The physical meaning is consistent with the physical quantity measured by the sensor group, and the measurement noise covariance matrix is ​​constructed online using the following formula:

[0015]

[0016] in For the first Only the factory noise variance of the sensor, To prevent the health floor value from being divided by zero; the adaptive Kalman fusion unit executes sequentially. , Prediction steps and information based New information covariance Gain The update steps, and the new information vector and its scalar components, The diagonal elements are synchronously output to the anomaly diagnosis and redundancy switching unit; the adaptive Kalman fusion unit, upon receiving the state covariance matrix reset command, outputs the covariance matrix... Reinitialize to the preset diagonal initial values.

[0017] As a preferred technical solution of the present invention, the load prediction unit takes the feature vector composed of the current outdoor temperature, the outdoor temperature of the previous hour, the outdoor temperature of the previous 24 hours, the time of day, the day of the week, the secondary network heat load of the previous hour, the secondary network heat load of the previous 24 hours, solar radiation and wind speed as input, and calls the gradient boosting decision tree model or lightweight recurrent neural network model trained in the cloud and then pruned and compressed and sent to the local machine to output the prediction sequence of the secondary network heat load in the future.

[0018] As a preferred embodiment of the present invention, the model prediction control unit constructs a rolling optimization objective function based on the following formula:

[0019]

[0020] in This refers to the secondary water supply temperature. The set temperature is given by the outdoor temperature compensation curve. For the increment of decision variables, For the prediction domain length, To control the length of the domain, , , The weight matrix and weighting coefficients are issued after being calibrated offline by the cloud. For the energy consumption estimation term of the circulating pump, based on the pump affine law, we get... At the rated operating point Nearby, expand it to the second order using Taylor expansion and take... The quadratic form, For the first Step pump frequency decision variables, coefficients , The objective function is obtained by fitting it in the neighborhood of the working point through offline regression; the objective function is solved by quadratic programming under the amplitude constraints of valve opening, circulating pump frequency and its rate of change; when the model prediction control unit receives the control reference tracking alignment command, it resets the previous step control quantity, which is used as the incremental reference, to the current actual value of the valve position feedback and the frequency feedback, and clears the hot start quantity of the previous cycle.

[0021] As a preferred embodiment of the present invention, the anomaly diagnosis and redundancy switching unit maintains a two-sided cumulative sum statistic for each sensor. and ,in The first of the new information vectors scalar component The new covariance matrix diagonal The square root of each element. For tolerance parameters; when or The system determines when a sudden fault has occurred in the corresponding sensor. This is a preset detection threshold.

[0022] As a preferred embodiment of the present invention, the anomaly diagnosis and redundancy switching unit is further configured as follows: for a sensor that is determined to have a sudden fault, its composite health value is directly clamped to the preset floor value within the current sampling period to achieve transient isolation; before determining that the sensor fault is resolved, a pre-recovery observation phase is entered, during which the residual factor, physical constraint compliance factor, and deviation of its measured value from the fused output of other sensors in the same group are continuously monitored within the pre-recovery observation window; only when the above indicators continuously meet the preset conditions is the linear ramp recovery phase entered, and its composite health value is linearly recovered from the preset floor value to the steady-state value naturally calculated by the four factors at a preset slope; if the above indicators exceed the limit again in either the pre-recovery observation phase or the linear ramp recovery phase, the composite health value of the sensor is immediately returned to the preset floor value and the observation time is restarted.

[0023] As a preferred embodiment of the present invention, the communication module includes a fieldbus communication unit and a cloud-based collaborative communication unit. The fieldbus communication unit uses Modbus RTU or Modbus TCP protocol to poll and collect slowly varying process quantities, and uses EtherCAT or Profinet real-time industrial Ethernet protocol to collect process quantities requiring high sampling rates or global clock synchronization. The cloud-based collaborative module includes a model retraining unit, a parameter distribution unit, and a historical data archiving unit. The cloud-based collaborative communication unit sends aggregated features of state estimation, health, control output, and energy consumption to the cloud-based collaborative module at a preset period. The model retraining unit performs offline retraining of the load prediction model and model predictive control weights and distributes them to the edge computing node via the parameter distribution unit. The edge computing node also includes a local database unit and a security and watchdog unit. The local database unit is used to save the original data and fused output when the link with the cloud is interrupted. The security and watchdog unit switches the control output to a backup PID regulation mode when the model predictive control unit fails to solve or the computing resources are overloaded.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This invention performs an online comprehensive evaluation of the residuals, noise statistics, communication integrity, and physical constraints of various similar sensors within an edge computing node. The inverse of the composite health factor is directly used in the construction of the Kalman filter measurement noise covariance matrix. This automatically assigns smaller weights to sensors experiencing drift, abrupt changes, or jamming during the fusion process, avoiding the contamination of control inputs by erroneous data from single-point measurements or equal-weighted averaging. Since the calculation method of the residual factor is completely consistent with the Kalman innovation, there is no conflict in dimensions or timing between health assessment and anomaly diagnosis. Under typical operating conditions, the standard deviation of the secondary network water supply temperature estimation can be reduced from approximately ±0.6℃ in single-point measurements to within ±0.08℃, resulting in a significant reduction in the frequency of regulating valve operation and controlling the fluctuation range of the secondary network supply and return water temperatures within ±0.5℃.

[0026] By offloading all control logic, such as load forecasting, model prediction rolling optimization, and fault diagnosis, to the edge nodes for execution, constrained quadratic programming is solved within a second-level cycle using valve opening and circulating pump frequency as decision variables. The control loop no longer relies on real-time communication with the cloud, and has the ability to proactively suppress secondary hydraulic disturbances and outdoor temperature step disturbances. The latency of the control loop is reduced from the original second level to the tens of millisecond level. The cloud and the edge only interact to aggregate features and model parameters, significantly reducing uplink bandwidth consumption. The edge side can independently maintain stable operation for a long time when the cloud link is interrupted.

[0027] The anomaly diagnosis and redundancy switching unit, based on the two-sided cumulative sum detection of Kalman innovation scalar components, instantly clamps the composite health status of the faulty sensor upon triggering and simultaneously resets the state covariance matrix. The model predictive control unit simultaneously aligns the incremental reference to the actual position of the field actuator based on valve position and frequency feedback, ensuring that the absolute control output at the moment of fault switching remains continuous with the actuator position before the fault, avoiding valve or pump frequency jumps caused by reference zeroing. After the fault is cleared, the health status recovery consists of a pre-recovery observation phase and a linear ramp recovery phase. If the index exceeds the limit in either phase, the clamping is retracted, preventing the system from repeatedly oscillating between isolation and recovery. Overall, the system maintains stable control quality under conditions of both sensor and communication anomalies, with reduced overall energy consumption compared to traditional PID schemes and a correspondingly extended mean time between failures (MTBF). Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall structure of an edge computing control system for a heat exchange station that integrates multi-sensor fusion according to the present invention.

[0029] Figure reference numerals: 100, Data acquisition module; 101, Primary side supply water temperature sensor group; 102, Primary side return water temperature sensor group; 103, Secondary side supply water temperature sensor group; 104, Secondary side return water temperature sensor group; 105, Primary side pressure sensor group; 106, Secondary side pressure sensor group; 107, Primary side instantaneous flow meter; 108, Secondary side instantaneous flow meter; 109, Outdoor environment sensor; 110, Indoor representative point temperature sensor group; 200, Edge computing node; 201, Multi-source data preprocessing unit; 202, Sensor health assessment unit; 203, Adaptive Kalman fusion unit; 204 205. Load Prediction Unit; 206. Model Prediction Control Unit; 207. Anomaly Diagnosis and Redundancy Switching Unit; 208. Local Database Unit; 209. Safety and Watchdog Unit; 300. Actuator Control Module; 301. Primary Side Electric Regulating Valve; 302. Secondary Side Circulating Pump Frequency Converter; 303. Secondary Side Makeup Pump Control Unit; 304. Electric Bypass Valve; 400. Communication Module; 401. Fieldbus Communication Unit; 402. Cloud Collaborative Communication Unit; 500. Cloud Collaborative Module; 501. Model Retraining Unit; 502. Parameter Distribution Unit; 503. Historical Data Archiving Unit; 504. Operation Visualization Unit. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can more clearly understand the technical solution of the present invention. However, the scope of protection of the present invention is not limited to the following embodiments.

[0031] like Figure 1 As shown in the figure, this embodiment discloses an edge computing control system for a heat exchange station integrating multi-sensor fusion. At the hardware level, it includes a data acquisition module 100 deployed on-site, an edge computing node 200 serving as the local computing and control core, an actuator control module 300 connected to the primary and secondary pipeline actuators, a communication module 400 connecting the on-site system and the cloud, and a cloud collaboration module 500 deployed on a remote server. The data acquisition module 100 sends multi-redundant sensor data to the edge computing node 200 via a fieldbus. After completing data fusion and control decisions, the edge computing node 200 drives the primary-side regulating valve and the secondary-side circulating pump via the actuator control module 300. The communication module 400 is responsible for low-frequency parameter synchronization between the edge node and the cloud collaboration module 500. In this embodiment, the edge computing node 200 can be an industrial-grade edge gateway with a neural network acceleration unit or a floating-point coprocessor to ensure that data fusion, model inference, and quadratic programming solutions can be completed within a second-level control cycle.

[0032] The fieldbus between the data acquisition module 100 and the edge computing node 200 can be configured in different tiers according to the requirements for synchronization and sampling rate. For slowly changing process quantities such as temperature and pressure, Modbus RTU carried by RS485 or Modbus TCP carried by TCP / IP can meet the requirements. In this case, the equivalent sampling rate of a single node in a single polling cycle is usually between 1Hz and 10Hz. For fast-changing process quantities or multi-channel global synchronous acquisition scenarios that require higher sampling rates and stricter clock synchronization, real-time industrial Ethernet such as EtherCAT and Profinet IRT can be used, with a global equivalent sampling rate of over 100Hz. The two buses can coexist in parallel on the same edge gateway, respectively handling the acquisition of different process quantities and the transmission of downlink commands.

[0033] The data acquisition module 100 includes a primary-side water supply temperature sensor group 101 for acquiring the primary-side water supply temperature, which consists of three Pt1000 platinum resistance thermometers redundantly installed at the inlet of the primary-side water supply pipe; a primary-side return water temperature sensor group 102 with the same structure installed on the primary-side return water pipe; a secondary-side water supply temperature sensor group 103 installed on the secondary-side water supply main pipe; and a secondary-side return water temperature sensor group 104 installed on the secondary-side return water main pipe, both of which also consist of three Pt1000 platinum resistance thermometers redundantly arranged to ensure that the redundancy of similar measurements can be maintained even if any one temperature sensor fails. The primary-side pressure sensor group 105 and the secondary-side pressure sensor group 106 each consist of two 4~20mA pressure transmitters, connected to the primary and secondary side main pipes respectively. The primary-side instantaneous flow meter 107 and the secondary-side instantaneous flow meter 108 can be electromagnetic or ultrasonic instruments, and their measurement results are used in energy balance and hydraulic constraint calculations. The outdoor environment sensor 109 integrates outdoor temperature, solar radiation, and wind speed measurement functions, serving as exogenous feature inputs for the load prediction model. The indoor representative point temperature sensor group 110 consists of two to six wireless temperature sensors, typically uploading indoor temperatures representing several user terminals to the edge computing node 200 via a LoRa wireless link, used to correct the outdoor temperature compensation curve or as an evaluation basis for comfort indicators.

[0034] Edge computing node 200 is the core carrier of the software logic of this invention. Internally, it is functionally divided into a multi-source data preprocessing unit 201, a sensor health assessment unit 202, an adaptive Kalman fusion unit 203, a load prediction unit 204, a model prediction control unit 205, an anomaly diagnosis and redundancy switching unit 206, a local database unit 207, and a security and watchdog unit 208. The multi-source data preprocessing unit 201 performs timestamp alignment, unit normalization, upper and lower limit pruning, and single-point glitch removal based on the 5σ criterion in each sampling period, and outputs the aligned measurement vector and sliding window buffer to subsequent units. The sensor health assessment unit 202, the adaptive Kalman fusion unit 203, the model prediction control unit 205, and the anomaly diagnosis and redundancy switching unit 206 form a strict dependency relationship and feedback loop through four sets of variables: state estimation, health, innovation sequence, and covariance matrix, constituting the core algorithm chain of this invention. The local database unit 207 adopts a ring buffer structure to store the raw data and fused output of the most recent 72 hours for use in network outage recovery and local analysis. The safety and watchdog unit 208 has an embedded hardware-level watchdog and implements amplitude and rate-of-change limiting on the control output. When numerical problems occur in the upper-level optimization solution or the edge CPU is overloaded, the control loop is automatically downgraded to a pre-calibrated safe PID mode to ensure that the system will not enter an uncontrolled state under any operating conditions.

[0035] The actuator control module 300 includes a primary-side electric regulating valve 301, a secondary-side circulating pump frequency converter 302, a secondary-side makeup water pump control unit 303, and an electric bypass valve 304. The primary-side electric regulating valve 301 has valve position feedback, and the secondary-side circulating pump frequency converter 302 has frequency feedback. These feedback signals are transmitted in real-time to the edge computing node 200 via the fieldbus, participating in the state estimation and optimization solution of the next cycle, and serving as the actual anchor point for the incremental reference of the model prediction control unit 205 during fault realignment. The communication module 400 includes a fieldbus communication unit 401 and a cloud collaborative communication unit 402. The former uses the aforementioned Modbus RTU / TCP or EtherCAT protocol to interact with the data acquisition module 100 and the actuator control module 300, while the latter uses the MQTT protocol via a TLS encrypted channel to interact with the cloud collaborative module 500. The cloud collaboration module 500 includes a model retraining unit 501, a parameter distribution unit 502, a historical data archiving unit 503, and a runtime visualization unit 504. The model retraining unit 501 uses aggregated data uploaded by edge nodes to update the load prediction model and model prediction control weights offline on a daily or weekly scale. The parameter distribution unit 502 distributes the updated model parameters and MPC weights to the edge side through a secure channel, thus forming a two-layer collaborative architecture of "edge autonomy and cloud evolution".

[0036] The following section further explains the specific implementation of the key units inside the edge computing node 200.

[0037] The adaptive Kalman fusion unit 203 independently establishes a discrete Kalman filter for each group of similar sensors. Specifically, primary-side supply water temperature, primary-side return water temperature, secondary-side supply water temperature, secondary-side return water temperature, primary-side pressure, secondary-side pressure, primary-side flow rate, and secondary-side flow rate each correspond to a Kalman filter. The state vector of each filter contains only the low-dimensional state components corresponding to the measured physical quantity (typically including the physical quantity and its first-order rate of change). Each filter operates in parallel on the same time grid and their fused outputs are aggregated to the outer layer of this unit. For each filter, the state transition equation adopts... , Measurement equations are for within groups Only sensors are unified as , ,in all sectors The state vector is projected onto the measurement space of each sensor. Since each sensor in the same group measures the same physical quantity, Degenerate into The scalar component of the corresponding physical quantity is shown in the figure. The key difference between this invention and conventional Kalman fusion lies in... Online matrix construction methods:

[0038]

[0039] in For the first Only the factory noise variance of the sensor, To prevent division by zero for the health floor value (typically 0.05), the variance corresponding to sensors with lower health is significantly amplified, and the filter gain is automatically mathematically assigned a smaller weight to them. The filtering process is performed in the conventional two-step prediction and update process, with the prediction step obtaining the prior estimate. with prior covariance The update step is in calculating new information. New information covariance With gain Based on this, the optimal state estimate for the current period is obtained. With covariance New information vector Its scalar components as well as diagonal elements The data is synchronously output to the sensor health assessment unit 202 and the anomaly diagnosis and redundancy switching unit 206. When this unit receives a state covariance matrix reset command from the anomaly diagnosis and redundancy switching unit 206, it will... The matrix is ​​reinitialized to a preset diagonal initial value to clear erroneous covariance coupling terms accumulated by the faulty sensor before and after the fault occurred.

[0040] For the sensor health assessment unit 202, it assesses the health of each sensor in the same sensor group. In each sampling period (Typically 1 second) Calculate four normalization factors. The residual factor is taken from the innovation vector output by the adaptive Kalman fusion unit 203. scalar component ,according to Calculation, where The allowable residual scale for this type of sensor is obtained through calibration during the field initialization phase. It is important to emphasize that the prior estimate for the current period is used here. Instead of the previous period's posterior estimate Therefore, the scalar on which the residual factor is based is exactly the same component of the Kalman innovation vector. The residual caliber used for health assessment and anomaly diagnosis is completely consistent at the variable level, avoiding system evolution bias introduced by using estimators from different time points. When the residual is smaller... The closer the measurement is to 1, the more significantly it will deviate from the fused output. It decreases exponentially, thus assigning extremely low weight to measurements of obvious anomalies at this factor level. The noise statistical factor is based on width. The standard deviation of the first difference is measured within a sliding window (typically 60). Compared with the factory noise standard of this sensor model The comparison, according to Calculations show that when the sensor experiences sudden changes or severe vibrations, the first-order difference standard deviation increases significantly. The corresponding decrease occurs. The communication integrity factor is based on the proportion of valid messages within the sliding window. Based on, according to Calculation, where This is the penalty index, typically set to 2. Physical constraint compliance factor. Based on the thermodynamic prior relationships of the heat exchange process, for example, under normal operating conditions, the primary side supply water temperature should be greater than the primary side return water temperature, and the secondary side supply water temperature should be greater than the secondary side return water temperature but less than the primary side supply water temperature; when the measured value of a certain sensor violates the above constraints... Set to the preset low value, otherwise take 1. The four factors are... The synthesis yields a composite health score, in which... In this embodiment, a typical example is taken , The resulting health level Between 0 and 1. When Below the health threshold When this happens, the sensor health assessment unit 202 marks the sensor as suspicious and reports it to the anomaly diagnosis and redundancy switching unit 206.

[0041] For load forecasting unit 204, in each optimization cycle, it receives outdoor temperature, current time period, weekday attribute, secondary network heat load of the previous hour and the previous 24 hours, solar radiation, and wind speed from the current and several historical times, and organizes them into a feature vector. Input into the local inference engine, output the future Step (typical) The secondary network heat load prediction sequence (corresponding to a 10-minute prediction domain with one step per minute) The prediction model used can be a gradient boosting decision tree model trained in the cloud and then pruned and quantized for compression, with no more than 200 nodes; or a long short-term memory neural network model with a small hidden dimension (typically no more than 16). Both can perform millisecond-level inference on mainstream ARM Cortex-A53-level edge processors. As equivalent alternatives, those skilled in the art can also use random forests, support vector regression, or heuristic methods based on historical template matching to replace the above models, as long as they can output load prediction sequences for multiple future steps, the function of this unit can be achieved.

[0042] For the model predictive control unit 205, rolling optimization is performed based on the lumped parameter thermodynamic model of the heat exchanger. The secondary side energy balance can be expressed as follows: Heat exchanger heat exchange ,in The logarithmic mean temperature difference can be approximated as linear near the operating point after linearization; the valve opening and pump frequency must meet the following requirements. as well as The amplitude and rate of change are constrained. The above equations are then used to control the period. After zero-order preserved discretization, the state-space form is obtained. ,in The load disturbance term is given by the load forecasting unit 204. Based on this, a rolling optimization objective function is constructed:

[0043]

[0044] in Based on the outdoor temperature compensation curve, this embodiment typically uses the prediction domain. control domain Control cycle Seconds. The above objective function combines comfort biases, control action costs, and energy consumption costs of different dimensions into a single cost functional in the form of a weighted sum of squares, with the weight matrix... , Energy trade-off coefficient It is responsible for the unit conversion between "℃²" and "%² / Hz²" and "kW·h", respectively. The values ​​are calibrated offline by the cloud-based collaborative module 500 based on historical operating data. The calibration is performed using a multi-objective weighted average, with comfort deviation under typical annual operating conditions, valve / pump frequency of operation, and energy consumption as trade-off indicators. A grid search or Bayesian optimization method based on cross-validation is used to complete the calibration, which supports distribution according to season, day / night, or operating mode. (Circulating pump energy consumption estimation item) The specific form is based on the affine law of water pumps. Considering that the cubic relation will make the objective function non-quadratic and violate the QP solution framework, the cubic relation is set at the rated operating point. Perform a Taylor expansion to second order in the vicinity, and take...

[0045]

[0046] The approximate form of the quadratic form, where For the prediction domain The pump frequency decision variable of the step, coefficient , Offline regression within the neighborhood of the work point The power-frequency data was fitted to the measured power-frequency dataset, and the constant term was omitted because it had no effect on the optimization problem. After the above simplification, the entire optimization problem is transformed into a standard quadratic programming problem, with the decision variable only including the valve opening. With pump frequency In the control domain The increment on the top, the total dimension is The solution can be obtained in milliseconds on mainstream edge processors using readily available solvers such as qpOASES or a custom-implemented active set method. The resulting optimal control sequence is applied only in the first step according to the rolling optimization principle; its valve opening and pump frequency commands are sent to the primary-side electric regulating valve 301 and the secondary-side circulating pump frequency converter 302 via the fieldbus communication unit 401. For applications with extremely limited computing resources, explicit model predictive control can also be used to store the piecewise affine control law obtained offline in the lookup table structure of the edge nodes to further compress the online solution time.

[0047] For reference processing in fault conditions, when this unit receives the control reference tracking and alignment command issued by the anomaly diagnosis and redundancy switching unit 206, it will not use the previous control quantity as the incremental reference. Simple zeroing—If this is done, and the valve position actually remains at a non-zero opening (e.g., 50%), the absolute control quantity applied in the new cycle... This will result in a jump of tens of percentage points, causing a sharp change in valve or pump frequency. Therefore, this unit utilizes the valve position feedback of the primary-side electric regulating valve 301 and the frequency feedback of the secondary-side circulating pump frequency converter 302 to... Force alignment to the actual execution position corresponding to such feedback signals, i.e., make Valve position feedback measured value Frequency feedback is used based on measured values; simultaneously, the warm-up values ​​from the previous cycle are cleared to avoid introducing solution biases with outdated QP intermediate variables. Based on this, MPC uses the aligned values... As the starting point for incremental calculation, with the current The QP is reconstructed based on the initial value, and the new absolute instruction obtained is continuous with the actual position of the actuator at the moment of failure, thereby achieving true non-disruptive realignment.

[0048] For the anomaly diagnosis and redundancy switching unit 206, it continuously consumes the innovation vector and diagonal elements of the innovation covariance output by the adaptive Kalman fusion unit 203, and maintains a two-sided cumulative sum statistic for each sensor. and ,in The first of the new information vectors scalar component The diagonal of the new information covariance matrix The square root of each element, tolerance parameter Typically, it's taken as 0.5. When or Time (typically taking the detection threshold) This indicates that the sensor has experienced a sudden malfunction.

[0049] For abrupt faults, this embodiment employs a processing strategy combining transient isolation and two-stage progressive recovery. For abrupt faults identified by the cumulative sum detector or physical consistency check, the measured value typically deviates significantly from the true physical quantity (e.g., a sudden jump from near the operating point of 60°C to the range boundary or remaining stuck). Continuing to allow it to enter fusion with any weight greater than the floor value would cause significant disturbances to the control loop. Therefore, the anomaly diagnosis and redundancy switching unit 206 will check the composite health of the sensor within the current sampling period when the fault is confirmed. Clamp to Thus The variance of the corresponding item in the matrix is ​​instantly amplified to a large order of magnitude determined by dividing the factory variance by the floor value, meaning the sensor essentially no longer contributes weight in subsequent fusion calculations. Simultaneously, this unit sends a state covariance matrix reset command to the adaptive Kalman fusion unit 203, causing... The matrix is ​​reinitialized to a preset diagonal initial value; a control reference tracking alignment command is sent to the model prediction control unit 205, causing it to... Alignment is made to the actual values ​​of valve position feedback and frequency feedback. Thanks to the continuity at the state level between the redundant sensors in the same group and the Kalman filter, the state estimation of the fused output will not show obvious steps due to the instantaneous removal of a certain sensor. This is also an important improvement of the present invention compared to the simple "single-point instrument + alarm shutdown" scheme.

[0050] The recovery of faulty sensors employs a two-stage, gradual process to prevent sensors that are not yet truly stable from skewing the fusion output again after weight increases, thus preventing repeated oscillations between isolation and recovery. The first stage is the pre-recovery observation stage: from the initial shielding time after the fault is triggered. (Typically 5 minutes) This phase begins, and its duration is [duration missing]. (Typically 5 to 10 minutes), within this window, this unit continuously monitors three consistency indicators—the sensor's residual factor. Is it continuously higher than the preset threshold? Physical constraint compliance factor Whether it consistently equals 1, and whether the deviation between its original measurement and the fusion results of other health sensors in the same group consistently falls below the tolerance. The second stage only begins when all three indicators remain within acceptable limits throughout the observation window. The second stage is the linear slope recovery stage: using a preset slope (typically...). , (Taking 60 to 300 seconds) will determine the composite health score. A linear recovery to the steady-state value calculated naturally from the four factors. During the slope operation, continue to monitor the aforementioned three consistency indicators. If any indicator exceeds the limit again, immediately... Back to The timing of the pre-recovery observation phase will be restarted. If necessary, once the cumulative number of triggers reaches the threshold, the sensor will be marked as a permanent fault and reported to the cloud for manual intervention.

[0051] In addition to cumulative sum-based mutation detection, this unit also performs parallel physical consistency checks and long-term mean drift detection based on thermodynamic prior relationships to improve the ability to identify chronic drift; chronic drift can be detected by the residual factor before reaching the cumulative sum threshold. With physical constraint factor The continued decline is reflected in the composite health score, thus naturally weakening it at the integration level.

[0052] This system operates in each control cycle. The system completes a full closed loop according to a fixed data and signal flow. The data acquisition module 100 sends the raw data from the field sensor groups 101 to 110 to the multi-source data preprocessing unit 201 at a sampling rate matching the selected fieldbus. The preprocessing unit 201 interpolates and aligns the data from each channel to the same grid and then outputs an aligned measurement vector. With a sliding buffer; the adaptive Kalman fusion unit 203 performs a prediction step for each group of sensors of the same type to obtain a prior estimate. With the new vector Sensor health assessment unit 202 The four factors and composite health score of each sensor are calculated as input. Adaptive Kalman fusion unit 203 structure The matrix is ​​then updated, and the optimal state estimate for the current period is output. ; Anomaly diagnosis and redundancy switching unit 206 receives and It also updates the cumulative and statistical data, and sends a notification to the health assessment unit 202 once a fault determination is triggered. Instantaneous clamping command, issuing a covariance matrix reset command to the adaptive Kalman fusion unit 203, and issuing a control baseline tracking alignment command to the model prediction control unit 205; the load prediction unit 204 uses the latest feature vector to call the local inference engine to provide future multi-step load predictions; the model prediction control unit 205 uses... For initial value, The feedforward disturbance is given by the outdoor temperature compensation curve. For reference, after alignment Using an incremental baseline, a quadratic programming problem is constructed and solved. The first step of the resulting optimal control sequence is sent via fieldbus communication unit 401 to the primary-side electric regulating valve 301 and the secondary-side circulating pump frequency converter 302. The valve position and frequency feedback returned by the actuator are incorporated into the state vector of the next cycle and used as a reference during fault realignment. The anchoring benchmark. At the same time, the cloud collaborative communication unit 402 in the communication module 400 sends the aggregated fusion status, health, control output and energy consumption indicators to the cloud collaborative module 500 at a preset low frequency period (typically once per minute). The model retraining unit 501 completes the retraining of the load prediction model and MPC weights on a daily or weekly scale. The parameter distribution unit 502 distributes the updated parameters to the edge computing node 200 through the MQTT secure channel to complete the cloud-edge collaborative iteration.

[0053] When the cloud-based collaborative link is interrupted, the edge computing node 200 continues to run independently using the latest locally cached model. The local database unit 207 retains the original data and fused output, and resumes the transmission of any data that was not uploaded once the link is restored. When the model prediction control unit 205 fails to solve the problem due to numerical reasons or the edge processor becomes overloaded due to sudden tasks, the safety and watchdog unit 208 automatically switches the control output to the backup safety PID mode and applies hard limiting to the output amplitude and rate of change, thereby ensuring that the system will not enter a state of out of control under various abnormal operating conditions.

[0054] The parameter values ​​involved in the above embodiments (e.g., sampling period) Sliding window width ,index / / / Healthy flooring Detection threshold Initial blocking duration Pre-recovery observation period Slope duration Prediction domain Control domain and energy consumption fitting coefficient , The values ​​(etc.) are typical values ​​given for ease of explanation, and those skilled in the art can make corresponding adjustments based on the capacity, thermal inertia, and sensor configuration of the on-site heat exchange station. The sensor type is not limited to the Pt1000 platinum resistance thermometer and electromagnetic flowmeter listed in this embodiment; it can also be replaced with thermocouples, thermistors, vortex flowmeters, or ultrasonic flowmeters, etc., with equivalent functions. Similarly, the hardware form of the edge computing node 200 is not limited to an industrial edge gateway; it can also be implemented as an industrial server with an accelerator, a programmable logic controller add-on module, or an embedded industrial control computer. The above variations and equivalent substitutions are all natural extensions of the present invention's concept, and any modifications and improvements that do not depart from the core technical solution of the present invention should fall within the protection scope of the present invention.

Claims

1. An edge computing control system for a heat exchange station integrating multi-sensor fusion, characterized in that, The system includes a data acquisition module, edge computing nodes, an actuator control module, a communication module, and a cloud collaboration module. The data acquisition module includes temperature sensor groups redundantly arranged in the primary water supply pipeline, primary water return pipeline, secondary water supply pipeline, and secondary water return pipeline; pressure sensor groups and instantaneous flow meters arranged on the primary and secondary sides; and an outdoor environment sensor group. The actuator control module includes a primary-side electric regulating valve with valve position feedback and a secondary-side circulating pump frequency converter with frequency feedback. The edge computing node independently establishes a discrete Kalman filter for each similar sensor group, and the state vector of each discrete Kalman filter is consistent with the physical quantity measured by its corresponding sensor group. The node includes a multi-source data preprocessing unit, a sensor health assessment unit, an adaptive Kalman fusion unit, a load prediction unit, a model prediction control unit, and an anomaly diagnosis and redundancy switching unit. The sensor health assessment unit is configured to, for each sensor in a similar sensor group, calculate a residual factor based on the residual between the value obtained by projecting the prior state estimate of the Kalman filter corresponding to the sensor's current period onto the corresponding observation row vector and the sensor's current measurement value; calculate a noise statistics factor based on the deviation of the standard deviation of the first-order difference within the sliding window from the factory noise baseline; calculate a communication integrity factor based on the proportion of valid messages within the sliding window; and calculate a physical constraint compliance factor based on the thermodynamic prior relationship of the heat transfer process. The above four factors are synthesized according to a preset index to obtain the composite health of the sensor; the adaptive Kalman fusion unit is configured to construct the measurement noise covariance matrix of the Kalman filter online with the reciprocal of the composite health as the ratio, so that the variance term corresponding to the sensor with a low health is amplified, and outputs the optimal state estimate, innovation vector and diagonal elements of the innovation covariance matrix for the current period accordingly; the model prediction control unit is configured to use the optimal state estimate as the initial value, the future multi-step load prediction sequence output by the load prediction unit as the feedforward perturbation, and based on the discrete state-space model of the heat exchanger lumped parameters, with the opening degree of the primary side regulating valve and the frequency of the secondary side circulating pump as decision variables, solve the constrained quadratic rolling... The optimization problem is addressed by applying only the first step instruction of the optimal control sequence to the field actuator via the actuator control module. The anomaly diagnosis and redundancy switching unit is configured to perform bilateral accumulation and detection on each scalar component of the innovation vector. When a sudden fault is detected in a sensor, the composite health of the sensor is clamped to a preset floor value within the current sampling period. Simultaneously, a state covariance matrix reset instruction is sent to the adaptive Kalman fusion unit, and a control benchmark tracking alignment instruction is sent to the model prediction control unit. The control benchmark tracking alignment instruction causes the model prediction control unit to reset the previous step control quantity, which serves as the incremental benchmark, to the actual value corresponding to the valve position feedback and the frequency feedback.After determining that the fault has been cleared, the system sequentially enters the pre-recovery observation phase and the linear ramp-up phase. If the verification factor exceeds the limit again during any phase, the composite health of the corresponding sensor is immediately reverted to the preset floor value.

2. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The temperature sensor groups on the primary water supply pipe, primary water return pipe, secondary water supply pipe, and secondary water return pipe are all composed of three redundantly installed Pt1000 platinum resistance thermometers; the primary pressure sensor group and the secondary pressure sensor group are both composed of two 4~20mA pressure transmitters; the outdoor environment sensor group includes an outdoor temperature sensor, a solar radiation sensor, and a wind speed sensor; the data acquisition module also includes an indoor representative point temperature sensor group for monitoring the temperature at the user end, and the indoor representative point temperature sensor group uploads the measurement data to the edge computing node wirelessly.

3. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The sensor health assessment unit synthesizes the composite health score according to the following formula: ; in For the first The composite health of the sensors alone , , , These are respectively the residual factor, the noise statistics factor, the communication integrity factor, and the physical constraint compliance factor, and the index. , , , satisfy ; When the composite health score is lower than a preset health score threshold, the corresponding sensor is marked as suspicious.

4. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 3, characterized in that, The residual factor is calculated according to Calculation, where For the new information vector The scalar component This is the prior state estimate of the current period of the Kalman filter corresponding to the sensor. This is the observation row vector corresponding to the sensor; the noise statistics factor is calculated as follows: Calculation; the communication integrity factor is calculated according to Calculation; the physical constraint compliance factor is set to a preset low value when the sensor measurement violates the thermodynamic prior constraint, otherwise it is set to 1; where This is the current measurement value of the sensor. To allow for residual scaling, The standard deviation of the first difference within the sliding window. This is the factory noise standard for this sensor model. For noise scale, The proportion of valid messages within the sliding window. This is the penalty index.

5. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The adaptive Kalman fusion unit independently establishes state transition equations for each similar sensor group. With measurement equation The state vector of each discrete Kalman filter contains only the physical quantity measured by the corresponding sensor group and its first-order rate of change; the adaptive Kalman fusion unit constructs the measurement noise covariance matrix online according to the following formula: ; in For the first Only the factory noise variance of the sensor, To prevent the health floor value from being divided by zero; the adaptive Kalman fusion unit executes sequentially. , Prediction steps and information based New information covariance Gain The update steps, and synchronously output the innovation vector and its scalar components, and the diagonal elements of the innovation covariance matrix to the anomaly diagnosis and redundancy switching unit; when the adaptive Kalman fusion unit receives the state covariance matrix reset command, it outputs the covariance matrix... Reinitialize to the preset diagonal initial values.

6. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The load prediction unit takes the feature vector composed of the current outdoor temperature, the outdoor temperature of the previous hour, the outdoor temperature of the previous 24 hours, the time of day, the day of the week, the secondary network heat load of the previous hour, the secondary network heat load of the previous 24 hours, solar radiation and wind speed as input, calls the gradient boosting decision tree model or lightweight recurrent neural network model trained in the cloud and then downloaded to the local machine after pruning and compression, and outputs the prediction sequence of the secondary network heat load in the future.

7. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The model prediction and control unit constructs a rolling optimization objective function based on the following formula: ; in This refers to the secondary water supply temperature. The set temperature is given by the outdoor temperature compensation curve. For the increment of decision variables, For the prediction domain length, To control the length of the domain, , and The weight matrix and trade-off coefficients are issued offline from the cloud based on historical operating data; the circulating pump energy consumption estimation term Pick The quadratic form, For the first Pump frequency decision variables, coefficients , Based on the affine law of water pumps At the rated operating point The result was obtained by nearby second-order Taylor expansion and offline regression fitting; The objective function is solved using a quadratic programming approach under the amplitude constraints of valve opening, circulating pump frequency, and their rate of change. When the model prediction control unit receives the control reference tracking alignment command, it resets the valve opening component in the previous step control quantity, which serves as the incremental reference, to the current actual value of the valve position feedback, resets the pump frequency component to the current actual value of the frequency feedback, and clears the hot start quantity of the previous cycle.

8. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The anomaly diagnosis and redundancy switching unit maintains a bilateral cumulative sum statistic for each sensor. and ,in For the first of the new information vectors scalar component The diagonal of the new information covariance matrix is ​​the first... The square root of each element. For tolerance parameters; when or The system determines when a sudden fault has occurred in the corresponding sensor. This is a preset detection threshold.

9. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The anomaly diagnosis and redundancy switching unit is further configured as follows: for a sensor determined to have experienced a sudden failure, the composite health of the sensor is directly clamped to the preset floor value within the current sampling period; during the pre-recovery observation phase, the residual factor, physical constraint compliance factor, and deviation of the sensor's original measurement value from the fusion results of other healthy sensors in the same group are continuously monitored within a preset observation window; only when the three indicators continuously meet the preset conditions within the observation window does it enter the linear ramp recovery phase, and the composite health of the sensor recovers from the preset floor value to the steady-state value naturally calculated by the four factors at a preset slope; if any of the above three indicators exceeds the limit again in either the pre-recovery observation phase or the linear ramp recovery phase, the composite health of the sensor is reverted to the preset floor value and the observation time is restarted.

10. The edge computing control system for a heat exchange station integrating multi-sensor fusion according to claim 1, characterized in that, The communication module includes a fieldbus communication unit and a cloud-based collaborative communication unit. The fieldbus communication unit uses Modbus RTU or Modbus TCP protocols to poll and collect slowly changing process quantities, and uses EtherCAT or Profinet real-time industrial Ethernet protocols to collect process quantities requiring high sampling rates or global clock synchronization. The cloud-based collaborative module includes a model retraining unit, a parameter distribution unit, and a historical data archiving unit. The cloud-based collaborative communication unit sends aggregated features of state estimation, health, control output, and energy consumption to the cloud-based collaborative module at preset intervals. The model retraining unit performs offline retraining of the load prediction model and model predictive control weights and distributes them to the edge computing node via the parameter distribution unit. The edge computing node also includes a local database unit and a security and watchdog unit. The local database unit is used to save the original data and fused output when the link with the cloud is interrupted. The security and watchdog unit switches the control output to a backup PID regulation mode when the model predictive control unit fails to solve or the computing resources are overloaded.

Images