Chiller fault diagnosis method based on reliable online data self-adaptation

Abstract

A kind of water chiller fault diagnosis method based on reliable online data self-adaption, collects real-time operation data of unit, carries out null value prediction and outlier correction using orthogonal state iteration method, and imports complete unit data into electrical-thermal fault decoupling model to carry out sensor state discrimination, eliminate fault sensor data and fault-free data prediction;Finally, the thermal fault detection model is used to distinguish the thermal state of the unit for fault-free prediction data, and the fault prediction, diagnosis result and auxiliary maintenance decision are obtained.The present application realizes the null value prediction and outlier correction of water chiller data, decoupling diagnosis when electrical fault and thermal fault coexist such as sensor measurement deviation, accurate prediction of water chiller thermal fault, reliable online data selection, online self-adaptive update of fault diagnosis model, and establishes maintenance decision database, greatly improves the intelligent level of water chiller, and can be quickly and efficiently generalized to other types of water chiller systems.

Description

Technical Field

[0001] This invention relates to a technology in the field of building air conditioning, specifically a fault diagnosis method for chiller units based on orthogonal state iteration method, electrical-thermal fault decoupling, and reliable online data adaptive strategy. Background Technology

[0002] Heating, ventilation, and air conditioning (HVAC) systems account for over 50% of a building's total energy consumption. Within HVAC systems, chillers account for the largest share of energy consumption, and their operational reliability directly impacts building energy efficiency. Existing chiller fault diagnosis methods generally target single faults, primarily alerting to hard faults affecting basic operation. Currently, there is a lack of methods for early prediction and rapid diagnosis of multiple coexisting soft faults, particularly a method that can decouple electrical-thermal faults and quickly generalize to multi-fault decoupling diagnosis across different chiller models. Furthermore, existing models lack the ability to self-update based on real-time data. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, such as inadequate handling of prediction biases caused by abnormal data acquisition in chiller units, significant decrease in prediction accuracy when operating conditions differ greatly from training conditions, leading to discrepancies between the final diagnosis and actual results; inability to quickly decouple and diagnose multiple coexisting soft faults, resulting in a significant drop in prediction accuracy when multiple faults are present; and lack of self-updating capabilities based on real-time data, preventing real-time correction of prediction results and causing a significant decrease in prediction accuracy when chiller unit operating conditions differ from training conditions. This invention proposes a chiller unit fault diagnosis method based on reliable online data adaptive adaptation. This method achieves chiller unit data null value prediction and outlier correction, decoupled diagnosis of electrical and thermal faults (such as sensor measurement deviations), accurate prediction of chiller unit thermal faults, reliable online data selection, online adaptive updating of the fault diagnosis model, and the establishment of a maintenance decision database. This significantly improves the intelligence level of chiller units and can be quickly and efficiently generalized and applied to other types of chiller unit systems.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a fault diagnosis method for chiller units based on reliable online data adaptation. The method involves collecting real-time operating data of the unit, using an orthogonal state iteration method for null value prediction and outlier correction, and inputting complete unit data into an electrical-thermal fault decoupling model for sensor status identification, removal of faulty sensor data, and prediction of fault-free data. Finally, a thermal fault detection model is used to determine the unit's thermal state from the fault-free prediction data, yielding fault prediction, diagnosis results, and auxiliary maintenance decisions.

[0006] The real-time operating data of the unit is obtained by reading the sensor data and operating parameters of the chiller unit and recording the data window of the chiller unit in real time. Specifically, it includes: operating status, alarm status, compressor suction temperature, discharge temperature, suction pressure, condensing pressure, compressor current, refrigerant water inlet and outlet temperature, cooling water inlet and outlet temperature, condenser and evaporator inlet and outlet pressure difference, electronic expansion valve and water flow regulating valve positions, start-up, shutdown and load increase / decrease temperature, superheat and subcooling data.

[0007] The orthogonal state iteration method refers to: setting the historical data collection period to α minutes, collecting real-time operating data of the unit, projecting the real-time operating data onto the Legendre orthogonal polynomial basis, iteratively updating the orthogonal basis coefficients according to the state-space expression, and finally reconstructing the unit operating data through the orthogonal basis coefficients to obtain the operating data curve approximated by the Legendre orthogonal polynomial at the current operating time, and using the predicted data values ​​to fill in the null values ​​in the data collection and the outliers removed by Hampel filtering.

[0008] The projection mentioned refers to obtaining the coefficients of the chiller unit operating data under the Legendre polynomial basis according to the definition of orthogonal projection, specifically including:

[0009] 1) Based on the historical data collection period α minutes, the measure function of the projected space of the chiller unit operating data is obtained as follows: Based on Legendre orthogonal polynomials that transform the running time from the range [-1, 1] to the range [t-α, t] Where: P n Let x be a Legendre orthogonal polynomial of order n, x be a certain operating data of the current chiller unit, such as suction temperature, discharge temperature, etc., and t be the operating time in minutes; polynomial p n (t, x) satisfies <p n (t) p m (t) >=δ n，m t is the running time, n and m are the nth and mth order polynomials respectively, and δ n，m To satisfy the condition n = m, δ n，m =1, otherwise δ n，m =0, and the symbols <·, ·> represent the inner product of polynomials in a given metric space.

[0010] 2) Calculate the coefficients of the chiller unit operating data projected onto the Legendre orthogonal polynomial basis. in: Let x be the coefficients of the nth Legendre orthogonal polynomial of the chiller unit operating data vector x at time t. Wherein: T CI T is the compressor suction temperature. COP is the compressor discharge temperature. RE P is the evaporation pressure. RC The condensation pressure is represented by the symbols <·, ·>. ω(t) Let f be the inner product of polynomials in a given metric space under the measure function ω(t). ≤t It is a function derived from the actual operating data of the chiller unit at times less than or equal to t.

[0011] The state-space expression refers to the update expression of the coefficients obtained by taking the partial derivative with respect to time t based on the calculation formula of the Legendre orthogonal polynomial basis coefficients: Wherein, the expression for matrix C is: The expression for matrix D is: n is the order of the Legendre orthogonal polynomial, and j is the dimension of the chiller unit's operating data.

[0012] The iterative update of the state-space expression refers to: discretizing the base coefficients of the Legendre orthogonal polynomial using the bilinear method. The state-space expression is used to calculate the state update at time t: Where: I is the identity matrix.

[0013] The Legendre orthogonal polynomial basis approximation refers to: using the orthogonal polynomial basis coefficients obtained by projection, and reconstructing the chiller unit operating data within the data collection time period α minutes using the Legendre orthogonal polynomial f(t), specifically: in: Let x be the basis coefficient vector of the Legendre orthogonal polynomial, x be the current operating data vector of the chiller unit, and n be the order of the Legendre orthogonal polynomial.

[0014] The Hampel filtering mentioned above refers to setting an outlier filtering window for the real-time operating data of the chiller system. For the data in each window, the median and mean absolute deviation are calculated. If the deviation is greater than three times the standard deviation of the data window, it is removed.

[0015] In the described orthogonal state iteration method, if the collected chiller unit data contains null or outlier values, the state space expression update step of the Legendre orthogonal polynomial basis coefficients at the current moment is ignored. When the operating data returns to normal, the update is performed again. After the Legendre orthogonal polynomial basis coefficients are updated, the latest coefficients are used to predict the chiller unit operating data within the time period α minutes, covering the missing data within the time period α minutes directly collected from the chiller unit. The covered complete data is then input into the electrical-thermal fault decoupling model.

[0016] The electrical-thermal fault decoupling model includes a sensor deviation fault detection model and a sensor fault-free data prediction model. The sensor deviation fault detection model separates sensor faults from real-time unit operation data, while the sensor fault-free data prediction model eliminates interference from sensor faults, completing sensor-thermal fault decoupling and obtaining fault-free prediction data. In the offline phase, historical data is used to train both the sensor deviation fault detection model and the sensor fault-free data prediction model. In the online phase, the most recent 20 minutes of real-time data from the chiller unit are collected, and data within the model training parameters are input into the sensor deviation fault detection model for detection, yielding predicted sensor faults. If the fault detection model outputs no fault, it is directly input into the thermal fault detection model for thermal fault identification. If the fault detection model outputs a fault, the sensor data predicted as faulty is input into the sensor fault-free data prediction model for prediction, obtaining the sensor fault-free data prediction value. The fault-free prediction value replaces the most recent 20 minutes of real-time chiller unit data input into the thermal fault detection model to eliminate interference from sensor faults in thermal fault diagnosis, thus enabling thermal fault diagnosis.

[0017] The sensor deviation fault detection model is a metric learning model based on adaptive density discrimination, which is trained in the following way:

[0018] S221: Encode the input sensor data and transform it into a representation vector through nonlinear transformation;

[0019] S222: For each sensor fault category, pre-define K cluster centers, use K-means clustering to minimize the distance between the representation vector and the cluster centers, and update the cluster center vector;

[0020] S223: Randomly sample a cluster and obtain the M-1 nearest clusters around it, and randomly sample D sensor data vectors from each cluster;

[0021] S224: Calculate the loss function for the current iteration.

[0022] Where: α is the separation interval between clusters. The average value of the sampled m-th cluster representation vector. Let m be the cluster number, and d be the representation vector of the sensor after being encoded by an artificial neural network. C(·) is the category label of the current representation vector.

[0023] Let be the variance of the representation vector of the m-th cluster. The positive subscript indicates that negative values ​​are discarded and only positive values ​​are taken.

[0024] S225: After a certain number of iterations, pause the training process, update the cluster center vector, and then resume gradient descent training.

[0025] The trained sensor deviation fault detection model calculates the category during the prediction phase.

[0026] Where: r n For the nth sensor data sample x n The representation vector μ obtained after network transformation l Let σ be the mean vector of the L closest clusters, and σ be the mean vector of the clusters during training. The average value of C(·) is the category label to which the current representation vector belongs.

[0027] The data within the model training condition parameter range refers to the following: when the unit is operating under model training conditions, it is determined whether the sensor values ​​meet a pre-set threshold condition. If the threshold is exceeded, a sensor fault is diagnosed. If the threshold is not exceeded or the unit is not operating under model training conditions, the data is input into the sensor deviation fault detection model for diagnosis. If a fault is diagnosed, a sensor fault is output, and the data is input into the sensor fault-free data prediction model for prediction.

[0028] The sensor fault-free data prediction model is an artificial neural network based on a reliable online data adaptive strategy. During the training phase, the network minimizes the mean square error between the predicted data and the real sensor data through gradient descent. During the prediction phase, it predicts the sensor value when there is no fault based on the faulty sensor data after encoding.

[0029] The aforementioned reliable online data adaptive strategy refers to: firstly, based on the data output after decoupling electrical-thermal faults, eliminating online data whose predicted entropy of the artificial neural network model exceeds a set threshold; secondly, given a parameter perturbation range, minimizing the model's maximum predicted entropy within that range. Following this training method, an artificial neural network model based on the reliable online data adaptive strategy is trained. This model is used to judge the thermal state of the unit operation, and the judgment results are stored in a diagnostic result database. This strategy only intervenes in the model's prediction phase, obtaining the information entropy of the output results based on the fault probability distribution of the online output of the trained model. The higher the information entropy, the more likely the model is to give incorrect classification results. Therefore, in the online update phase, the model performs gradient descent updates based on entropy loss to reduce the model's entropy during the prediction phase. When applying the online data sharpness perception and adaptation strategy, the model only updates the lower-level normalization layer, specifically updating only the statistical normalization parameters and affine parameters, to reduce the impact on the stability of the model's prediction results. The statistical normalization parameters are estimated based on real-time data, while the affine parameters γ and β are calculated using gradient descent based on the entropy loss of the real-time data.

[0030] Furthermore, to mitigate the prediction accuracy degradation caused by class imbalance, batch size issues, and batch data distribution problems in the online data adaptation process, samples with entropy loss less than a set threshold must be selected during the gradient descent update process using entropy loss. Selecting samples with small entropy loss reduces the number of samples with large gradients, making the update process more stable and reducing the occurrence of model collapse. Simultaneously, the entropy loss optimization adopts a minimax optimization approach. Given the perturbation magnitude of the model parameters, a minimax strategy is used, that is, first maximizing the entropy of the output result under the perturbation, and then minimizing the output entropy. This training method encourages the model to enter the flat region of the entropy loss surface, resulting in better generalization ability.

[0031] The sensor fault-free data prediction model, based on a reliable online data adaptive strategy, trains by adjusting only the normalized layer in the artificial neural network. Specifically, the input parameters are: intake temperature, exhaust temperature, intake pressure, condensing pressure, compressor current, refrigerant inlet and outlet temperatures, cooling water inlet and outlet temperatures, condenser and evaporator inlet and outlet pressure differences, load increase / decrease temperatures, superheat, and subcooling parameters. The normalized parameter update in the Lth layer is specifically: μ←E[f L (T CI T CO P RE P RC …;Θ)], Wherein: T CI T is the compressor suction temperature. CO P is the compressor discharge temperature. RE P is the evaporation pressure.RC σ is the condensation pressure, μ is the mean vector in the normalized layer, and σ is the mean vector in the normalized layer. 2 f is the variance vector in the normalized layer. L (·) is the encoding function of the first L layers of the network, and E(·;Θ) is the prediction entropy of the artificial neural network model when the network parameter is Θ.

[0032] The entropy of the output of the trained artificial neural network: The network parameters are optimized by minimizing the loss L(x): L(T) CI T CO P RE P RC …)=minS(T CI T CO P RE P RC …)E R (T CI T CO P RE P RC …;Θ), Where: index function T CI T is the compressor suction temperature. CO P is the compressor discharge temperature. RE P is the evaporation pressure. RC Let γ be the condensing pressure, ∈ be the perturbation of the neural network parameters, ρ be the upper limit of the perturbation, f(·) be the encoding function of the network, E(x; Θ) be the predicted entropy of the artificial neural network model when the network parameter is Θ and the input chiller unit has no faults (x); E0 is the set upper limit of the entropy threshold. The γ and β parameters are updated as follows: Where: γ is the scaling parameter of the normalization layer, β is the offset parameter of the normalization layer, and L is the network loss.

[0033] The aforementioned thermal fault diagnosis model is an artificial neural network based on a reliable online data adaptive strategy. This network obtains the fault category probability through a softmax transformation after nonlinear transformation based on the compressor inlet and outlet pressures, evaporation pressure, and condensation pressure. Based on the pre-stored unit thermal state discrimination results and the current unit state, it recommends maintenance suggestions and assists in decision-making.

[0034] The diagnostic results include: healthy, sub-healthy, and faulty. Among them, when the system experiences refrigerant leakage, condenser blockage, evaporator blockage, or sensor fixed deviation, the system indicates a sub-healthy or warning state. When the system experiences electronic expansion valve jamming, high or low pressure in the system exhaust, high exhaust temperature, compressor oil leakage, low pressure difference between refrigerant water and cooling water, or sensor disconnection, the system indicates an alarm state.

[0035] The aforementioned auxiliary maintenance decision-making includes: preventive auxiliary maintenance decision-making and fault maintenance auxiliary decision-making. Specifically, the preventive auxiliary maintenance decision-making function is set to time, reminding the user to perform preventive maintenance at fixed maintenance time nodes. The fault maintenance auxiliary decision-making generates a maintenance suggestion decision tree based on the corresponding historical fault information of the chiller unit, which is linked to the corresponding maintenance information management database, thereby providing maintenance auxiliary decisions for fault information and reminding the user to perform corresponding maintenance operations.

[0036] This invention relates to a system for implementing the above-mentioned method, comprising: a chiller unit operation data acquisition and storage unit, a fault diagnosis unit, and a host computer communication unit, wherein: the chiller unit operation data acquisition unit acquires data collected by the chiller unit PLC and stores it in a MySQL or Influx database; the fault diagnosis unit obtains real-time operation data of the chiller unit for the past 20 minutes from the database, performs outlier and null value replacement through an orthogonal state iteration method, decouples electrical and thermal faults, and then uses an artificial neural network based on a reliable online data adaptive strategy to determine thermal faults, thereby obtaining thermal fault prediction results; the host computer communication unit encapsulates the fault prediction results and auxiliary decision-making content and sends them to the host computer.

[0037] Technical effect

[0038] This invention presents a method for predicting null values ​​and correcting outliers in chiller unit sensors based on orthogonal state iteration, and an artificial neural network model based on a reliable online data adaptive strategy for predicting fault-free sensor data and diagnosing system thermal faults in chiller units. Compared to existing technologies, this invention, by using the method for predicting null values ​​and correcting outliers in chiller unit sensors based on orthogonal state iteration, enables the fault diagnosis system to obtain more accurate sensor data predictions. Compared to directly using data interpolation or outliers, the diagnostic accuracy is improved by 3.3% and 7.7%, respectively. Using the artificial neural network model based on a reliable online data adaptive strategy for predicting fault-free sensor data improves accuracy by 9.8% compared to directly using faulty sensor data. Furthermore, using the artificial neural network model based on a reliable online data adaptive strategy for thermal fault identification improves accuracy by 4.3% compared to directly using traditional artificial neural network models, even when there is a significant difference between online data conditions and training data. Attached Figure Description

[0039] Figure 1 This is a flowchart of the present invention;

[0040] Figure 2 The flowchart is a method for sensor null value prediction and outlier correction of chiller units based on orthogonal state iteration.

[0041] Figure 3 This is a structural diagram of the sensor deviation fault detection model;

[0042] Figure 4 This is a structural diagram of the sensor fault-free data prediction model.

[0043] Figure 5 This is a structural diagram of a thermal fault detection model;

[0044] Figure 6 Diagram of the fault simulation experiment system;

[0045] Figure 7 A graph showing the predicted inlet water temperature curve for condenser #2 using the orthogonal state iteration method;

[0046] Figure 8 A schematic diagram for diagnosing the fixed deviation results of the inlet water temperature sensor of the No. 1 condenser in the electrical-thermal decoupling module;

[0047] Figure 9 This is a curve showing the predicted value of the outlet water temperature sensor for the No. 1 condenser of the electrical-thermal decoupling module. Detailed Implementation

[0048] like Figure 1 As shown in the figure, this embodiment relates to an intelligent control and health management method for HVAC and its chiller units, which includes the following steps:

[0049] S1: The system reads chiller unit sensor data and operating parameters in real time via RS485 communication, and records the chiller unit's data window for the past 20 minutes. It uses an orthogonal state iteration method for null value prediction and outlier correction, and stores the corrected data in the diagnostic database for use in the diagnostic process. The implementation process of this step is described in detail below:

[0050] The data transmitted via RS485 communication mainly includes the chiller system operating status, alarm status, compressor suction temperature, discharge temperature, suction pressure, condensing pressure, compressor current, refrigerant inlet and outlet temperatures, cooling water inlet and outlet temperatures, condenser and evaporator inlet and outlet pressure differences, electronic expansion valve and water flow regulating valve positions, start-up, shutdown, and load increase / decrease temperatures, superheat, and subcooling. The values ​​read from RS485 are mapped and converted. After each read is completed, the mapped result is stored in the database. The master station initiates the next RS485 query at fixed intervals. If multiple chiller units are configured, data from multiple chiller units is read simultaneously during a single conversion process.

[0051] In the orthogonal state iteration method, if the collected chiller unit data contains null or outlier values, the update step of the state-space expression for the Legendre orthogonal polynomial basis coefficients at the current moment is ignored. The update is performed only after the operating data returns to normal. After the Legendre orthogonal polynomial basis coefficients are updated, the latest coefficients are used to predict the chiller unit operating data for the time period α minutes, overwriting the missing data within the time period α minutes directly collected from the chiller unit, and then outputting the complete overwritten data. The update formula for the Legendre orthogonal polynomial basis coefficients in orthogonal state iteration is: in: n is the order of the Legendre orthogonal polynomial, and j is the dimension of the chiller unit operating data. The expression for the chiller unit operating data reconstruction process is as follows: in Let x be the basic coefficient vector of the Legendre orthogonal polynomial, t be the current operating data vector of the chiller unit, n be the operating time, and n be the order of the Legendre orthogonal polynomial.

[0052] Outliers are removed using the Hampelle filtering method. This method primarily uses the chiller unit's operating data from the past minute as the window size for filtering outliers. For each data point within the window, the number of digits and the mean absolute deviation are calculated. If the deviation exceeds three times the standard deviation of that data point, it is replaced with a null value. The window is then moved to the next data point, and the process is repeated.

[0053] S2: Based on the real-time unit operation data stored in the database, sensor faults are separated using a sensor deviation fault detection model. Interference from sensor faults is eliminated using a sensor fault-free data prediction model, thus decoupling sensor-thermal faults and obtaining fault-free prediction data. Specifically, this includes:

[0054] S21: Read historical data from the database, perform steady-state discrimination on the historical data, and then extract steady-state data to train the sensor deviation fault detection model;

[0055] S22: Train a sensor deviation fault detection model and a sensor fault-free data prediction model based on historical data;

[0056] S23: Read real-time data from the database to obtain the real-time data of the chiller unit for the most recent 20 minutes;

[0057] S24: Input the data within the model training condition parameter range into the sensor deviation fault detection model for detection, and obtain the predicted sensor fault.

[0058] S25: Input the sensor data predicted as faulty into the sensor fault-free data prediction model for detection, and obtain the sensor fault-free data prediction value;

[0059] S26: After detecting a sensor fault, use the fault-free predicted value to replace the sensor value and input it into the thermal fault detection model to eliminate the interference caused by the sensor fault and diagnose the thermal fault.

[0060] In step S22, the sensor deviation fault detection model adopts a metric learning model based on adaptive density discriminant, and its training steps include:

[0061] S221: Use an artificial neural network to encode the input sensor data, so that the input sensor data is transformed into a representation vector through a nonlinear transformation;

[0062] S222: For each sensor fault category, pre-define K cluster centers, use K-means clustering to minimize the distance between the representation vector and the cluster centers, and update the cluster center vector;

[0063] S223: Randomly sample a cluster and obtain the M-1 nearest clusters around it, and randomly sample D sensor data vectors from each cluster;

[0064] S224: Calculate the loss function for the samples in the current iteration according to the following formula, where: α is the separation interval between clusters. The average value of the sampled m-th cluster representation vector. Let m be the cluster number, and d be the representation vector of the sensor after being encoded by an artificial neural network. C(·) is the category label of the current representation vector. Let be the variance of the representation vector of the m-th cluster. The positive subscript indicates that negative values ​​are discarded and only positive values ​​are taken.

[0065] S225: After a certain number of iterations, pause the training process, update the cluster center vector, and then resume gradient descent training.

[0066] The sensor deviation fault detection model selected its category during the prediction phase according to the following formula, where: r n For the nth sensor data sample x n The representation vector μ obtained after network transformation l Let σ be the mean vector of the L closest clusters, and σ be the mean vector of the clusters during training. The average value, C(·) is the class label of the current representation vector, and the final output class is...

[0067] The data within the model training condition parameter range refers to: first, determining whether the unit is operating within the model training condition; that is, if the unit is operating within the model training condition, determining whether the sensor values ​​meet a pre-set threshold condition. If the threshold is exceeded, a sensor fault is diagnosed; if the threshold is not exceeded or the unit is not operating within the model training condition, the data is input into the sensor deviation fault detection model for diagnosis. If a fault is diagnosed, a sensor fault is output and input into the sensor fault-free data prediction model for prediction.

[0068] The sensor fault-free data prediction model employs an artificial neural network model based on a reliable online data adaptive strategy. During the training phase, the mean square error of the prediction is minimized through gradient descent. In the prediction phase, given faulty sensor data, the artificial neural network encodes the predicted sensor values ​​when fault-free, and this prediction is then passed to the thermal fault detection model.

[0069] S3: Based on the fault-free prediction data, use the thermal fault diagnosis model to judge the thermal state of the unit operation and store the judgment results in the diagnosis result database.

[0070] The aforementioned thermal fault diagnosis model employs an artificial neural network based on a reliable online data adaptive strategy. This network transforms input features such as compressor inlet and outlet pressures, evaporation and condensation pressures through the artificial neural network model, and finally obtains the fault category probability through a softmax transformation.

[0071] The aforementioned thermal fault diagnosis model updates its parameters based on real-time data, specifically through gradient descent updates using entropy loss, reducing the model's entropy during the prediction phase. When applying the online data sharpness perception and adaptation strategy, the model only updates the lower-level normalized feature modulation layer, such as... Figure 5 As shown, it specifically includes:

[0072] ① The model input parameters are suction temperature, discharge temperature, suction pressure, condensing pressure, compressor current, refrigerant water inlet and outlet temperatures, cooling water inlet and outlet temperatures, condenser and evaporator inlet and outlet pressure differences, load increase / decrease temperatures, superheat, and subcooling parameters. The normalized parameter update in layer L is obtained from the output of layer L and the following formula, where: T CI T is the compressor suction temperature. CO P is the compressor discharge temperature. RE P is the evaporation pressure. RC For condensation pressure, the update method for the normalized parameters in the normalized layer is specifically: μ←E[f L (T CI T CO P RE P RC …;Θ)],

[0073] ② Based on the trained network, obtain the entropy of the network output: Optimize the network parameters by minimizing the loss L(x): L(T) CI T CO P RE P RC …)=minS(T CI T CO P RE P RC …)E SA (T CI T CO P RE P RC …;Θ), Where: index function S(T) CI T CO P RE P RC …) can be obtained from the following formula: The γ and β parameters are updated as follows:

[0074] During the gradient descent update process using entropy loss, samples with entropy loss less than a set threshold are selected. Given the perturbation magnitude of the model parameters, a minimax strategy is used, that is, first maximizing the entropy of the output result under the perturbation, and then minimizing the output entropy.

[0075] The diagnostic results stored in the diagnostic results database are mainly divided into three states: healthy, sub-healthy, and faulty. Specifically, when the system experiences refrigerant leakage, condenser blockage, evaporator blockage, or sensor fixed deviation, the system indicates a sub-healthy or warning state. When the system experiences electronic expansion valve failure, high or low pressure in the system exhaust, high exhaust temperature, compressor oil leakage, low pressure difference between refrigerant water and cooling water, or sensor disconnection, the system indicates an alarm state.

[0076] A fault simulation experiment was conducted on a water-cooled chiller unit through specific practical experiments. The system consists of two screw compressors, two shell-and-tube condensers, one plate heat exchanger, and an electronic expansion valve. The unit contains two refrigeration cycle loops, with the evaporator shared by both loops. The refrigerant is R22, with a nominal charge of 100 kg, a rated cooling capacity of 400 kW, and a rated power of 107 kW. The system structure diagram is shown below. Figure 6 As shown in the table below, the fault simulation items include electrical faults such as fixed deviations of condenser inlet and outlet water temperature and pressure sensors, and fixed deviations of evaporator inlet and outlet water temperature and pressure sensors. Electrical-thermal coupling faults include all electrical faults and pairwise combinations of refrigerant leakage, reduced condenser water flow, and reduced evaporator water flow.

[0077]

[0078] The orthogonal state iteration method was used for null value prediction and outlier correction of chiller unit operating data. Compared with no prediction correction, the average accuracy was improved by 7.7%; compared with interpolation using values ​​near the current time, the average accuracy was improved by 3.3%; and compared with the traditional artificial neural network model, the average accuracy was improved by 4.1%. The orthogonal state iteration method was used to predict the inlet water temperature curve of condenser #2 in the experimental system, as shown below. Figure 7 As shown.

[0079]

[0080] Using an electrical-thermal fault decoupling model, the average accuracy is improved by 9.8% compared to methods without sensor fault isolation and fault-free data prediction. The diagnosis of the fixed deviation results of the No. 1 condenser inlet water temperature sensor in the electrical-thermal decoupling module is as follows: Figure 8 As shown, the predicted value of the fault-free data from the outlet water temperature sensor of condenser #1 is as follows: Figure 9 As shown.

[0081]

[0082]

[0083] The thermal fault detection model employs an artificial neural network based on a reliable online data adaptive strategy. Compared to using support vector machines, it achieves an average accuracy improvement of 6.8%; compared to using random forests, it achieves an average accuracy improvement of 8.5%; and compared to using traditional artificial neural network models, it achieves an average accuracy improvement of 4.3%.

[0084]

[0085] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for fault diagnosis of chiller units based on reliable online data adaptive analysis, characterized in that, Real-time operating data of the unit is collected, and the orthogonal state iteration method is used to perform null value prediction and outlier correction. The complete unit data is then input into the electrical-thermal fault decoupling model to identify sensor status, remove faulty sensor data, and predict fault-free data. Finally, the thermal fault detection model is used to identify the thermal status of the unit based on the fault-free prediction data, and to obtain fault prediction, diagnosis results, and auxiliary maintenance decisions. The real-time operating data of the unit is obtained by reading the sensor data and operating parameters of the chiller unit and recording the data window of the chiller unit in real time. Specifically, it includes: operating status, alarm status, compressor suction temperature, discharge temperature, suction pressure, condensing pressure, compressor current, refrigerant water inlet and outlet temperature, cooling water inlet and outlet temperature, condenser and evaporator inlet and outlet pressure difference, electronic expansion valve and water flow regulating valve position, start-up, shutdown and load increase / decrease temperature, superheat and subcooling data. The electrical-thermal fault decoupling model includes: a sensor deviation fault detection model and a sensor fault-free data prediction model. The sensor deviation fault detection model separates sensor faults from the real-time operating data of the unit, and the sensor fault-free data prediction model eliminates the interference of sensor faults, completes the decoupling of sensor-thermal faults, and obtains fault-free prediction data. The aforementioned thermal fault diagnosis model is an artificial neural network based on a reliable online data adaptive strategy. This network obtains the fault category probability through a softmax transformation after nonlinear transformation based on the compressor inlet and outlet pressures, evaporation pressure, and condensation pressure. Based on the pre-stored unit thermal state discrimination results and the current unit state, it recommends maintenance suggestions and assists in decision-making. The orthogonal state iteration method refers to: setting the historical data collection period length as... Every minute, real-time operating data of the unit is collected, the real-time operating data is projected onto the Legendre orthogonal polynomial basis, and the orthogonal basis coefficients are iteratively updated according to the state-space expression. Finally, the unit operating data is reconstructed through the orthogonal basis coefficients to obtain the operating data curve approximated by the Legendre orthogonal polynomial at the current operating moment. The predicted data values ​​are used to fill in the null values ​​in the data collection and the outliers removed by Hampel filtering. The projection mentioned refers to obtaining the coefficients of the chiller unit operating data under the Legendre polynomial basis according to the definition of orthogonal projection, specifically including: 1) Based on the historical data collection period The measure function for the projected space of the chiller unit's operating data is obtained after minutes. According to the running time from Convert within range Legendre orthogonal polynomials ,in: Let n be a Legendre orthogonal polynomial. This refers to a specific operating data point of the current chiller unit. The running time is in minutes; polynomial satisfy , Let n be the running time, and m be the nth-order polynomial and m-order polynomial, respectively. To meet the conditions hour Other cases ,symbol The inner product of polynomials in a given metric space; 2) Calculate the coefficients of the chiller unit operating data projected onto the Legendre orthogonal polynomial basis. ,in: For at any time Data vector related to chiller unit operation The coefficients of the nth Legendre orthogonal polynomial. = ,in: This refers to the compressor suction temperature. This refers to the compressor discharge temperature. For evaporation pressure, The condensation pressure is represented by the symbol. To find the measure function within a given metric space. The inner product of the following polynomials For chiller units at less than or equal to A function based on the actual runtime data at any given time; The state-space expression refers to: based on the formula for calculating the basis coefficients of the Legendre orthogonal polynomial, for time... Taking the partial derivative yields the updated expression for the coefficients: ,in: The expression for a matrix is: , The expression for a matrix is: , For the order of Legendre orthogonal polynomials, This refers to the data dimensions for chiller unit operation; The iterative update of the state-space expression refers to: discretizing the base coefficients of the Legendre orthogonal polynomial using the bilinear method. The state-space expression is calculated. Real-time status updates: ,in: It is the identity matrix; The Legendre orthogonal polynomial basis approximation refers to: based on the orthogonal polynomial basis coefficients obtained by projection, using the Legendre orthogonal polynomial basis... Reconstructing the data collection period The chiller unit's operating data within the specified minutes is as follows: ,in: Let the vector be the basis coefficients of the Legendre orthogonal polynomial. This is the current operating data vector of the chiller unit. For the order of Legendre orthogonal polynomials; The Hampel filtering mentioned above refers to: setting an outlier filtering window for the real-time operating data of the chiller system; calculating the median and mean absolute deviation for the data in each window; and removing data when the deviation is greater than three times the standard deviation of that data window. In the aforementioned orthogonal state iteration method, if the collected chiller unit data contains null or outlier values, the state-space expression update step for the Legendre orthogonal polynomial basis coefficients at the current moment is ignored. The update is performed only after the operating data returns to normal. After the Legendre orthogonal polynomial basis coefficients are updated, the latest coefficients are used to predict the time period. Chiller unit operating data within minutes, covering the time period directly collected from the chiller units. Missing data within a minute will be replaced by complete data that has been overwritten and input into the electrical-thermal fault decoupling model.

2. The chiller unit fault diagnosis method based on reliable online data adaptive analysis according to claim 1, characterized in that, The electrical-thermal fault decoupling model uses historical data to train the sensor deviation fault detection model and the sensor fault-free data prediction model in the offline stage. In the online stage, it collects the real-time data of the chiller unit in the last 20 minutes and inputs the data within the model training condition parameter range into the sensor deviation fault detection model for detection to obtain the predicted sensor fault. If the fault detection model outputs no fault, it is directly input into the thermal fault detection model for thermal fault identification; if the fault detection model outputs a fault, the sensor data predicted to be faulty is input into the sensor fault-free data prediction model for prediction, and the sensor fault-free data prediction value is obtained. The fault-free predicted value is used to replace the real-time data of the chiller unit in the most recent 20 minutes to input into the thermal fault detection model in order to eliminate the interference of sensor failure on thermal fault diagnosis and to diagnose thermal faults.

3. The chiller unit fault diagnosis method based on reliable online data adaptive method according to claim 1 or 2, characterized in that, The sensor deviation fault detection model is a metric learning model based on adaptive density discrimination, which is trained in the following way: S221: Encode the input sensor data and transform it into a representation vector through nonlinear transformation; S222: For each sensor fault category, preset... Given cluster centers, use K-means clustering to minimize the distance between the representation vector and the cluster center, and update the cluster center vector; S223: Randomly sample a cluster and obtain its nearest neighbors. Identify 1 cluster and randomly sample from each cluster. A vector of sensor data; S224: Calculate the loss function for the current iteration. ,in: The separation interval between clusters For the first sample The average value of the cluster representation vectors, For the first The nth cluster, the th The representation vector of each sensor after being encoded by an artificial neural network The category label to which the current representation vector belongs. For the first The variance of the representation vectors of each cluster, with positive subscripts indicating that negative values ​​are discarded and only positive values ​​are taken; S225: After a certain number of iterations, pause the training process, update the cluster center vector, and then resume gradient descent training. The trained sensor deviation fault detection model calculates the category during the prediction phase. ,in: For the nth sensor data sample The representation vector obtained after network transformation For this The mean vector of the closest cluster. For training The average value, This is the category label to which the current representation vector belongs.

4. The chiller unit fault diagnosis method based on reliable online data adaptive method according to claim 1 or 2, characterized in that, The sensor fault-free data prediction model is an artificial neural network based on a reliable online data adaptive strategy. During the training phase, the network minimizes the mean square error between the predicted data and the real sensor data through gradient descent. During the prediction phase, it predicts the sensor value when there is no fault based on the faulty sensor data after encoding. The aforementioned reliable online data adaptive strategy refers to: firstly, based on the data output after decoupling electrical-thermal faults, eliminating online data whose predicted entropy of the artificial neural network model exceeds a set threshold; secondly, given a parameter perturbation range, minimizing the maximum model predicted entropy within that perturbation range; and training an artificial neural network model based on the reliable online data adaptive strategy according to the above training method, using this model to judge the thermal state of the unit operation, and storing the judgment results in the diagnostic result database. This strategy only intervenes in the prediction phase of the model. It obtains the information entropy of the output result based on the fault probability distribution of the online output of the trained model. The higher the information entropy, the more likely the model is to give incorrect classification results. Therefore, in the online update phase, the model performs gradient descent updates based on entropy loss to reduce the model's entropy during the prediction phase. When applying the online data sharpness perception and adaptation strategy, the model only updates the lower-level normalization layer, specifically the statistical normalization parameters and affine parameters, to reduce the impact on the stability of the model's prediction results. The statistical normalization parameters are estimated based on real-time data, while the affine parameters... , It is calculated using gradient descent based on the entropy loss of real-time data; The aforementioned sensor fault-free data prediction model is trained by adjusting only the normalized layer in the artificial neural network based on a reliable online data adaptive strategy. Specifically, the input parameters are: intake temperature, exhaust temperature, intake pressure, condensing pressure, compressor current, refrigerant water inlet and outlet temperatures, cooling water inlet and outlet temperatures, condenser and evaporator inlet and outlet pressure differences, load increase / decrease temperatures, superheat, and subcooling parameters. The update of normalized parameters in the layer is as follows: , ,in: This refers to the compressor suction temperature. This refers to the compressor discharge temperature. For evaporation pressure, For condensation pressure, The mean vector in the normalization layer. The variance vector in the normalized layer. For the network front The encoding function of the layer, For network parameters At that time, the prediction entropy of the artificial neural network model; The entropy of the output of the trained artificial neural network: By minimizing loss To optimize network parameters: , , where: index function , This refers to the compressor suction temperature. This refers to the compressor discharge temperature. For evaporation pressure, For condensation pressure, This represents the perturbation of the neural network parameters. This is the upper limit of the disturbance amount. For the network's encoding function, For network parameters At that time, input the fault-free data vector of the chiller unit. The predictive entropy of an artificial neural network model; The upper limit of the set entropy threshold; , The parameters are updated as follows: , ,in: The scaling parameters for the normalization layer. The offset parameter of the normalized layer. This is a network loss.

5. A system for implementing the chiller unit fault diagnosis method based on reliable online data adaptation as described in any one of claims 1-4, characterized in that, include: The system includes a chiller unit operation data acquisition and storage unit, a fault diagnosis unit, and a host computer communication unit. Specifically, the chiller unit operation data acquisition unit collects data from the chiller unit PLC and stores it in a MySQL or Influx database. The fault diagnosis unit obtains real-time operation data of the chiller unit from the database, replaces outliers and null values ​​using an orthogonal state iteration method, decouples electrical and thermal faults, and then uses an artificial neural network based on a reliable online data adaptive strategy to determine thermal faults and obtain thermal fault prediction results. The host computer communication unit encapsulates the fault prediction results and auxiliary decision-making content and sends them to the host computer.

Images