Central air conditioning fault prediction method based on multi-source data fusion

Abstract

The application discloses a central air conditioner fault prediction method based on multi-source data fusion, relates to the technical field of air conditioner fault monitoring, and specifically comprises a closed-loop fault prediction system including a weak feature adaptive detection module, a multi-source data dynamic fusion module, a fault explainable attribution positioning module, a verification feedback module and a feedback scheduling center; a model-specific weak feature library is called to focus on, enhance and determine the weak features related to early faults; the multi-source features are dynamically weighted, redundantly identified and fused; based on the explicit mapping relationship and the feature contribution quantification results, the fault type, fault position and confidence are output; then, the consistency verification is performed in combination with the subsequent operation state, after-sales records and / or artificial review results, and the collaborative regulation instructions are formed through the feedback scheduling center to realize the closed-loop optimization of the model and rules. The method can improve the identification capability of the early weak faults of the central air conditioner, the fusion effectiveness and the explainability of the fault positioning.

Description

Technical Field

[0001] This invention relates to the field of air conditioning fault monitoring technology, and in particular to a method for predicting central air conditioning faults based on multi-source data fusion. Background Technology

[0002] As a core temperature control device in homes, commercial buildings, and industrial plants, the operational stability of central air conditioning systems directly impacts human comfort, production continuity, and energy efficiency. With the development of IoT and AI technologies, intelligent fault prediction and diagnosis has become a core development direction in the central air conditioning operation and maintenance field. By integrating multi-source sensors such as temperature, vibration, current, and pressure, and combining data fusion and algorithm analysis, early warning and precise fault location can be achieved. This significantly reduces the risk of sudden downtime, lowers maintenance costs, and extends equipment lifespan, representing a key requirement for the intelligent upgrading of modern air conditioning systems.

[0003] For example, Chinese invention patent CN114897271B discloses a predictive maintenance method for central air conditioning based on fault propagation in a digital twin environment. This method involves building a digital twin system for the central air conditioning system, including the physical central air conditioning equipment, a virtual 3D model, and a mathematical model. Secondly, it utilizes a design structure matrix modeling method to establish a multi-domain fault propagation model based on a "functional principle structure." Finally, based on real-time operational data and change prediction algorithms, it predicts potential high-risk faults, highlights them in the twin model as early warnings, and generates proactive predictive maintenance information, which is then fed back to maintenance personnel to provide a basis for maintenance decisions.

[0004] For example, Chinese invention patent CN118840629A discloses a predictive maintenance system for central air conditioning equipment based on multimodal data fusion, comprising: an infrared data acquisition module that monitors the infrared imaging and operating noise of the air conditioning equipment based on the operating environment of the central air conditioning system, captures hotspot signals using infrared sensors, and records sound changes using a microphone array to obtain an initial signal set; an anomaly feature analysis module that optimizes the hotspot areas in the infrared image based on the initial signal set, highlights anomaly features, performs frequency analysis on the sound signal through Fourier transform, extracts key frequency bands, and generates an anomaly feature set; and a feature fusion module that dynamically adjusts the weights of key features of the image and sound based on the anomaly feature set, optimizes the data fusion effect, and evaluates the data. The applicability of the system is assessed to form a fused feature body. The time-series analysis module, based on the fused feature body, uses an LSTM network to capture and learn the temporal changes of the features, match the characteristics of the feature data, and optimize memory utilization and computational efficiency to generate a time-dependent model. The lightweight training module, based on the time-dependent model, uses knowledge distillation technology to adjust the LSTM network architecture, optimize computational complexity by reducing computational layers and refining execution paths, and performs LSTM network compression and performance tuning to obtain a lightweight operation framework. The fault response module, based on the lightweight operation framework, monitors and analyzes the central air conditioning operation data, compares historical performance parameters, identifies potential deviations and anomalies, determines fault modes and severity, extracts key maintenance information, and generates fault mode identification results.

[0005] However, in the process of implementing the inventive technical solutions in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems: Existing technologies lack targeted feature enhancement and weak feature detection design, making it difficult to achieve early warning. For example, in the case of refrigerant micro-leakage in household multi-split air conditioners, the temperature difference and pressure change signals are usually weak, and existing technologies often miss these signals because they fail to capture them. Furthermore, existing technologies employ a fixed weight allocation strategy, resulting in poor multi-source data fusion effects. They also lack a scenario-based redundancy removal mechanism, leading to severe interference from invalid data, resulting in insufficient effective information density in the fusion results and decreased recognition accuracy. For example, in commercial cabinet air conditioners in factory scenarios, which are easily affected by vibrations from surrounding equipment, the inclusion of irrelevant data in the fusion often leads to misjudgments of fan malfunctions.

[0006] Furthermore, existing multi-source fusion models are mostly black-box architectures, lacking feature contribution analysis and predictive logic tracing functions. They cannot clearly identify the cause of faults or the basis for fault location, requiring after-sales personnel to conduct blind troubleshooting, significantly reducing maintenance efficiency. For example, when commercial chillers experience abnormal temperature differences, the model only indicates a refrigeration system fault, but cannot explain whether it is caused by heat exchanger blockage or refrigerant leakage. Simultaneously, the functional modules of existing technologies often operate independently, lacking a unified feedback scheduling and collaborative adjustment mechanism. This prevents continuous optimization of model parameters and strategies based on posterior results, resulting in poor system adaptability and difficulty in adapting to feature drift caused by changes in scenarios and equipment aging. For instance, after years of use, the compressor vibration baseline of a residential wall-mounted air conditioner changes, but existing technologies still detect based on the initial threshold, leading to a significant increase in false alarm rates. Summary of the Invention

[0007] To address the technical problems of existing technologies, such as weak early-stage minor fault identification capability, poor multi-source data fusion effect, lack of interpretability in fault attribution, and lack of closed-loop optimization mechanism, this invention provides a central air conditioning fault prediction method based on multi-source data fusion. The technical solution is as follows: A central air conditioning fault prediction method based on multi-source data fusion includes the following steps: S1, establishing a closed-loop fault prediction system, which includes a weak feature adaptive detection module, a multi-source data dynamic fusion module, a fault interpretability and attribution localization module, a verification feedback module, and a feedback scheduling center. The closed-loop fault prediction system is deployed on the controller, edge gateway, and / or cloud platform according to the central air conditioning unit model and application scenario; S2, using the weak feature adaptive detection module to process the central air conditioning unit's operating data, constructing and calling a model-specific weak feature library, focusing on, enhancing, and judging weak fault features related to early faults, and outputting suspected fault signals; S3, using the multi-source data dynamic fusion module to achieve optimized fusion of suspected fault signals and their corresponding weak fault features, establishing a three-dimensional mapping relationship between data value, fault mode, and model scenario, so as to dynamically assign weights to multi-source data according to the current model scenario and fault mode, and to handle redundancy. The remaining data is identified and removed to obtain fused feature results; the three-dimensional mapping relationship between data value, fault mode, and model scenario is trained by machine learning and calibrated in combination with air conditioner fault mechanism to improve the dynamic weighting adaptation capability of different models and different fault modes; S4, using the fault explainable attribution localization module, based on the explicit mapping relationship of features, fault type, and fault location, as well as the feature contribution quantification result, fault prediction results are generated, and prediction logic traceability information is recorded. The fault prediction results include at least candidate fault type, fault location, and confidence information; S5, using the verification feedback module, combined with the subsequent operating status of the equipment, after-sales records, and / or manual review results, the consistency of the fault prediction results is verified and standardized feedback information is generated; S6, using the feedback dispatch center to receive standardized feedback information, and classify, prioritize, and analyze the problem attribution of standardized feedback information to form coordinated adjustment instructions for different modules.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention constructs a closed-loop fault prediction system comprising a weak feature adaptive detection module, a multi-source data dynamic fusion module, a fault interpretability and attribution localization module, a verification feedback module, and a feedback scheduling center. This system enables the early identification of weak faults in central air conditioning systems. By setting up a weak feature adaptive detection module and combining it with a model-specific weak feature library, the system can focus on, enhance, and determine weak signs related to early faults. This allows for the earlier detection of progressive, low-intensity fault signals that are difficult to identify using traditional thresholding and conventional classification methods.

[0009] This invention establishes a three-dimensional mapping relationship between data value, fault modes, and machine model scenarios, dynamically assigns weights to data from different sources, and identifies and removes redundant data to avoid irrelevant data interfering with diagnostic results, thereby improving the fusion quality and diagnostic stability under different machine models and operating conditions.

[0010] This invention constructs an explicit mapping relationship between features, fault types, and fault locations, and combines the feature contribution quantification results to not only output candidate fault types, fault locations, and confidence levels, but also record prediction logic traceability information, which facilitates manual review, after-sales handling, and maintenance decisions.

[0011] This invention introduces standardized feedback from subsequent equipment operating status, after-sales records, and manual verification results. This feedback is then categorized and analyzed by the feedback dispatch center, further enabling coordinated adjustment and continuous optimization of various functional modules. This improves the accuracy, stability, interpretability, and long-term online operation capability of central air conditioning fault prediction. The feedback dispatch center further categorizes, prioritizes, and analyzes problem attribution, achieving targeted optimization of weak feature detection strategies, multi-source fusion strategies, and attribution localization strategies. Compared to solutions relying solely on single training or static model parameters, this invention can continuously correct the model and rules based on different models, application scenarios, and subsequent feedback, reducing the adverse effects of model drift, scenario switching, and changes in operating conditions on prediction performance.

[0012] The closed-loop fault prediction system involved in this invention can be deployed on controllers, edge gateways and / or cloud platforms according to the central air conditioning unit model and application scenario, taking into account both real-time processing requirements and cloud optimization capabilities, and is more suitable for the layered deployment and large-scale application of actual central air conditioning systems. Attached Figure Description

[0013] Figure 1 This is a diagram illustrating the closed-loop fault prediction architecture provided in an embodiment of this application. Figure 2 A flowchart of the weak feature adaptive detection module provided in the embodiments of this application; Figure 3 A flowchart of the multi-source data dynamic fusion module provided in the embodiments of this application. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0015] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0016] Embodiment 1 of the present invention: This embodiment takes a residential multi-split central air conditioning unit as the object and performs fault prediction on the operating status of the equipment under cooling, heating and dehumidification conditions in a residential weak interference scenario.

[0017] S1, establish a closed-loop fault prediction system, such as Figure 1The diagram shows the closed-loop fault prediction system architecture. This system includes a weak feature adaptive detection module, a multi-source data dynamic fusion module, a fault interpretability and attribution localization module, a verification feedback module, and a feedback dispatch center. These modules are deployed on the controller, edge gateway, and / or cloud platform, depending on the central air conditioning unit model and application scenario. Optionally, the controller is used for on-site data acquisition and local control execution; the edge gateway is used for data preprocessing, weak feature extraction, and local analysis with high real-time requirements; and the cloud platform is used for cross-device, cross-model, and / or cross-application scenario fusion analysis, model updates, and policy distribution.

[0018] The system includes the following modules: a weak feature adaptive detection module for preprocessing, noise reduction, enhancement, and weak fault feature extraction of at least one of the raw data collected during the operation of the central air conditioning system, including temperature, current, pressure, and vibration; a multi-source data dynamic fusion module for confidence assessment, weight allocation, redundancy removal, and fusion calculation of features from different sensor sources, time scales, and operating conditions; a fault explainable attribution localization module for attribution analysis of candidate fault modes based on the model and application scenario, outputting candidate fault types, fault locations, and corresponding confidence levels; a verification feedback module for receiving subsequent equipment operating status, after-sales records, and / or manual review results, verifying fault prediction results, and generating standardized feedback information; and a feedback scheduling center for classifying, sorting, and attribution analysis of the standardized feedback information, and issuing parameter correction, rule update, and / or model iteration instructions to the aforementioned modules to form a closed-loop fault prediction mechanism.

[0019] Preferably, the structured data fields for interaction between modules include at least the following: device identifier, model code, application scenario identifier, timestamp, feature identifier, original feature value, original deviation, percentage abnormal deviation, candidate fault mode code, fault location identifier, confidence level, and feedback identifier.

[0020] S2 uses a weak feature adaptive detection module to process the central air conditioning operation data, calls a preset model-specific weak feature library, focuses on, enhances and judges weak fault features related to early faults, and outputs suspected fault signals.

[0021] like Figure 2The flowchart shown is for the weak feature adaptive detection module. Specifically, the weak feature adaptive detection module preprocesses at least one of the raw data collected during the operation of the central air conditioning system, including temperature, current, pressure, vibration, and oil pressure. Preprocessing includes at least one of the following: DC removal, normalization, outlier suppression, missing value completion, and time series alignment, to reduce the impact of zero drift, noise interference, outliers, and missing sampling on the weak fault identification results, thereby obtaining a feature sequence suitable for subsequent weak feature extraction and judgment. Based on this, according to the central air conditioning unit model, application scenario, and candidate fault mode, preset model-specific weak feature detection methods are invoked. The feature library focuses on and enhances weak fault features with low amplitude, low frequency, and / or those easily masked by environmental interference that are associated with early faults. It selects feature dimensions, target frequency bands, and abnormal change directions with high correlation to the current candidate fault mode from a preset weak feature library specific to the model. Through parameter linkage adjustment and / or scene adaptive compensation, it improves the distinguishability of early weak fault features relative to background noise. Combined with an adaptive threshold judgment mechanism, it identifies abnormal features. That is, when the target feature in the feature sequence meets the preset abnormal triggering conditions, it outputs the corresponding suspected fault signal and sends the suspected fault signal to the multi-source data dynamic fusion module.

[0022] After preprocessing, the weak feature adaptive detection module reads the target feature dimension, target frequency band, abnormal change direction, and reference enhancement parameters corresponding to the current candidate fault mode from the preset weak feature library for the current central air conditioner, based on the model identifier, application scenario, and candidate fault mode. It then performs targeted focusing and enhancement on the weak features related to the early fault.

[0023] Targeted focusing and enhancement include at least one or more of the following methods: (1) For fault features that are low frequency or submerged by background noise, bandpass filtering, low-pass filtering, narrowband tracking filtering or wavelet decomposition are performed on the original feature sequence or the preprocessed feature sequence according to the target frequency band recorded in the model-specific weak feature library, so as to retain the frequency components related to the candidate fault mode and suppress random disturbances and environmental interference in irrelevant frequency bands.

[0024] In one specific embodiment, when the candidate fault mode is early wear of the compressor bearing, the vibration target frequency band corresponding to the fault mode in the model-specific weak feature library is called, and bandpass filtering and envelope analysis are performed on the vibration signal to extract the envelope energy, peak factor and frequency band energy ratio within the fault frequency band; when the candidate fault mode is refrigerant micro-leakage, the low-frequency frequency band related to slow drift of evaporation temperature and low fluctuation of suction pressure in the library is called, and trend separation and low-frequency residual extraction are performed on the temperature and pressure sequences to highlight the continuous slow shift characteristics.

[0025] (2) For minor fault characteristics that are easily masked by environmental interference, first establish a normal operation compensation mechanism based on the application scenario, and then enhance the compensation residual. The application scenario should include at least one of the following: high temperature environment, low load operation, frequent start-stop, low load at night, and seasonal change.

[0026] Specifically, historical features of the core operating parameters of the equipment are extracted from historical data to obtain predicted values ​​under normal operating conditions, based on factors such as ambient temperature, load rate, set temperature, and running time. Then, the real-time measured values ​​are compared with the predicted values ​​to obtain the scenario compensation residual. The scenario compensation residual is used as compensation to reduce the masking effect of external operating condition fluctuations on fault judgment.

[0027] For example, in high-temperature and high-humidity scenarios, the natural rise in condensing pressure can easily mask subtle anomalies caused by minor condenser blockage. In this case, a condensing pressure compensation prediction model under normal operating conditions is first established based on ambient temperature, fan speed, and load rate. Model training needs to be based on historical normal operating data, focusing on extracting the correlation characteristics between ambient temperature and condensing pressure, the adjustment characteristics of fan speed on condensing pressure, and the adaptation characteristics of load rate and condensing pressure under this scenario. The model is trained using these historical core features to output the predicted condensing pressure value under the current high-temperature and high-humidity conditions. Then, the difference between the real-time measured condensing pressure value and this predicted value is calculated to obtain the scenario compensation residual. This residual is used as the basis for compensation to offset the natural rise in condensing pressure caused by the high-temperature and high-humidity environment. The fluctuation reflected in the residual represents the subtle fault characteristics such as minor blockage that may exist in the condenser after excluding environmental interference, thus effectively reducing the masking effect of external operating condition fluctuations on fault diagnosis and avoiding missed fault detection.

[0028] A model-specific weak fault feature library stores sets of weak fault features related to early faults across different models, application scenarios, and fault modes, along with corresponding reference thresholds, feature weights, and judgment rules. The weak fault feature set includes the weak fault features themselves, along with their corresponding feature dimensions, target frequency bands, and abnormal change directions. Feature dimensions include at least one of time-domain features, frequency-domain features, time-frequency features, correlation features, and deviation features; abnormal change directions include at least one of increasing anomalies, decreasing anomalies, intensified fluctuation anomalies, and periodically enhanced anomalies. During retrieval, based on the current central air conditioning unit's model identifier, application scenario, and operating conditions, weak fault features matching the central air conditioning unit's model, application scenario, and candidate fault mode are selected from the model-specific weak fault feature library.

[0029] The model-specific weak feature library categorizes historical operating data into hierarchical groups based on preset central air conditioning model category tags, application scenario tags, and fault mode tags, forming corresponding normal sample sets and early fault sample sets. After preprocessing the raw operating data in the sample sets, time-domain features, frequency-domain features, time-frequency features, correlation features, and deviation features are extracted from temperature, pressure, current, vibration, oil pressure, and / or speed data to form a candidate weak fault feature set. The candidate weak fault features are then screened, retaining those with high correlation to the target early fault and stable characterization capabilities under the corresponding model and application scenario. The feature configuration includes normal reference values, initial judgment thresholds, feature weights, and judgment rules. After generating corresponding feature entries, they are written into the model-specific weak feature library. The judgment rules include at least preset abnormal triggering conditions.

[0030] In this embodiment, a model-specific weak feature library is established using model category labels, application scenario labels, and fault mode labels as indexes to create multi-layered feature entries. Each feature entry includes at least the following: model identifier; application scenario identifier; candidate fault mode identifier; sensor type; feature dimension identifier; target frequency band; and direction of abnormal change.

[0031] The screening process includes, but is not limited to, model matching screening, which removes feature entries that are not applicable to the current model based on the model identifier. In one specific embodiment, centrifugal units and screw units differ in vibration frequency, oil pressure response, and load change patterns, so only feature entries that are consistent with the current model or shared by similar models are retained. Scene matching screening removes feature entries that do not match the current scene based on the current application scene label. In one specific embodiment, in a high-temperature and high-humidity machine room environment, pressure deviation features verified by high-temperature scene modeling are prioritized for retention; in a low-load nighttime operation scenario, features of gradual change trend and residual features are prioritized for retention.

[0032] The adaptive threshold determination mechanism can unify the determination criteria for different types of features, convert each feature entering the fault determination into a comparable anomaly representation quantity, and compare it with the corresponding determination threshold; when any target feature or combination of target features meets the preset anomaly triggering condition, a corresponding suspected fault signal is generated. The suspected fault signal includes at least one of the following: candidate fault type, trigger feature identifier, anomaly degree information, and time stamp.

[0033] In one specific embodiment, for each selected target feature, the abnormality representation quantity of the feature is first calculated based on the normal reference value, normal fluctuation range, and abnormal change direction corresponding to that feature. Preferably, for increasing abnormal features, the abnormality representation quantity is the positive deviation of the current feature value from the normal reference value; for decreasing abnormal features, the abnormality representation quantity is the negative deviation of the current feature value from the normal reference value; for bidirectional abnormal features, the abnormality representation quantity is the absolute deviation between the current feature value and the normal reference value; for trend-type features, the abnormality representation quantity is the slope of change within the sliding window, the cumulative deviation, or the duration of continuous offset. For each target feature, the threshold is not fixed, but is adaptively updated based on the device model, scenario, operating condition, and recent normal operating status. The adaptive threshold includes, but is not limited to, the initial threshold corresponding to the current device model; and the compensation coefficient corresponding to the current application scenario. Current load rate range; current ambient temperature range; characteristic baseline during the most recent stable operating period; equipment aging level or operating time. Preset anomaly triggering conditions include at least one or more of the following: single-feature instantaneous trigger, i.e., the abnormal characterization of a target feature exceeds the corresponding judgment threshold; single-feature continuous trigger, i.e., a target feature exceeds the corresponding judgment threshold for M consecutive analysis windows; single-feature cumulative trigger, i.e., in the most recent N analysis windows, the number of times a target feature exceeds the judgment threshold reaches a set number; multi-feature combined trigger, i.e., the sum of the weighted abnormal characterization of multiple target features exceeds the combined threshold; master-slave joint trigger, i.e., the master feature exceeds the threshold, and at least one auxiliary feature simultaneously shows an anomaly in a predetermined direction.

[0034] In one specific embodiment, taking a residential wall-mounted central air conditioning unit as an example, a model-specific weak feature library can pre-store weak fault features related to early faults of that model. Preferably, early faults include at least refrigerant micro-leakage, compressor winding overheating, and early wear of the fan bearing; correspondingly, weak fault features can include at least: for refrigerant micro-leakage, features such as slight pressure drop, abnormal indoor-outdoor temperature difference, and changes in current harmonic distortion rate are selected; for compressor winding overheating, features such as winding temperature rise, current fluctuation, and local vibration changes are selected; for early wear of the fan bearing, features such as abnormal vibration sideband, speed fluctuation, and auxiliary current features are selected. In actual use, based on the current equipment model identifier, application scenario, and operating conditions, the corresponding feature set and judgment parameters are retrieved from the model-specific weak feature library for subsequent weak feature focusing, enhancement, and anomaly judgment.

[0035] Preferably, the weak feature adaptive detection module first preprocesses the raw data collected by the sensor. Preferably, the raw data sequence within a single sampling window is used as the processing object, and the raw data sequence is... The sampling window length can be n=1024. For the original data, DC removal, normalization, outlier removal, and linear interpolation of adjacent points are performed sequentially to eliminate zero drift, unify data scale, and suppress interference from abnormal sampling points. To enhance early anomalous features with low amplitude and low frequency, wavelet packet decomposition can be performed on the preprocessed data. Preferably, a 3-level wavelet packet decomposition can be performed using the db4 wavelet basis, and an enhancement factor is applied only to the wavelet packet coefficients corresponding to the target frequency band, while the coefficients of non-target frequency bands remain unchanged. The enhancement factor can be preset by those skilled in the art, and the enhancement factor can be 1.3. The target frequency band may include the frequency bands corresponding to the second and third harmonics of the current, as well as the low-frequency components corresponding to slow abnormal temperature fluctuations. After reconstructing the enhanced wavelet packet coefficients, the enhanced time-frequency domain signal is obtained.

[0036] S3 utilizes a multi-source data dynamic fusion module to optimize and fuse suspected fault signals and their corresponding weak fault features, establishing a three-dimensional mapping relationship between data value, fault mode, and aircraft model scenario. This allows for dynamic weighting of multi-source data based on the current aircraft model scenario and fault mode, as well as the identification and removal of redundant data to obtain fused feature results.

[0037] like Figure 3 The diagram shows the flowchart of the multi-source data dynamic fusion module. This module optimizes and fuses suspected fault signals and their corresponding weak fault features. Specifically, after receiving the suspected fault signals and corresponding multi-source candidate features output by S2, a three-dimensional mapping relationship is established between data value, fault mode, and machine model scenario. Data value characterizes the effective contribution of each data source or feature dimension to fault identification in the current machine model scenario and fault mode. The machine model scenario includes at least one of machine type, application scenario, and operating condition. The fault mode includes at least one of refrigerant micro-leakage, compressor winding overheating, and early wear of fan bearings. The three-dimensional mapping relationship between data value, fault mode, and machine model scenario refers to the structured association of the contribution relationships of different data sources or different feature dimensions in different machine model scenarios and fault modes. This allows for the rapid retrieval of the corresponding data value assessment results and weighting rules after the current machine model scenario and fault mode are determined.

[0038] In one specific embodiment, the three-dimensional mapping relationship between data value, fault modes, and machine model scenarios uses labeled historical operating data as training samples. These labels include at least machine model category labels, application scenario labels, operating condition labels, fault mode labels, fault stage labels, and influence weight labels. The training samples include at least normal states, refrigerant micro-leakage states, compressor winding overheating states, and fan bearing early wear states. Operating condition labels include, but are not limited to, load rate ranges, ambient temperature ranges, start-stop states, day / night cycles, and seasonal information.

[0039] Based on the three-dimensional mapping relationship between data value, fault modes, and machine model scenarios, candidate features from different sources such as temperature, pressure, current, vibration, and / or oil pressure are evaluated for data value and dynamically weighted. Features with high correlation to the current machine model scenario and fault mode, good stability, and significant diagnostic contribution are given higher weights. Subsequently, redundancy identification and elimination are performed on the weighted multi-source features, removing data with high information redundancy, low supplementary gain, or insufficient correlation with the current fault mode, while retaining features with strong complementarity to obtain the fused feature result. The three-dimensional mapping relationship between data value, fault modes, and machine model scenarios is obtained through machine learning training and calibrated in conjunction with the fault mechanism of central air conditioning systems to improve the dynamic weighting adaptability and interpretability of the fusion results across different machine models, application scenarios, and fault modes.

[0040] Preferably, redundancy identification and removal of weighted multi-source features can be performed by analyzing the redundancy between sources of the retained data. When the information redundancy between two or more data sources is higher than a preset threshold, the core features are retained and their feature weights are reduced, thereby reducing the interference of duplicate information on the fusion results.

[0041] In one specific embodiment, the three-dimensional mapping relationship between data value, fault modes, and machine model scenarios can be constructed through machine learning training. Preferably, historical operating data is used as training samples; the machine model category, application scenario, and fault mode corresponding to each data source are used as inputs, and the influence weight of each feature on the target fault identification is used as the output to train an initial mapping model. After training, the model output results are calibrated in conjunction with the fault mechanism of central air conditioning to avoid weighting bias caused by pure data-driven approaches that are inconsistent with the actual fault mechanism.

[0042] In one specific embodiment, to identify redundant relationships between multi-source data, the correlation coefficient between each pair of retained data sources can be calculated as the information redundancy degree to evaluate its information redundancy level. Let the feature sequences corresponding to the two data sources be X=(x1,x2,...,x...). n ) and Y=(y1,y2,...,y n The linear correlation between the two can be quantified using the Pearson correlation coefficient, which can be calculated using the following formula: in and These are the means of sequences X and Y, respectively. The Pearson correlation coefficient r ranges from [-1, 1], with a larger |r| indicating a stronger correlation. When the absolute value of the Pearson correlation coefficient is higher than the preset redundancy threshold, it is determined that there is high homogeneous redundancy between the two data sources. In this case, the data source with higher data value or higher fault sensitivity is preferred, and the other data source is removed or its preset weight is reduced. When the absolute value of the Pearson correlation coefficient is lower than the preset redundancy threshold, it is determined that the two data sources have good complementarity and can participate in subsequent fusion.

[0043] In one specific embodiment, after the model training obtains the initial weight mapping relationship, it is then calibrated by combining it with the central air conditioning fault mechanism mapping relationship preset by those skilled in the art. For example, in the refrigerant leakage fault mode, the influence weight of pressure drop and temperature difference abnormality related features is increased by a preset ratio; in the compressor winding overheating fault mode, the influence weight of winding temperature rise and current fluctuation related features is increased by a preset ratio; in the fan bearing wear fault mode, the influence weight of vibration sideband abnormality and speed fluctuation related features is increased by a preset ratio. The central air conditioning fault mechanism mapping relationship preset by those skilled in the art can be obtained by analyzing the refrigeration cycle mechanism, electrothermal effect mechanism, and mechanical vibration mechanism of the central air conditioning system, and by combining the unit design data, historical operating data, fault maintenance records, simulated fault test results, and expert experience rules for inductive modeling; the fault mechanism mapping relationship at least characterizes the correlation between different fault modes and corresponding sensitive data sources, key fault features, abnormal change directions, and weight correction coefficients.

[0044] S4. Using the fault explainable attribution localization module, based on the explicit mapping relationship between features, fault type and fault location and the feature contribution quantification results, a fault prediction result is generated and the prediction logic traceability information is recorded. The fault prediction result includes at least candidate fault type, fault location and confidence information.

[0045] Preferably, to achieve interpretable attribution of fault types and fault locations, an explicit mapping relationship between features, fault types, and fault locations can be pre-constructed to describe the correspondence between different fused features and different fault types and fault locations. By invoking the explicit mapping relationship, while outputting candidate fault types, the corresponding fault locations can be further determined, and the correlation basis between features and fault types and fault locations can be provided, thereby improving the interpretability and traceability of fault prediction results.

[0046] In one specific embodiment, abnormal current harmonic distortion rate can correspond to compressor winding fault, and the fault location can be located in the compressor winding; abnormal pressure fluctuation can correspond to refrigerant leakage fault, and the fault location can be located in the evaporator interface or pipeline connection; abnormal vibration sideband can correspond to fan fault, and the fault location can be located in the fan bearing. Therefore, through explicit mapping relationships, not only can candidate fault types be output, but also their corresponding fault locations, as well as the mapping basis between relevant features and faults.

[0047] In one specific embodiment, the process of constructing an explicit mapping relationship between features, fault types, and fault locations involves collecting and organizing tagged historical operating samples. These samples include at least machine type tags, application scenario tags, operating condition tags, fault type tags, and fault location tags. The fault type tags include at least one of refrigerant leakage, compressor winding overheating, and fan bearing wear. The fault location tags include at least one of compressor windings, evaporator interfaces, pipe connections, and fan bearings. The historical operating samples are preprocessed, and candidate features are extracted from data such as temperature, pressure, current, vibration, oil pressure, and / or speed to form a candidate feature set. The candidate features include at least one or more of the following: time-domain features, frequency-domain features, time-frequency features, trend deviation features, residual features, and correlation features.

[0048] Furthermore, to quantify the contribution of each fused feature to the candidate fault prediction result, let the feature set involved in fault prediction be F={f1,f2,...,f...} m The fusion weights for each feature are W = {w1, w2, ..., w}. m The anomaly coefficient A = {a1, a2, ..., a} is the characteristic anomaly coefficient. m}, where a i ∈[0,1].

[0049] Based on the fusion weights and anomaly coefficients of each feature, the contribution C of each feature to the current candidate fault result is calculated. i This is to obtain the feature contribution metric results. Preferably, the first... i Contribution C of each feature i It can be represented as: Among them, C i w represents the relative contribution of the i-th feature to the current candidate fault result. i This represents the weight of the feature in the fusion process, a. i This represents the anomaly coefficient corresponding to the feature.

[0050] When outputting fault prediction results, the fault explainable attribution localization module also records prediction logic traceability information simultaneously. The prediction logic traceability information includes at least the feature set involved in the attribution, the abnormality information of each feature, the feature contribution information, the candidate fault type, the fault location and confidence information, so as to support subsequent verification, manual review and fault traceability analysis.

[0051] In one specific embodiment, the anomaly coefficient a i Used to characterize the degree of abnormality of the i-th feature relative to its fault determination threshold. Preferably, a i The percentage-based anomaly deviation D of this feature can be used as a basis. i With the corresponding fault determination threshold T i Perform the calculation, and a i The value range of a is limited to [0,1]. The anomaly coefficient a i It can be determined by the following formula: = , where D i T represents the percentage-based abnormal deviation of the i-th feature, characterizing the degree of deviation of the current value of the i-th feature from its normal reference value or compensated reference value. i This is the percentage-based fault determination threshold corresponding to the i-th feature. When D i When Ti is not reached, a i Less than 1; when D i When Ti is reached or exceeded, a i Approaching or taking 1.

[0052] Furthermore, percentage-type abnormal deviation D i It can be derived from the original deviation Δ i Normalization yields the result; where the original deviation Δ i The original engineering dimensions are retained for operation and maintenance display, log recording, and subsequent traceability analysis, while the percentage-type anomaly deviation Di is used for anomaly degree calculation and feature contribution quantification.

[0053] S5 utilizes the verification feedback module to combine the equipment's subsequent operating status, after-sales records, and / or manual review results to perform consistency verification on the fault prediction results and generate standardized feedback information.

[0054] Preferably, the consistency verification includes at least the verification of subsequent equipment operation data, i.e., determining whether the abnormalities related to the target fault continue to appear or further manifest within a preset observation time window; after-sales record verification, i.e., confirming the actual fault type and fault location based on maintenance work orders, inspection records, and parts replacement records; and manual review verification, i.e., the fault prediction results are confirmed by maintenance personnel in conjunction with on-site detection results. Furthermore, the verification feedback module can also verify the sensor status and data link status to identify data quality problems caused by sensor drift, communication anomalies, or sampling anomalies.

[0055] In one specific embodiment, after S4 outputs the fault prediction result, the verification feedback module first generates a verification task corresponding to the prediction result. The verification task includes at least the device identifier, model code, application scenario identifier, prediction time, candidate fault type, predicted fault location, and prediction confidence level. The subsequent running verification unit retrieves the corresponding target verification feature set from the preset fault verification rule base based on the current candidate fault type and predicted fault location. Within the observation time window, subsequent equipment operation data is continuously collected, and the target verification features are analyzed window by window. The resulting verification quantities include at least the persistence rate, i.e., the number or percentage of times the target anomaly occurs continuously within the observation time window; the enhancement level, i.e., whether the target anomaly has further increased relative to the prediction time; the accompanying consistency, i.e., whether auxiliary anomalies consistent with the fault mechanism occur simultaneously; and the event responsiveness, i.e., whether protection shutdown, derating operation, or system alarm is triggered.

[0056] In one specific embodiment, when the candidate fault type is refrigerant micro-leakage, the target verification feature set includes a continuous decrease in suction pressure, a decrease in evaporation temperature, an increase in superheat, and a gradual decrease in compressor current; when the candidate fault type is compressor winding overheating, the target verification feature set may include a continuous increase in winding temperature rise, aggravated current fluctuations, increased current imbalance, and abnormal casing temperature; when the candidate fault type is early wear of the fan bearing, the target verification feature set may include an increase in vibration sideband energy, an enhancement of envelope spectrum peak value, aggravated speed fluctuations, and abnormal motor current pulsation.

[0057] In one specific embodiment, all the above-mentioned verification quantities are uniformly converted into standardized feedback information. This standardized feedback information includes, but is not limited to, device identifier, model code, application scenario identifier, prediction time, observation time window, candidate fault type, predicted fault location, prediction confidence level, verification source, verification result, deviation type, cause explanation, severity, manual intervention identifier, and feedback identifier. This standardized feedback information provides a unified, comparable, and quantifiable post-verification basis for subsequent feedback to the dispatch center.

[0058] S6 utilizes the feedback dispatch center to receive standardized feedback information, and performs classification processing and problem attribution analysis on the standardized feedback information to form coordinated adjustment instructions for different modules.

[0059] In one specific embodiment, after receiving standardized feedback information, the feedback scheduling center first performs unified formatting on the equipment identifier, model code, application scenario identifier, prediction time, deviation type, verification result, and feedback identifier in the standardized feedback information, and establishes association relationships with the corresponding prediction logic traceability information, model version information, and rule version information to form corresponding feedback task objects. The feedback task object includes at least the equipment identifier, model code, application scenario identifier, prediction result information, verification result information, deviation type information, attribution feature set, feature contribution information, data quality status, and manual intervention status.

[0060] The feedback dispatch center categorizes feedback tasks based on the verification results. Feedback tasks are classified into one or more of the following categories: consistent, partially consistent, inconsistent, and pending confirmation. Further, based on deviation type, feedback tasks are subdivided into one or more of the following categories: unbiased, false alarms, missed alarms, attribution bias, location bias, and data quality anomalies. Finally, aggregation and statistics are performed based on machine model code, application scenario identifier, fault type, and model version to identify systemic problems occurring in clusters.

[0061] After classification, the feedback dispatch center performs problem attribution analysis on the feedback task objects. Specifically, it executes collaborative adjustment instructions based on the deviation type, verification source, and key attribution feature set. The collaborative adjustment instructions are used to adjust the observation time window, after-sales record parsing rules, manual review trigger conditions, and data quality anomaly judgment rules.

[0062] The feedback dispatch center tracks the execution process of the coordinated adjustment instructions and records the instruction issuance time, execution module, execution status, execution completion time, and parameter changes. After the coordinated adjustment instructions are executed, the center determines whether the adjustment instructions are effective based on the subsequent feedback results. If effective, the corresponding correction rules or model update results are written into the rule base or model management record. If ineffective, the center triggers a re-attribution analysis or upgrades the task to manual expert review.

[0063] In one specific embodiment, the preset decision logic is pre-established by the feedback dispatch center based on historical prediction results, verification results, after-sales records, manual review results, and post-adjustment effect evaluation data. It is based on fault mechanisms, expert experience, model rules, scenario rules, parameter boundaries, and safety constraints. Then, the feedback dispatch center generates corresponding collaborative adjustment instructions according to the preset decision logic.

[0064] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0065] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0066] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0067] The above description is only an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting central air conditioning faults based on multi-source data fusion, characterized in that, Includes the following steps: S1. Establish a closed-loop fault prediction system. The closed-loop fault prediction system includes a weak feature adaptive detection module, a multi-source data dynamic fusion module, a fault explainable attribution localization module, a verification feedback module, and a feedback scheduling center. The closed-loop fault prediction system is deployed on the controller, edge gateway, and / or cloud platform according to the model and application scenario of the central air conditioning unit. S2, using the weak feature adaptive detection module to process the central air conditioning operation data, construct and call the model-specific weak feature library, focus on, enhance and judge the weak fault features related to early faults, and output the suspected fault signal; S3, the multi-source data dynamic fusion module is used to optimize and fuse suspected fault signals and their corresponding weak fault features, and establish a three-dimensional mapping relationship between data value, fault mode and model scenario, so as to dynamically assign weights to multi-source data according to the current model scenario and fault mode, and identify and remove redundant data to obtain fusion feature results. The data value, fault modes, and three-dimensional mapping relationship of model scenarios are trained by machine learning and calibrated in combination with air conditioner fault mechanisms to improve the dynamic weighting and adaptation capability of different models and different fault modes. S4. Using the fault explainable attribution localization module, based on the explicit mapping relationship between features, fault type and fault location and the feature contribution quantification result, a fault prediction result is generated, and prediction logic traceability information is recorded. The fault prediction result includes at least candidate fault type, fault location and confidence information. S5. Using the verification feedback module, combined with the subsequent operating status of the equipment, after-sales records and / or manual review results, the consistency of the fault prediction results is verified and standardized feedback information is generated. S6. The standardized feedback information is received by the feedback scheduling center, and the standardized feedback information is classified, prioritized, and analyzed to form coordinated adjustment instructions for different modules.

2. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 1, characterized in that, In the closed-loop fault prediction system, the controller is used for on-site data acquisition and local control execution, the edge gateway is used for data preprocessing, weak feature extraction and local analysis with high real-time requirements, and the cloud platform is used for cross-device, cross-model and / or cross-application scenario fusion analysis, model update and strategy distribution.

3. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 1, characterized in that, The central air conditioning operation data includes at least one of temperature, current, pressure, vibration and / or oil pressure. The central air conditioning operation data is preprocessed, and the preprocessing includes at least one of DC removal, normalization, outlier suppression, missing value completion and time series alignment. Based on the model, application scenario, and candidate fault modes of the central air conditioning system, the target feature dimension, target frequency band, abnormal change direction, and judgment rules corresponding to the current candidate fault mode are read from the model-specific weak feature library. The weak fault features related to the early fault are targeted and enhanced, and the corresponding suspected fault signal is output when the preset abnormal triggering conditions are met.

4. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 3, characterized in that, The model-specific weak feature library is used to store a set of weak fault features related to early faults in the model, application scenario and fault mode, as well as the reference threshold, feature weight and judgment rules corresponding to the weak fault features; The targeted focusing and enhancement of the weak fault features includes at least one of the following methods: Perform bandpass filtering, lowpass filtering, narrowband tracking filtering, or wavelet decomposition on the original or preprocessed feature sequence based on the target frequency band. To address the subtle fault characteristics that are easily masked by environmental interference, a normal operation compensation mechanism is established based on the application scenario, and enhancement processing is performed based on the scenario compensation residual between real-time measured values ​​and predicted values.

5. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 1, characterized in that, The data value, fault mode, and three-dimensional mapping relationship of model scenario are obtained by training with labeled historical operating data and calibrated in combination with the fault mechanism of central air conditioning. The tags include at least the model category tag, application scenario tag, operating condition tag, fault mode tag, and impact weight tag; The aforementioned model scenario includes at least one of the following: model category, application scenario, and operating condition; Based on the aforementioned three-dimensional mapping relationship, data value assessment and dynamic weighting are performed on candidate features from different data sources.

6. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 5, characterized in that, Redundancy identification and removal are performed on the weighted multi-source features, including: Perform inter-source redundancy analysis on the retained data to evaluate the redundancy relationship between different data sources based on information redundancy. When the information redundancy between two or more data sources exceeds a preset redundancy threshold, retain information from any one data source. When the information redundancy is lower than the preset redundancy threshold, all data source information is retained to obtain the fused feature result.

7. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 1, characterized in that, An explicit mapping relationship between features, fault types and fault locations is pre-constructed, and the contribution of each feature to the current candidate fault result is calculated based on the fusion weight and anomaly coefficient of each fusion feature involved in fault prediction, so as to obtain the feature contribution quantification result. The prediction logic traceability information includes at least the feature set involved in the attribution, information on the degree of anomaly of each feature, information on the contribution of each feature, candidate fault type, fault location, and confidence level information.

8. The central air conditioning fault prediction method based on multi-source data fusion as described in claim 1, characterized in that, The consistency check includes at least: Equipment subsequent operation data verification, that is, determining whether the target fault-related anomalies continue to appear or further manifest within the preset observation time window; After-sales record verification, that is, confirming the actual fault type and fault location based on repair work orders, inspection records and / or parts replacement records; Manual verification and validation, which involves maintenance personnel confirming the fault prediction results in conjunction with on-site testing results; The standardized feedback information includes at least the equipment identifier, model code, application scenario identifier, prediction time, observation time window, candidate fault type, predicted fault location, prediction confidence level, verification source, verification result, deviation type, and feedback identifier.

9. The central air conditioning fault prediction method based on multi-source data fusion according to claim 1, characterized in that, After receiving standardized feedback information, the feedback scheduling center first performs unified formatting processing on the standardized feedback information, and establishes association relationships with the corresponding prediction logic traceability information, model version information and rule version information to form feedback task objects.

10. The central air conditioning fault prediction method based on multi-source data fusion according to claim 1, characterized in that, The coordinated adjustment command is used to adjust at least one of the following: observation time window, after-sales record parsing rules, manual review triggering conditions, and data quality anomaly judgment rules; The feedback dispatch center tracks the execution process of the coordinated adjustment instructions and records the instruction issuance time, execution module, execution status, execution completion time, and parameter changes.

Images