Anti-forgetting central air conditioner fault diagnosis continuous incremental learning method and system

Abstract

The invention discloses an anti-forgetting central air conditioner fault diagnosis continuous incremental learning method and system, and belongs to the field of building heating ventilation air conditioner intelligent operation and maintenance. The method comprises the following steps: firstly, acquiring operation data of the central air conditioner on line, and constructing an information enhanced feature sample; then, dynamically diagnosing and quantifying the mastery degree of the model on various fault knowledge periodically from three dimensions of category confusion degree, decision ambiguity and learning loss, and identifying knowledge weak points; further, adaptively adjusting a sampling weight according to a diagnosis result, and constructing a mixed training batch considering new and old knowledge and weak point reinforcement by adopting a three-in-one experience playback strategy; and finally, updating the model through a parameter-efficient adapter fine tuning technology and a composite loss function considering classification and knowledge distillation. According to the method, the problems of data drift and disastrous forgetting of a traditional data-driven model under a dynamic working condition are solved, and long-term, stable and efficient operation of the fault diagnosis model is realized.

Description

Technical Field

[0001] This invention belongs to the field of cloud monitoring and intelligent operation and maintenance technology for building HVAC systems, specifically involving a continuous incremental learning method and system for diagnosing central air conditioning faults that is resistant to forgetting. Background Technology

[0002] Central air conditioning systems are an indispensable core facility in modern large public and commercial buildings, and are typically the largest contributor to building energy consumption. According to the International Energy Agency (IEA), building operation accounts for approximately 30% of global final energy consumption, with HVAC systems accounting for a particularly significant proportion. If system malfunctions are not detected and addressed promptly, they can lead to an additional 15% to 30% in energy consumption, significantly increasing operating costs. Therefore, developing an efficient and reliable central air conditioning fault diagnosis system is crucial for building energy conservation and improving system reliability.

[0003] Traditional central air conditioning fault diagnosis methods mostly rely on physical mechanism models or expert experience to determine fault states. Mechanism models often struggle to construct universal models applicable to various operating conditions when faced with complex system coupling and nonlinear characteristics; while expert experience is costly to acquire and difficult to update rules, exhibiting poor adaptability to new or complex faults. With the rise of artificial intelligence in recent years, more and more research is adopting data-driven methods for fault diagnosis. These methods automatically learn fault patterns based on historical operating data, overcoming the limitations of manual modeling. However, in practical applications, the operating status of central air conditioning systems constantly changes due to seasonal variations, equipment aging, or system maintenance, causing the distribution of input data to shift over time. This phenomenon is known as "data drift." Traditional data-driven models often assume that the training and testing data distributions are consistent, lacking the ability to adapt to changes. Once deployed online, their diagnostic performance easily degrades over time, even leading to serious misjudgments. Researchers have used continuous online training to enable fault diagnosis models to better adapt to new data, but this also reduces the model's ability to judge older data. As the seasons change, when the diagnostic system encounters the same operating conditions again, it struggles to recognize patterns. The phenomenon that this model loses old knowledge while learning new data is called the "knowledge forgetting" problem. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, the present invention aims to provide a continuous incremental learning method and system for central air conditioning fault diagnosis that is resistant to forgetting. This method is an effective way to both follow dynamic changes in operating conditions and gradually adapt to new data while retaining historical diagnostic capabilities. In other words, it is a diagnostic system with online continuous learning capabilities, enabling it to function reliably in actual engineering scenarios for a long time and solving the performance degradation problem caused by data drift and knowledge forgetting in traditional data-driven models.

[0005] This invention provides the following technical solution: A cloud-based, continuously learning, and forgetting-resistant central air conditioning fault diagnosis system with continuous learning capabilities is disclosed. This system integrates a model training difficulty assessment mechanism and a fault adaptive playback strategy, enabling it to continuously receive new operating condition data while retaining historical knowledge during model operation in the cloud. By dynamically adjusting the focus of cloud-based continuous learning and rationally organizing historical sample playback, the system can achieve long-term, stable, and highly accurate fault identification, suitable for the intelligent cloud-based operation of building central air conditioning systems in complex and changing environments. It includes: Data interface module: Used to receive real-time data in JSON format from the building's central air conditioning system; Task buffer module: Instantiates a thread-safe object in program memory as an asynchronous decoupled buffer pool for front-end data reception and back-end model training; Background processing module: Creates an independent thread that, after initialization, enters an infinite loop to wait for and process data in the task queue; Data storage module: Used to store samples covering all known fault types, the latest operational data, and historical data; Model management module: includes student models for continuously receiving new knowledge and updating parameters online, and teacher models that serve as reference benchmarks for retaining old knowledge in student models.

[0006] Incremental learning in a continuous incremental learning system for central air conditioning fault diagnosis that resists forgetting includes the following steps: Step 1: Online Data Acquisition and Feature Extraction The continuous learning system receives real-time operational data from the central air conditioning system through an interface module, which is recorded as the original sample. Through an automated online processing workflow, it is transformed into an enhanced feature sample with higher information density and better revealing the system's operational status. This step enhances the data by integrating physical mechanisms and statistical data, considering three dimensions: steady-state performance, dynamic stability, and long-term evolutionary trends. This provides high-quality, highly discriminative input for subsequent incremental learning models. The specific process is as follows: Step 1.1: Extract thermodynamic features reflecting the system state. From the collected raw data, the continuous learning system will analyze the data at each time point... Temperature and pressure measurements are used to calculate two physical characteristics related to the refrigeration cycle state: superheat. With supercooling The specific calculation method is as follows: (1); (2); in, and These are the evaporator outlet temperature and the condenser outlet temperature, respectively. and These are the compressor's intake and exhaust pressures, respectively. These are the saturation temperatures under the corresponding pressure. These two parameters can intuitively reflect whether the air conditioning system is operating in a stable and energy-saving range, and are important thermal reference indicators for diagnosing faults.

[0007] Step 1.2: Extract time-series features reflecting the dynamic characteristics of the air conditioning system. Besides instantaneous snapshots, the dynamic behavior of the air conditioning system also contains rich fault information. This step extracts time-series features reflecting past... By using time windows of sampling points, the two types of time series features mentioned above are extracted to capture the short-term fluctuations and long-term trends of the air conditioning system.

[0008] Step 1.2.1: Extract operational volatility characteristics. This characteristic is used to quantify the smoothness of the air conditioning system's operation. This step assesses the stability of the system operation by calculating the volatility quantification index of one or more key thermodynamic parameters within a preset time window.

[0009] Step 1.2.2: Extracting Gradual Trend Features. To identify gradual faults, this step analyzes the direction and rate of change of one or more key performance indicators within a preset time window to extract trend-based quantitative indicators, thereby capturing the slow changes in the system.

[0010] Step 1.3: Construct Enhanced Feature Samples. The continuous learning system concatenates the extracted physical and temporal features with the original data vector to form the final enhanced feature samples. This sample will then be fed into the subsequent learning process: (3); in, This is the set of thermodynamic features extracted in step 1.1. This refers to the set of dynamic characteristic features extracted in step 1.2.

[0011] Through this step, the raw data is given a clear physical meaning and dynamic evolution information.

[0012] The system counts data in batches. Whenever the cumulative number of batches processed reaches the set number, steps 2 and 3 are executed sequentially. Between two diagnostic cycles, the system directly enters step 4 to perform model update iteration.

[0013] Step 2: Periodically trigger dynamic diagnosis of knowledge gaps in the model. Establish a dynamic diagnostic mechanism to periodically and quantitatively assess the model's mastery of knowledge for each type of fault.

[0014] Step 2.1: Utilize Class Confusion Conduct a misdiagnosis risk assessment of the model. Statistical validation is used to determine if the set belongs to the fault category. The proportion of samples misclassified by the model into other categories is used for calculation. To highlight the risk, a higher penalty weight is assigned to cases misclassified as "healthy". ): (4); in, It is to verify centralized faults The total number of samples, It's a malfunction. Misclassified as a category The number of samples.

[0015] Step 2.2: Utilize the prediction confidence margin The decision ambiguity of the model is evaluated. This is measured by calculating the confidence margin of the model for correctly classified samples. The confidence margin is the difference between the probability of the model predicting the correct class and the probability of predicting the second-highest confidence class. The smaller this difference, the more ambiguous the decision boundary. (5); in, It is a correctly classified fault. The sample set, For the model to sample Predicted as the correct category The probability, add the minimum value to the denominator. To prevent division by zero, usually Less than 10 -6 .

[0016] Step 2.3: Utilize the average loss value Perform an overall learning difficulty assessment of the model. Calculate the number of faults in the model on the validation set. The average cross-entropy loss value on the samples: (6); in, Let cross-entropy be the loss function. For real labels, This is the model's predicted output.

[0017] The three indicators mentioned above are weighted and summed to form a comprehensive dynamic difficulty index. : (7); in, , , These are adjustable weighting coefficients, set according to the importance of each item in the actual case. The sum of the three coefficients is 1. The higher the score, the weaker the model's knowledge of that fault, and the more important it is to focus on learning that fault category during continuous learning.

[0018] Step 3: Automatically adjust the sampling ratio based on the learning difficulty. The continuous learning system calculates the various faults in Step 2. The index, via a temperature coefficient The Softmax function is used to obtain the dynamic sampling weights for each type of fault. : (8); The weight and difficulty This is proportional, allowing subsequent training resources to automatically focus on the model's knowledge weaknesses, i.e. Figure 4 The mechanism-targeted training library is shown.

[0019] Step 4: Construction of a three-part experience replay and hybrid batch architecture based on dynamic diagnosis. Construct a "three-part" experience replay architecture (e.g., ...). Figure 4 The architecture consists of three data pools with different functions but the same goal. Before each model iteration update, the continuous learning system extracts samples from these three pools according to a specific strategy and combines them into training data for continuous learning.

[0020] Step 4.1: Construction of a permanent baseline data pool. This pool is used to store representative historical fault data covering typical operating conditions. Once established, this data pool is not easily modified to combat "catastrophic forgetting." During continuous learning, the data pool is uniformly and randomly sampled to ensure that various typical fault conditions are continuously reviewed.

[0021] Step 4.2: Construction of the Dynamic Frontier Knowledge Data Pool. This data pool is constructed using a first-in, first-out sliding window buffering method, receiving and caching the latest enhanced feature samples processed in Step 1 in real time. To accommodate "data distribution drift," the data pool is uniformly and randomly sampled during continuous learning to ensure that the latest knowledge is learned in a balanced manner.

[0022] Step 4.3: Mechanism-Targeted Enhancement Data Pool Construction. This data pool achieves "precise and efficient gap filling," storing the most easily confused fault samples. The sampling method for this pool is dynamic weighted sampling, with the weights determined in Step 3 based on... Dynamic sampling weights for score calculation The higher the weight, the greater the sampling probability.

[0023] Step 4.4: Constructing Hybrid Continuous Learning Data. At the beginning of each continuous learning iteration, continuous learning training data is sampled proportionally from the three data pools mentioned above.

[0024] Step 5: Parametrically Efficient Model Forward Propagation and Adapter Activation. The mixed data constructed in Step 4 is input into a parametrically efficient diagnostic model for forward propagation (e.g., ...). Figure 3 (As shown). The diagnostic model comprises a backbone network with frozen parameters, whose... Layer output features are The model also incorporates a parameter-trainable adapter module in parallel. During forward propagation, its feature update method is as follows: (9); in, For general-purpose adapters, To match the difficulty of learning Bound expert adapter, This is the activation function. To activate the switch, when a determination is made When the most urgent need is to optimize the fault, It is 1 if it is true, otherwise it is 0.

[0025] Step 6: Calculate the composite loss for continuous learning. Calculate the total loss using the composite loss function on the output of the model's forward propagation. : (10); in, The classification loss is calculated over the entire mixed training data; The knowledge distillation loss is calculated only on historical data samples in the mixed data. Its weighting coefficient. By minimizing the current model ( ) and a periodically updated model ( The output logits() between , The KL divergence is used to obtain: (11); Step 7: Efficient backpropagation and model parameter update. Based on the total loss calculated in Step 6... Backpropagation is then performed. When updating model parameters, the backbone network parameters remain frozen; only the parameters of the general adapter and the expert adapter activated in step 5 are optimized and updated. After the update is complete, the continuous learning system returns to step 1 and repeats the process. Figure 2 As shown, continuous online incremental learning is carried out.

[0026] Furthermore, the present invention also provides a cloud architecture system that implements the above method. The system includes an API interface for receiving real-time data, a task queue for caching samples to be processed, and a background training thread independent of the API response, thereby ensuring the real-time performance and efficiency of the entire continuous incremental learning process.

[0027] The method of this invention first collects central air conditioning operation data online and constructs information-enhanced feature samples by integrating thermodynamic mechanisms and dynamic time series analysis. Then, it periodically diagnoses and quantifies the model's mastery of various fault knowledge from three dimensions: category confusion, decision ambiguity, and learning loss, identifying knowledge weaknesses. Furthermore, it adaptively adjusts the sampling weights based on the diagnostic results and adopts a three-body experience replay strategy that includes permanent benchmarks, dynamic frontiers, and mechanism targeting to construct a hybrid training batch that takes into account both new and old knowledge and the reinforcement of weaknesses. Finally, it updates the model through parameter-efficient adapter fine-tuning technology and a composite loss function that takes into account both classification and knowledge distillation.

[0028] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows: 1) Strong adaptability and effective response to "data drift": This invention constructs a dynamic frontier knowledge region to continuously integrate the latest operating condition data into the training process. Combined with enhanced feature engineering that can capture system dynamics and gradual trends, the diagnostic model can automatically adapt to changes in operating conditions caused by seasonal changes, equipment aging, etc., avoiding the performance degradation problem caused by inconsistent data distribution in traditional static models. 2) Outstanding resistance to forgetting and overcoming “catastrophic forgetting”: This invention designs a three-body experience replay architecture that includes a permanent benchmark knowledge base, and introduces knowledge distillation loss oriented towards historical samples when calculating the composite loss. This forces the model to continuously consolidate its memory of historical typical failures while learning new knowledge, effectively solving the problem of catastrophic forgetting in incremental learning. 3) High learning efficiency and "precise gap filling": This invention designs a dynamic diagnosis (DDQ) mechanism for model knowledge weaknesses, which can periodically and multidimensionally quantify the model's mastery of each fault and generate targeted sampling weights accordingly. This allows training resources to be intelligently and selectively invested in the parts of the model that need the most strengthening, greatly improving learning efficiency and the final diagnostic accuracy of the model. 4) Low online deployment cost and high practicality: This invention adopts an efficient learning strategy of freezing the backbone network and only fine-tuning the parameters of the lightweight adapter module, which significantly reduces the computational resource overhead required for continuous online training; combined with the asynchronous cloud system architecture, this invention is easy to deploy at low cost in actual engineering projects and has strong engineering application value. Attached Figure Description

[0029] Figure 1 This is an overall architecture diagram of the system in an embodiment of the invention; Figure 2 This is a flowchart of the incremental learning method in an embodiment of the invention; Figure 3 This is a schematic diagram of the structure of the parameter-efficient model in the embodiment of the invention; Figure 4 This is an architectural diagram of the three-body experience playback data in an embodiment of the invention; Figure 5 This is the diagnostic performance confusion matrix on the old knowledge dataset (C1) in this embodiment of the invention; Figure 6 In this embodiment of the invention, the diagnostic performance confusion matrix on the new knowledge dataset (C2) is shown. Figure 7 This is a graph showing the changing trend of the overall diagnostic accuracy of the new and old knowledge datasets in an embodiment of the invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0032] The system is first initialized. When the continuous incremental learning system of this invention is started, the background service performs a series of initialization operations to prepare for subsequent online learning.

[0033] The system's hardware and software environment and modules are loaded. In this implementation case, the system is deployed on a cloud server running Linux, using Python 3.10 and the PyTorch 2.60 deep learning framework. (Refer to...) Figure 1 When the system starts, it loads and instantiates the core modules, including: Data Interface Module (API): Starts a web service using Python's Flask web framework, listens on a preset port, and opens an HTTP POST interface to receive real-time data in JSON format from the building's central air conditioning system.

[0034] Task buffer module (task queue): Instantiates a thread-safe object in program memory as an asynchronous decoupled buffer pool for front-end data reception and back-end model training.

[0035] Background processing module (background training thread): Creates an independent thread that is the "control center" of the system. After initialization, it will enter an infinite loop to wait for and process the data in the task queue.

[0036] Next, the model is loaded and prepared. The cloud system loads the current diagnostic model from the local path. The cloud system will create two instances of this model: Student Model: A model instance used to continuously receive new knowledge online and update its parameters.

[0037] Teacher model: A periodically fixed copy of the student model whose parameters remain stable during knowledge distillation, serving as a reference benchmark for the student model to retain old knowledge.

[0038] Finally, initialize the experience replay pool. (Refer to...) Figure 4 To implement the core "three-body" experience replay strategy of this invention, the system initializes three independent double-ended queues in memory: Permanent Baseline Data Pool: This knowledge base is populated by reading a predetermined number of samples covering all known fault types from a historical baseline dataset. The pool has a fixed capacity and its contents remain unchanged during runtime to combat catastrophic forgetting.

[0039] Dynamic Frontier Knowledge Data Pool: Initialized as an empty queue with a fixed capacity limit of 2048, it will adopt a first-in-first-out strategy to store the latest runtime data.

[0040] Mechanism-targeted data pool: Initialized as a large-capacity empty queue to store historical data and to cooperate with subsequent weighted sampling strategies for precise "gap-filling" training.

[0041] like Figure 2 As shown, the specific steps of the incremental learning method are as follows: Step 1: Online Data Acquisition and Feature Engineering Data reception and queuing. When the building's central air conditioning system collects new operational data, it encapsulates it in JSON format and sends it to the initiated API interface via an HTTP POST request. Upon receiving the data, the API interface directly adds it to the task queue and immediately returns a success response.

[0042] Step 1.1: Extract thermodynamic features reflecting the system state. The background training thread retrieves a segment of raw data from the task queue. A two-dimensional array of shape [15, 19] represents 19 sensors at 15 time steps. This array is then subjected to feature enhancement operations to generate more informative enhanced feature samples. The specific steps are as follows: For each time step, based on the intake and exhaust temperatures and pressures at that moment, the thermodynamic property calculation library CoolProp is called to calculate the superheat (SH) and subcooling (SC) using formulas (1) and (2).

[0043] Step 1.2: Extract time-series features that reflect the dynamic characteristics of the air conditioning system.

[0044] Step 1.2.1: Calculate the fluctuation characteristics: Select the exhaust pressure As the key parameter, calculate its standard deviation over 15 time steps. As shown in formula (12), it is used to quantify the stability of system operation: (12); in, This represents the average exhaust pressure within that time window. A higher standard deviation indicates that the air conditioning system is operating unstably.

[0045] Step 1.2.2: Calculate trend characteristics: Select compressor power consumption Using energy efficiency as the core indicator, a univariate linear regression was performed on data from 15 time steps to calculate its slope. As shown in formula (13): (13); in, Within the time window The compressor power consumption at each sampling point. A small, consistently positive slope ( A value greater than 0 may indicate that the efficiency of the air conditioning system is gradually deteriorating, and can provide an early warning in the early stages of a fault, which cannot be detected by fluctuation characteristics alone.

[0046] Step 1.3: Constructing Enhanced Feature Samples: The four newly extracted features are concatenated with the original 19 features, as shown in formula (3), to form an enhanced feature sample with the shape [15,23]. : ; Enhance feature samples Stored in the dynamic frontier knowledge data pool and the mechanism-targeted data pool.

[0047] The system counts data processed in batches. Whenever 100 batches are processed, a model knowledge gap diagnosis process is periodically triggered, executing steps 2 and 3 sequentially. Between two diagnosis cycles, the system skips these two steps and proceeds directly to step 4 to perform the regular model update iteration.

[0048] Step 2: Dynamic Diagnosis of Weak Points in Model Knowledge. This step involves a comprehensive evaluation of the current student model every 100 integer cycles.

[0049] Evaluation is performed on a validation set. The system uses a fixed validation set reserved during initialization. Perform a complete prediction on the student model and record the true label of each validation sample. Model predicted labels Prediction probability vector and cross-entropy loss value .

[0050] Step 2.1: Assess the risk of misdiagnosis using class confusion. This involves analyzing each fault mechanism present in the dataset. And it is quantitatively evaluated from the following three complementary perspectives.

[0051] In the evaluation of this implementation case, the validation set contained 100 fault samples of "refrigerant leakage concurrently with condenser scaling". After model diagnosis, 85 were correctly identified, but 10 were misclassified as healthy, and another 5 were misclassified as another concurrent fault, "refrigerant leakage concurrently with evaporator scaling". Risk weights were set. =2.0, other misjudgment weights =1.0, then the confusion level of this fault is: ; in, This represents the number of samples that were mistakenly identified as healthy. This represents the number of samples misclassified as other categories. That is the total number of samples.

[0052] Step 2.2: Calculate the decision ambiguity assessment (confidence margin) : Calculated according to formula (5). This index is used to measure the model's confidence in its correct decision; the smaller the difference, the more ambiguous the decision boundary. Still taking the fault of "refrigerant leakage and condenser scaling" as an example ( For example, suppose one sample in the validation set is correctly identified, and the model predicts it as the correct class ( The probability of ) The probability of a sample being predicted as having a confidence level of 0.85, and the probability of being predicted as belonging to the second-highest confidence category (such as "refrigerant leak") being 0.10, is therefore the confidence margin for this sample is (0.85 − 0.10) = 0.75. The final... It is the average of the inverse of the confidence margin of all correctly classified failure samples of this type. The larger the value, the more ambiguous the overall decision is.

[0053] Step 2.3: Calculate the overall learning difficulty assessment (average loss) ): Calculated according to formula (6). This index directly calculates all faults. The average cross-entropy loss value of the samples intuitively reflects the model's overall cognitive bias regarding the fault.

[0054] Finally, the final difficulty index is synthesized by weighting and summing the scores of the three dimensions mentioned above using formula (7) to obtain the final difficulty index for each type of fault. Generate a comprehensive, quantitative, dynamic difficulty index. .

[0055] Step 3: Automatically adjust the sampling ratio based on the learning difficulty. The system will then use the data generated in Step 2.3. The set of fractions is transformed into a probability distribution using a Softmax function with a temperature coefficient T (Equation (8)). This distribution represents the dynamic sampling weights for various fault types. .

[0056] Step 4: Three-Body Experience Replay and Hybrid Batch Construction Based on Dynamic Diagnosis. Constructing the hybrid data batch. At the beginning of each training iteration, the system draws samples from the "three-body" experience replay pool according to a preset ratio (e.g., 1:1:2) to construct a final hybrid data batch. For example, for a batch of size 64: Sixteen samples were randomly selected uniformly from the permanent benchmark knowledge base to review old knowledge.

[0057] Sixteen samples were randomly selected uniformly from the dynamic frontier knowledge region for learning new knowledge.

[0058] The sampling weights calculated in step 3 are derived from the mechanism-targeted training library. 32 samples were randomly selected using a weighted average method to "fill in the gaps".

[0059] Step 5: Model forward propagation and adapter activation.

[0060] Input Preparation: A mixed data batch of size 64, constructed according to the 1:1:2 ratio in step 4 (i.e., 16 baseline samples, 16 new samples, and 32 target samples), is fed into the student model. Taking a sample from the r_c_e.xls file in this batch as an example, this sample represents a triple composite fault of "leakage + condenser + evaporator," and its feature vector after feature engineering in step 1 is denoted as... .

[0061] Features through the backbone network: the feature vector First, a residual module from a pre-trained backbone network based on a CNN with its parameters frozen is used.

[0062] Parallel feature processing: The mixed data batches constructed in step 4 are input into the student model for forward propagation. (Refer to...) Figure 3 And Equation (9), characteristics of the intermediate layer of the backbone network It will flow through three paths simultaneously: 1) Path 1 (Direct Connection): The core features are passed directly through an identity mapping to ensure lossless inheritance.

[0063] 2) Path Two (Universal Adapter): Sent into the general adapter module In this embodiment, It is a micro-LP consisting of two fully connected layers with 16 hidden layers and ReLU as the activation function. It is responsible for extracting universal features common to all fault types.

[0064] 3) Path Three (Expert Adapter Activation): Based on the DDQ evaluation results in step 2, the system has determined the dynamic difficulty index of the "r_c_e" triple fault. The highest among all current fault types. Therefore, the system activates the switch. =1, precisely activating the expert adapter module associated with the fault. The structure of this module is similar to... While identical, their parameters are independently trained to learn the unique characteristics of the specific complex condition "r_c_e". All other samples in the batch, corresponding to different faults, will activate their respective expert adapters. Step 6: Calculate the composite loss function. To achieve the "connection between previous and subsequent knowledge," the system employs a composite loss function. Optimization is performed as shown in formula (10): Classification loss For all 64 samples in the batch, the final predicted output is compared with the true label, and the standard cross-entropy loss is calculated.

[0065] Knowledge distillation loss The knowledge distillation loss is calculated only for 48 historical data samples from the "Permanent Benchmark Knowledge Base" and "Mechanism-Targeted Training Library" in the batch. In this embodiment, the temperature coefficient T for knowledge distillation is set to 2.0. The student model is constrained to retain old knowledge by calculating the KL divergence of the logits of the current student model and the teacher model's outputs on these 48 samples.

[0066] Total loss: According to formula (10), the two losses above are weighted and added together (in this embodiment, the weighting coefficients are...). (Setting it to 1.0), we obtain the total loss. .

[0067] Step 7: Based on the calculated total loss Backpropagation is performed. In this embodiment, the Adam optimizer is configured with a learning rate of 1e-4. Since the backbone network parameters are frozen, gradient calculation and parameter updates will only apply to the universal adapter shared by all samples. And the individual expert adapters activated in this batch (such as those in the previous examples). On the parameters of ).

[0068] The system includes a batch counter. In this embodiment, the threshold of the counter is set to 500. After the system completes 500 batches of training updates, the program will perform a weight copy operation, deeply copying the current parameters of all adapter modules (including general and all expert adapters) in the student model to the corresponding modules in the teacher model, thereby completing the iteration of the teacher model.

[0069] After completing a full parameter update, the background training thread returns to step 1, continuously processes new data, and repeats the closed-loop process of "self-diagnosis, targeted learning, and reviewing the old to learn the new" mentioned above, thereby achieving long-term, stable and efficient fault diagnosis in a dynamically changing environment.

[0070] To verify the effectiveness of the method of the present invention, incremental learning experiments were conducted. Figure 5 and Figure 6 The final diagnostic performance of this method on the old knowledge dataset (C1) and the new knowledge dataset (C2) after the incremental learning task is presented. The confusion matrix shows that the model achieves extremely high diagnostic accuracy on both the old and new data, proving that this method effectively avoids knowledge forgetting while learning new knowledge. Figure 7The graph shows that throughout the incremental learning process, the model's diagnostic accuracy on the C1 dataset (solid blue line) remained consistently high, while its diagnostic accuracy on the C2 dataset (dashed orange line) steadily improved. This trend chart strongly demonstrates that the method proposed in this invention possesses continuous, stable, and efficient incremental learning capabilities.

[0071] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A central air conditioning fault diagnosis persistent incremental learning method against forgetting, characterized in that, The method comprises the following steps: Step 1, collecting the operation data of the central air conditioning system and converting it into enhanced feature samples; Counting in units of processed data batches, and when the cumulative processing reaches the set batch number, sequentially executing steps 2 and 3, and between the two diagnosis cycles, the system directly enters step 4 to perform model update iteration; Step 2, calculate the comprehensive dynamic difficulty index of each fault ; Step 3, generating ATB target sampling weights based on individual failures , generating ATB target sampling weights; Step 4, constructing a mixed data batch to generate training data; Step 5, selectively activating the adapter and inputting the training data into the diagnostic model for forward propagation; Step 6, the total loss is calculated by the composite loss function on the output result of the model forward propagation ; Step 7, perform backpropagation based on the total loss calculated in step 6 ; and perform an optimized update of the parameters of the adapter After the update is completed, the system returns to step 1 and continues online incremental learning. 2.The anti-forgetting central air conditioning fault diagnosis persistent incremental learning method according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Extracting thermodynamic characteristics reflecting the state of the central air conditioning system: From the collected raw operating data of the central air conditioning system, the system extracts thermodynamic characteristics based on each time point. Temperature and pressure measurements are used to calculate physical characteristics related to the refrigeration cycle state: heat. With supercooling ; Step 1.2, extracting time-series features reflecting dynamic characteristics of the air conditioning system: extracting time-series features from a time window containing past sampling points to capture short-term fluctuations and long-term trends of the air conditioning system; Specifically as follows: Step 1.2.1, extracting running volatility features: by calculating one or more thermal parameter volatility quantification indicators within a preset time window, the stability of system operation is evaluated; Step 1.2.2, extracting gradual trend features: by analyzing the change direction and rate of one or more performance indicators within a preset time window, trend quantification indicators are extracted to capture system changes; Step 1.3, Constructing enhanced feature samples: The extracted physical features and timing features are concatenated with the original data vector to form enhanced feature samples , enhanced feature samples will be fed into the subsequent learning process. 3.The anti-forgetting central air conditioning fault diagnosis persistent incremental learning method according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1, Calculate the class confusion degree by the proportion of samples belonging to the fault in the statistical verification set that are misjudged as other categories by the model , to evaluate the misdiagnosis risk of the model ; Step 2.

2. Confidence margin of correctly classified samples by the computational model Performing decision ambiguity assessment of the model; Step 2.3, the average cross-entropy loss value of the model on all samples belonging to the fault in the validation set Step 2.4, the overall learning difficulty evaluation of the model Step 2.5, the overall learning difficulty evaluation of the model Step 2.4, class confusion confidence margin and average cross-entropy loss value perform weighted sum to form comprehensive dynamic difficulty index . 4.The anti-forgetting central air conditioning fault diagnosis persistent incremental learning method according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1, constructing a permanent reference data pool: the permanent reference data pool is used to store historical working condition fault data, and the permanent reference data pool adopts uniform random sampling; Step 4.2, dynamic front knowledge data pool construction: the dynamic front knowledge data pool is constructed in a first-in first-out sliding window buffer manner, and enhanced feature samples are received and cached in real time The dynamic front knowledge data pool adopts uniform random sampling; Step 4.3, Mechanism-targeted hardening data pool construction: The mechanism-targeted hardening data pool stores the most confusing fault samples, and the mechanism-targeted hardening data pool adopts dynamic weighted sampling, and the weight is a dynamic sampling weight ; Step 4.4, mixed continuous learning data construction: at the beginning of each continuous learning iteration, continuously learning training data is sampled from the permanent reference data pool, dynamic frontier knowledge data pool and mechanism targeted reinforcement data pool in equal proportions.

5. The incremental learning method of the anti-forgetting central air conditioning fault diagnosis persistent incremental learning system according to any one of claims 1-4, characterized in that, Based on the continuous learning system implementation, the system integrates model training difficulty evaluation mechanism and fault self-adaptive playback strategy, and can receive new working condition data during the cloud operation of the model, while maintaining the memory of historical knowledge; It includes: Data interface module: used to receive JSON format real-time data from building central air conditioning system; Task buffer module: instantiates a thread-safe object in program memory as an asynchronous decoupling buffer pool for front-end data reception and back-end model training; Background processing module: create a separate thread that will enter an infinite loop after initialization to wait and process data in the task queue; Data storage module: used to store samples covering all known fault types, the latest operation data and historical data; Model management module: including a student model for online continuous reception of new knowledge and parameter update, and a teacher model as a reference benchmark for the student model to retain old knowledge.

Images