Air conditioner quality inspection model training method and device and medium

Abstract

This disclosure relates to a training method, apparatus, and medium for an air conditioner quality inspection model, belonging to the field of air conditioning technology. The method includes: acquiring air conditioner production process parameters, operational performance indicators, and environmental adaptability data; preprocessing the data to obtain standard feature data; training a gradient boosting regression tree model using a training set; using the trained gradient boosting regression tree model to predict the validation set and determine the threshold for judging abnormal air conditioners; inputting the test set into the trained gradient boosting regression tree model to obtain the mean squared error for each air conditioner; comparing the mean squared error of each air conditioner with the abnormal air conditioner judgment threshold to determine whether the current air conditioner is normal; and calculating the precision, recall, and F1 score of the boosting regression tree model. With the help of this solution, the defect rate of air conditioners can be effectively reduced.

Description

Technical Field

[0001] This disclosure belongs to the field of air conditioning technology, and specifically relates to an air conditioning quality inspection model training method, device and medium. Background Technology

[0002] Currently, air conditioner manufacturers often employ quality inspection and analysis methods based on fixed standards. These methods involve manual or semi-automated equipment performing sampling or full inspection of individual machines, essentially a "passive" quality screening approach. A defect detected in this way means the cost of rework or scrapping has already been incurred. While this approach has low technical barriers and doesn't require significant investment in sensors and data processing resources, it can meet practical needs when production batches are small or product models are limited. However, its over-reliance on "human experience + fixed thresholds" means it can only detect preset defect types, such as leak detection, performance, noise, and safety testing, and cannot uncover hidden problems caused by combinations of parameters. Furthermore, once a batch defect occurs, it's difficult to quickly trace whether the cause is a material quality issue, an assembly process problem, or a deviation by testing personnel or equipment. Summary of the Invention

[0003] This disclosure proposes a training method, device, and medium for an air conditioner quality inspection model, aiming to solve the problem of high product defect rates caused by "post-processing" product quality inspection.

[0004] According to a first aspect of this disclosure, a method for training an air conditioner quality inspection model is provided. The method includes: acquiring production process parameters, operational performance indicators, and environmental adaptability data of air conditioners; preprocessing the production process parameters, operational performance indicators, and environmental adaptability data to obtain standard feature data, wherein the standard feature data is divided into a training set, a validation set, and a test set. The training set consists of quality inspection data of normal air conditioners, the validation set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners, and the test set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners; training a gradient boosting regression tree model using the training set; predicting the validation set using the trained gradient boosting regression tree model to determine the abnormal air conditioner judgment threshold; inputting the test set into the trained gradient boosting regression tree model to obtain the mean square error of each air conditioner; comparing the mean square error of each air conditioner with the abnormal air conditioner judgment threshold to determine whether the current air conditioner is normal; and calculating the precision, recall, and F1 score of the boosting regression tree model based on the number of truly abnormal air conditioners after manual re-inspection, the number of misjudged normal air conditioners, and the number of abnormal air conditioners missed in the test set, thereby completing the training of the air conditioner quality inspection model.

[0005] In some embodiments, the preprocessing of production process parameters, operational performance indicators, and environmental adaptability data to obtain high-quality feature data includes: associating production process parameters, operational performance indicators, and environmental adaptability data through the unique serial number of the air conditioner to obtain an air conditioner data archive sample, wherein the air conditioner data archive sample includes the unique serial number of the air conditioner and multiple feature variables; when the missing rate of the feature variables of the air conditioner data archive sample is less than or equal to the missing threshold, the mean or mode of the feature is used to fill the gaps using other complete samples of the same model and under the same operating conditions; when the missing rate of the feature variables of the air conditioner data archive sample is greater than the missing threshold, the entire air conditioner data archive sample is removed.

[0006] In some embodiments, if a certain characteristic of an air conditioning data archive sample is significantly outside the normal range or physical limits, it is corrected using the median of that characteristic from other valid samples in the same batch.

[0007] In some embodiments, if the value of the feature variable is numerical, the value of the feature variable is converted into a distribution with a mean of 0 and a standard deviation of 1.

[0008] In some embodiments, derived features are constructed based on feature variables.

[0009] In some embodiments, when the feature variable is a category, the category value is one-hot encoded.

[0010] In some embodiments, determining the abnormal air conditioner judgment threshold includes: determining the mean square error of each air conditioner; calculating the mean and standard deviation based on the mean square error; and determining the abnormal air conditioner judgment threshold based on the mean and standard deviation.

[0011] In some embodiments, determining the abnormal air conditioner judgment threshold based on the mean and standard deviation includes:

[0012] According to the formula: The threshold for determining abnormal air conditioning was calculated. ;

[0013] in, This represents the mean. It represents the standard deviation.

[0014] In some embodiments, after determining whether the current air conditioner is normal, the method further includes: using SHAP analysis to diagnose the abnormal air conditioner and determine the root cause of the air conditioner malfunction.

[0015] In some embodiments, before training the air conditioner quality inspection model, the method further includes: adding the quality inspection data of the misjudged normal air conditioners to the training set, and using the new training set to train the gradient boosting regression tree model; using the trained gradient boosting regression tree model to predict the new validation set, redetermining the threshold for judging abnormal air conditioners, and continuously performing model optimization iterations.

[0016] According to a second aspect of this disclosure, an air conditioner quality inspection model training device is provided, comprising: an acquisition module for acquiring production process parameters, operational performance indicators, and environmental adaptability data; a preprocessing module for preprocessing the production process parameters, operational performance indicators, and environmental adaptability data to obtain high-quality feature data, wherein the high-quality feature data is divided into a training set, a validation set, and a test set, wherein the training set consists of quality inspection data of normal air conditioners, the validation set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners, and the test set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners; and a training module for training gradients using the training set. The system comprises the following modules: a boosted regression tree model; a prediction module, used to predict the validation set using the trained boosted regression tree model to determine the threshold for identifying abnormal air conditioners; an input module, used to input the test set into the trained boosted regression tree model to obtain the mean squared error for each air conditioner; a comparison module, used to compare the mean squared error of each air conditioner with the threshold for identifying abnormal air conditioners to determine whether the current air conditioner is normal; and a calculation module, used to calculate the precision, recall, and F1 score of the boosted regression tree model based on the number of truly abnormal air conditioners after manual re-inspection, the number of normally normal air conditioners misjudged, and the number of abnormal air conditioners missed in the test set, thus completing the training of the air conditioner quality inspection model.

[0017] According to a third aspect of this disclosure, an air conditioning quality inspection model training device is provided, comprising: a memory; and

[0018] A processor coupled to the memory is configured to execute the air conditioning quality inspection model training method described above based on instructions stored in the memory.

[0019] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the air conditioning quality inspection model training method as described above.

[0020] The advantages of this solution are: by using deep learning on product quality inspection datasets, it gains the ability to automatically intercept suspected defective products, and through early warning by the model, closed-loop feedback and adjustment of production parameters, it transforms "post-event processing" into "process prevention", significantly reducing the defect rate; it reveals the complex and implicit correlations between parameters, and digs out deep-seated patterns, thereby quickly locating the root cause of batch defects; the model learns and optimizes by learning from new data, and continuously iterates to adapt to new process and material changes, keeping the detection accuracy within a stable range. Attached Figure Description

[0021] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure.

[0022] This disclosure can be more clearly understood with reference to the accompanying drawings and the following detailed description.

[0023] Figure 1 This is a flowchart illustrating a method for training an air conditioning quality inspection model according to some embodiments of the present disclosure.

[0024] Figure 2 This is a block diagram illustrating an air conditioning quality inspection model training apparatus according to some embodiments of the present disclosure.

[0025] Figure 3 This is a block diagram illustrating an air conditioning quality inspection model training apparatus according to other embodiments of the present disclosure.

[0026] Figure 4 This is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure. Detailed Implementation

[0027] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.

[0028] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0029] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or utility.

[0030] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0031] In all the examples shown and discussed herein, any specific values ​​should be understood as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0032] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0033] like Figure 1 As shown, the training method for the air conditioning quality inspection model includes steps S110 to S170.

[0034] In step S110, the production process parameters, operating performance indicators, and environmental adaptability data of the air conditioner are obtained.

[0035] We selected the product lifecycle indicators that best reflect the core performance of air conditioners and are easily affected by the production process, covering three types of quality inspection data: production process, motion performance, and environmental adaptability. Based on the quality inspection results or historical fault records, we clearly labeled the samples as "normal" or "abnormal".

[0036] The production process parameters of each product (including indoor and outdoor units) are obtained from the MES (Manufacturing Execution System), which is the product's "genetic data".

[0037] Based on end-of-line inspection data, combined with commodity inspection and sampling inspection processes, operational performance indicators are obtained, which constitute the product's "health check data".

[0038] Multi-condition environmental adaptability data are obtained from laboratories using enthalpy difference, noise reduction, and EMI testing; this is the product's "strain data".

[0039] By using the unique serial number (SN) of the air conditioner product, the three types of data are linked to construct a complete data file for each product, which is represented as an n-dimensional data sample, where each dimension corresponds to a specific feature parameter.

[0040] In step S120, the production process parameters, operating performance indicators, and environmental adaptability data are preprocessed to obtain standard feature data. The standard feature data is divided into a training set, a validation set, and a test set. The training set consists of quality inspection data of normal air conditioners, the validation set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners, and the test set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners.

[0041] For each feature parameter of all samples, calculate its mean, variance, standard deviation, maximum value, minimum value, median, and mode.

[0042] If a certain feature variable of a sample has missing values, and the missing rate is ≤10%, then the mean (for numerical indicators) or mode (for other categories of indicators) of the same complete sample under the same model and working conditions will be used to fill the missing value; if the missing rate is >10%, then the entire sample will be removed.

[0043] If a certain indicator of a sample significantly exceeds the normal range or physical limit, it is corrected by using the median of that indicator from other valid samples in the same batch.

[0044] To eliminate the influence of dimensions, numerical data is standardized to ensure that different feature values ​​are on the same order of magnitude, transforming the feature parameters into a distribution with a mean of 0 and a standard deviation of 1. ;

[0045] in These are the original feature values ​​of a sample. This is the mean of the feature across all valid samples after cleaning. This represents the standard deviation of the feature.

[0046] To enhance the model's sensitivity to anomalous patterns and its ability to identify normal patterns, the following measures were taken: Derivative features are constructed to extract features that reflect the performance status of the air conditioner, including cooling and heating efficiency to determine cooling and heating capacity, pressure attenuation rate to determine whether there is micro-leakage, and performance attenuation rate of the same machine under different operating conditions.

[0047] Categorical features are encoded. Each categorical value of a categorical feature is converted into a unique binary vector, enabling the model to effectively utilize categorical features for training and prediction.

[0048] In step S130, the gradient boosting regression tree model is trained using the training set.

[0049] Define a one-dimensional prediction result Yi∈{0,1}. If the sample prediction result is normal, then Yi=0; otherwise, =1.

[0050] Construct three sample datasets. Divide the data into a training set (containing only normal samples, used to train the model to learn normal operating patterns), a validation set (containing known normal and abnormal samples, used for parameter tuning and selection of anomaly detection thresholds), and a test set (composed of test samples that mix normal and abnormal samples, used to evaluate the model's ability to detect unknown anomalies), ensuring that the model only learns "normal patterns" and avoids interference from abnormal data.

[0051] Initialize the model parameters.

[0052] Objective: The regression target is the mean squared error (MSE).

[0053] Learning_Rate: Used to control the learning step size; a higher value results in faster learning but may miss the optimal solution; a lower value requires more trees to converge but usually yields a model with better generalization ability.

[0054] Max_Depth: The maximum depth of the decision tree, used to control the complexity of the model; a smaller depth results in a simpler model that is less prone to overfitting, but may have insufficient learning ability; a larger depth results in strong learning ability, but is more likely to capture noise and special patterns, leading to overfitting.

[0055] N_Estimators: The total number of trees to be built; too few may result in insufficient learning; too many may lead to excessively long training times and potential overfitting.

[0056] Colsample: Feature sampling rate, the percentage of features used per tree.

[0057] Subsample: The sampling rate of training samples, used to improve generalization ability.

[0058] First, train the regression model using a training set containing only normal samples, starting from the input. Predicting targets .

[0059] In step S140, the trained gradient boosting regression tree model is used to predict the validation set and determine the threshold for abnormal air conditioning judgment.

[0060] The trained model is then used to predict the entire validation set (including normal and abnormal samples) to obtain the MSE for each sample. By fitting the input-output relationship of normal samples, the pattern of normal products is learned.

[0061] For each feature value of each sample in the validation set, calculate the reconstruction error between the predicted value and the true value. Then calculate the MSE of the entire sample: For normal samples, the model can map the feature values ​​of the samples well, and the reconstruction error is small; while for abnormal samples, because they deviate from the normal pattern, the reconstruction error is large.

[0062] Set an error threshold. Perform statistical analysis based on the MSE of all samples in the training and validation sets: calculate the MSE distribution of all samples and obtain the mean. Standard deviation Given the true labels of each sample in the validation set, find the minimum MSE of the validation set. To the maximum value As a possible threshold range.

[0063] If MSE approximately follows a normal distribution, according to In principle, approximately 99.7% of normal samples have an MSE less than " "; Considering that the anomaly is manifested by a significantly large error, a one-sided confidence interval is adopted, and an anomaly judgment threshold can be set." If the MSE distribution deviates significantly from normality, then the empirical quantile method should be used instead, taking the 99th or 99.7th quantile as the threshold.

[0064] In step S150, the test set is input into the trained gradient boosting regression tree model to obtain the mean square error of each air conditioner.

[0065] In step S160, the mean square error of each air conditioner is compared with the abnormal air conditioner judgment threshold to determine whether the current air conditioner is normal.

[0066] Feature Importance Analysis: By using SHAP values ​​or the model's built-in feature importance parameter "feature_importances", key features affecting anomaly detection can be identified, helping to quickly locate the most likely source of the problem and see which dimensions contribute the most to the residuals, thereby guiding on-site investigation and quickly locating the key features that lead to high residuals.

[0067] Closed-loop feedback verification verifies the model's effectiveness, and feature importance analysis identifies the root cause of anomalies.

[0068] In step S170, based on the number of truly abnormal air conditioners after manual re-inspection, the number of normally normal air conditioners misjudged, and the number of abnormal air conditioners missed in the test set, the precision, recall, and F1 score of the improved regression tree model are calculated, and the training of the air conditioner quality inspection model is completed.

[0069] The identified defective machines are manually re-inspected, and if necessary, disassembled for on-site testing to confirm the fault, and the labels are corrected accordingly.

[0070] Based on the newly confirmed samples, incremental training is performed to update the model, and TH is recalibrated according to the error distribution of the updated normal samples.

[0071] Analyze the commonalities in the input characteristics of abnormal samples. Is there a general deviation from the standard for a certain assembly parameter? Is it concentrated on a certain production line, a certain shift, or a certain supplier's materials? Through multi-dimensional correlation analysis, we can help locate the potential root cause.

[0072] To quantify model performance, the results are evaluated using precision, recall, and their harmonic mean (F1).

[0073] Precision = TP / (TP + FP), a value of 1 indicates that all samples judged as anomalous are true anomalous.

[0074] Recall = TP / (TP + FN), a value of 1 indicates that all real anomalies were successfully detected;

[0075] F1 = 2 * (Precision * Recall) / (Precision + Recall), where 1 indicates that the model performs well on both.

[0076] In this model, TP stands for True Positive Instances, which is the number of correctly detected anomalous samples; FP stands for False Positive Instances, which is the number of normal samples incorrectly labeled as anomalous; and FN stands for False Negative Instances, which is the number of true anomalies missed. If the model frequently classifies normal products as anomalous (high FP), it will lead to unnecessary rework and increase quality inspection costs. In this case, the model should be optimized to ensure that samples classified as anomalous are likely to be true anomalies.

[0077] Model parameter optimization iteration: In order to cope with the slow drift in the production process (such as the replacement of key component suppliers, mold wear, process route adjustment, etc.), new normal samples confirmed by manual or high confidence mechanism are periodically included in the training set, and the model is retrained in combination with historical data (including representative abnormal samples).

[0078] Two embodiments are provided below to further illustrate the concept of this disclosure.

[0079] Example 1. A detailed explanation is provided using quality inspection data reported via an IoT module from a specific brand and model of a smart wall-mounted air conditioner over a specific time period (e.g., one week):

[0080] The following quality inspection indicators, which best reflect the core performance of air conditioners and are easily affected by the production process, were selected to form a dataset X: Production process parameters for each air conditioner were obtained from the MES, including sealing pressure value P_Seal (MPa), evaporator length deviation ΔL_Evap (mm), condenser width deviation ΔL_Cond (mm), and compressor assembly torque Torq1~Torq4 (N·m).

[0081] Based on the data from the whole machine end inspection line, combined with the commercial inspection and sampling inspection process, the operating performance indicators are obtained, including the compressor exhaust temperature, operating power, internal and external fan speeds, and inlet and outlet temperatures under rated operating conditions.

[0082] Multi-condition environmental adaptability data were obtained from the enthalpy difference chamber and the anechoic chamber, including the simulated annual comprehensive energy efficiency value (APF) calculated based on 4 cooling temperature points and 5 heating temperature points, the cooling energy efficiency ratio (EER_T1) and heating energy efficiency ratio (COP_T1) under rated operating condition T1, the high-speed noise (Noise_high) (dB), and the average power (Paver_T2) and (Paver_T3) (W) under low-temperature heating condition T2 and high-temperature cooling condition T3.

[0083] By using the unique serial number (SN) of the air conditioner product, the three types of data are linked to construct a complete data file for each product, which is represented as an n-dimensional data sample, where each dimension corresponds to a specific feature parameter.

[0084] Data cleaning and preprocessing. For each feature variable of all samples, calculate descriptive statistics such as mean, variance, standard deviation, maximum, minimum, median, and mode.

[0085] If a certain feature variable of a sample has missing values, and the missing rate is ≤10%, then the mean (for numerical indicators) or mode (for other category indicators) of the same complete sample under the same model and working conditions is used to fill the missing values; if the missing rate is >10%, then the sample is removed entirely.

[0086] If a certain indicator of a sample significantly exceeds the normal range or physical limit, it is corrected by using the median of that indicator from other valid samples in the same batch. For example, if a sample's R_Fan is 1950 rpm (far exceeding the normal range of 850-1100 rpm), it is determined to be a measurement error, and the median R_Fan of the samples in the same batch (900 rpm) is used instead.

[0087] To eliminate the influence of dimensions, numerical data is standardized to ensure that different feature values ​​are on the same order of magnitude, transforming the feature parameters into a distribution with a mean of 0 and a standard deviation of 1. ,in These are the original feature values ​​of a sample. This is the mean of the feature across all valid samples after cleaning. This represents the standard deviation of the feature.

[0088] Taking the compressor discharge temperature T_Comp as an example, the mean of the normal sample Celsius For example, if the T_Comp of a sample is 58.8℃, after standardization, it is 2.58.

[0089] Derivative features are constructed for X to extract features reflecting the air conditioner's operating status, enhancing the model's sensitivity to abnormal modes and its ability to identify normal modes. Here, features or combinations of feature values ​​with specific significance are selected: cooling efficiency (EER) and heating efficiency (COP), pressure drop rate per unit time during pressure holding tests (Pres_Drop_Rate), energy efficiency degradation rate (EER_Drop_Rate) relative to rated operating conditions under high-temperature conditions, and evaporator and condenser installation dimensional errors (ΔL) reflecting assembly consistency. Where m is the air flow rate (Kg / s), Cp is the specific heat capacity (KJ / (Kg*K)), Pin is the input power (KW), and ΔT is the temperature difference between the inlet and outlet (°C). ; and These represent the initial and final pressures per unit time, respectively. ; and These represent the energy efficiency ratios under rated operating conditions and high-temperature operating conditions, respectively. ; and These represent the assembly dimensional deviations of the evaporator and condenser, respectively.

[0090] Encode categorical features. Each categorical value of a categorical feature is converted into a binary vector, enabling the model to effectively utilize categorical features for training and prediction. For example, work conditions T1, T2, and T3 are categorical features and can be encoded as [1,0,0], [0,1,0], [0,0,1], etc., respectively.

[0091] Define a one-dimensional prediction target If the sample prediction result is normal, then ;otherwise Note that certain operating indicators or steady-state performance parameters should not be used as features to predict another performance indicator. For example, using power and speed to predict compressor temperature will cause the model to "cheate." This is because power and speed are themselves indicators that need to be monitored; features should primarily come from the "cause" (such as the production process, assembly quality, and environmental conditions) rather than the "effect."

[0092] Three sample datasets are constructed. The data is divided into a training set (containing only normal samples, accounting for 80% of all normal samples, used to train the model to learn normal operating patterns), a validation set (containing the remaining normal samples and known abnormal samples, used for parameter tuning and selection of anomaly detection thresholds), and a test set (a mixture of normal and abnormal samples, used to evaluate the model's ability to detect unknown anomalies). This ensures the model only learns "normal patterns" and avoids interference from abnormal data. When abnormal samples are limited, normal samples are used for training, and abnormal samples are only used for validation and testing.

[0093] Initialize the following basic model parameters: Objective = 0; Learning_Rate = 0.1; Max_Depth = 3; N_Estimators = 100; Colsample = 0.8; Subsample = 0.8.

[0094] Training and Fitting. Install third-party library functions to build a basic model, and then train and fit it. Model Training: First, train the regression model using a training set containing only normal samples, starting from the input... Predicting targets Fitting: The trained model is then used to predict the entire validation set (including normal and abnormal samples) to obtain the MSE for each sample. By fitting the input-output relationship of normal samples, the pattern of normal products is learned.

[0095] For each feature value of each sample in the validation set, calculate the reconstruction error between the predicted value and the true value. Then calculate the MSE of the entire sample: For normal samples, the model can map the feature values ​​of the samples well and the reconstruction error is small; while for abnormal samples, because they deviate from the normal pattern, the reconstruction error will be larger.

[0096] The calculated MSE of the training set is 0.02, indicating that the model fits the normal samples well. The MSE of the normal samples in the validation set is 0.03, which is close to that of the training set, indicating that the model has good generalization ability and no obvious overfitting. The MSE of the abnormal samples in the validation set is 1.27, which is significantly higher than that of the normal samples, indicating that the model has a large prediction bias for abnormal samples and can be used for anomaly detection.

[0097] Set an error threshold. Calculate the MSE distribution of normal samples and obtain the mean. Standard deviation MSE approximately follows a normal distribution, according to In principle, approximately 99.7% of normal samples have an MSE less than [a certain value]. Considering that the abnormal behavior is characterized by significantly larger errors, a one-sided confidence interval is adopted, and the abnormal judgment threshold TH=1.87 is set. This threshold can ensure that the misclassification rate of normal samples is less than 0.3%.

[0098] Anomaly detection: Calculate the MSE of each sample in the test set. If the MSE of a machine is greater than TH, it is predicted as an "anomaly suspect machine". =1); predictions less than or equal to are normal ( ).

[0099] Feature Importance Analysis. SHAP values ​​identify key features that influence anomaly detection, helping to quickly pinpoint the most likely source of the problem. For example, if a batch of anomalous samples has a high SHAP value, the evaporator size should be checked first.

[0100] Closed-loop feedback verification validates model effectiveness and pinpoints the root cause of anomalies through feature importance analysis: Abnormal machines are manually re-inspected, and if necessary, disassembly and testing are conducted to confirm the fault, thus correcting the labels. Based on newly confirmed samples, incremental training is performed to update the model, and the TH (Head-of-Size) is recalibrated according to the error distribution of the updated normal samples. Commonalities in input features of abnormal samples are analyzed: Does a certain assembly parameter generally deviate from the standard? Is it concentrated on a specific production line, shift, or supplier's materials? Multi-dimensional correlation analysis helps pinpoint potential root causes. For example, normal samples have compressor temperature prediction errors ≤ ±1.5℃, stable sealing pressure at 2.5MPa, and evaporator dimensions within ±0.2mm. Abnormal samples, however, have prediction errors as high as 6℃, sealing pressure reduced to 2.2MPa, and evaporator size deviations of 0.5mm. Given the high importance of the "sealing pressure" feature, this suggests a potential assembly defect in the sealing system, requiring a focused investigation of related process steps.

[0101] To quantify model performance, results are evaluated using precision, recall, and their harmonic mean (F1): Precision = TP / (TP + FP), a value of 1 indicates that all samples judged as anomalous are true anomalous; Recall = TP / (TP + FN), a value of 1 indicates that all true anomalous samples were successfully detected; F1 = 2 * (Precision * Recall) / (Precision + Recall), a value of 1 indicates that the model performs well in both metrics.

[0102] In this model, TP stands for True Positive Cases, which is the number of correctly detected anomalous samples; FP stands for False Positive Cases, which is the number of normal samples incorrectly labeled as anomalous; and FN stands for False Negative Cases, which is the number of true anomalies missed. If the model frequently classifies normal products as anomalous (high FP), it will lead to unnecessary rework and increase quality inspection costs. In this case, the model should be optimized to improve accuracy and ensure that samples classified as anomalous are likely to be true anomalies.

[0103] During optimization, increasing the decision threshold to make the model more "cautious" can reduce false positives and improve precision, but it may also miss more real anomalies, leading to a decrease in recall. Conversely, decreasing the threshold to make the model more "sensitive" can capture more real anomalies and improve recall, but it may misclassify more normal samples, leading to a decrease in precision. Therefore, a comprehensive consideration is needed, using F1 to balance the two and avoid the bias caused by a single metric.

[0104] Model parameter optimization iteration. To cope with slow drift in the production process (such as replacement of key component suppliers, mold wear, process route adjustment, etc.), new normal samples confirmed by manual or high confidence mechanisms are periodically included in the training set, and the model is retrained by combining historical data (including representative abnormal samples).

[0105] Example 2

[0106] Assuming quality inspection data is collected from 10 air conditioners of the same model under specific operating conditions, the raw data is shown in the table below:

[0107]

[0108] The table contains missing values ​​(-) and obvious outliers (noise value 999).

[0109] Missing Value Handling: SN003's "Cooling Power" is missing: We need to calculate the average of this indicator for the other 9 units. Average = (3500+3550+3480+3520+3530+3490+3540+3510+3560) / 9 = 3523.3W. Therefore, fill the missing value of SN003 with 3523. SN007's "Pressure Value" is missing: The missing rate is 1 / 10 = 10%, and it is filled with the average value according to the same rules. Average of other effective pressure values ​​= (2.1+2.2+1.9+2.5+2.0+2.3+2.15+2.05+2.25) / 9 = 2.16MPa. Therefore, fill the missing value of SN007 with 2.16.

[0110] Outlier Correction. SN006's "Noise" value is 999dB: This is clearly an outlier exceeding physical limits (air conditioner noise shouldn't be that loud). Correction is made using the median of this indicator from other valid samples in the same batch. All noise values ​​are sorted: 43, 44, 44, 45, 45, 45, 46, 46, 47 → the median is 45dB. Therefore, SN006's noise value is corrected from 999 to 45.

[0111] Data standardization. We perform Z-score standardization on numerical features (such as cooling power and operating current) to eliminate the influence of dimensions. Taking "cooling power" as an example: calculate the mean (μ) and standard deviation (σ): μ≈3523, σ≈25.5. Standardize the cooling power (3500) of SN001: (3500 - 3523) / 25.5≈-0.90. Repeat this process for all numerical features, and finally each feature is transformed into a distribution with a mean of 0 and a standard deviation of 1.

[0112] After cleaning and preprocessing, the data becomes complete, clean, and of the same magnitude, laying a solid foundation for model training.

[0113] Derivative feature construction. To enhance the model's ability to identify anomalies, derived features with physical meaning are artificially constructed.

[0114] Cooling efficiency (EER). Formula: Cooling efficiency = Cooling power / (Operating current × Voltage). Assuming a rated voltage of 220V, the calculated cooling efficiency of SN001 is: 3500 / (5.8 × 220) ≈ 2.74. This value is significantly lower than the normal value (usually > 3.0), which may indicate problems such as poor heat exchange, refrigerant leakage, or low compressor efficiency. This derived characteristic reflects core performance more directly than power or current alone.

[0115] Pressure decay rate: Formula: (Initial pressure - Stabilized pressure) / Initial pressure. Assuming a continuous time series of pressure data, its decay rate can be calculated. For example, if the pressure slowly decreases from 2.1 MPa to 2.05 MPa, the decay rate = (2.1 - 2.05) / 2.1 ≈ 2.38%. A relatively high decay rate is a strong indication of micro-leakage in the system. This feature is highly sensitive for detecting such potential faults. By constructing derived features, the raw data is transformed into a more interpretable and sensitive indicator of abnormal states.

[0116] Categorical Feature Encoding. "Production Team" is a categorical feature (Class Label), which the model cannot directly process as text like "A", "B", and "C". It needs to be converted to a numerical form. One-hot encoding is used to convert it into a binary vector, as shown in the table below.

[0117]

[0118] The "Production Team" feature of SN001 (Team A) changes from A to (1, 0, 0); the "Production Team" feature of SN002 (Team B) changes from B to (0, 1, 0); and the "Production Team" feature of SN006 (Team C) changes from C to (0, 0, 1). After encoding, the model can understand categorical features and learn the potential correlation between different teams and product quality (e.g., whether the products of Team C are generally less efficient).

[0119] Final preprocessed sample example (SN001): After all the above steps, a raw, messy data record has been transformed into a clean, rich feature vector that can be efficiently learned by the model. The final processed sample is shown in the table below.

[0120]

[0121] The above demonstrates how systematic data preprocessing can transform raw, potentially erroneous industrial data into a set of high-quality features that can effectively drive a quality monitoring AI model.

[0122] It should be noted that if the (missing rate) is greater than 10%, the entire sample will be removed.

[0123] When faced with a sample, the following steps are required to determine and handle its missing values:

[0124] Identification and Calculation: First, check the completeness of each feature value in the sample, count the number of features with missing values, and calculate their proportion relative to the total number of features.

[0125] Judgment and Decision: If the missing rate is ≤10%, it is considered mildly missing. It is assumed that most of the information in this sample is still complete and valid, possessing preservation value. Therefore, imputation methods using the mean or mode are employed to repair it, allowing it to continue participating in subsequent modeling analyses.

[0126] If the missing data rate is >10%, it is considered severely missing. This means that a significant portion of the information in the sample is lost. If imputation methods are still used in this case, a large amount of uncertainty and inaccurate information will be artificially introduced, greatly reducing the overall credibility and reliability of the sample. Using an unreliable sample to train a model introduces noise that is likely to far outweigh its value, causing the model to learn incorrect patterns and affecting the final prediction accuracy and reliability. Therefore, the safest and most scientific approach is to discard the sample, removing it entirely from the dataset to ensure that the dataset used for training is as pure and reliable as possible.

[0127] For example: There is an air conditioner dataset with 10 feature dimensions. Now there is a sample record SN123, whose data is as follows:

[0128]

[0129] In this sample, we can see that the values ​​of four features—high pressure, inlet / outlet temperature difference, ambient temperature, and valve opening—are missing (represented by -). The missing value rate is calculated as follows: number of missing features = 4, total number of features = 10, missing value rate = (4 / 10) * 100% = 40%. Since a missing value rate of 40% is much higher than the specified 10% threshold, we will not attempt to fill these missing values ​​with the mean or median of other samples. This sample will be considered invalid. The entire row of sample SN123 will be deleted from the dataset to ensure that the data used for further analysis and modeling does not contain unstable data with such a high missing value rate.

[0130] Data volume reduction: Performing this operation will reduce the total number of samples in the dataset. When data is already scarce, it's necessary to assess the impact of removing a large number of samples on model training.

[0131] Cause Investigation: A high missing data rate is usually not a coincidence. This may indicate a specific malfunction in the data acquisition system (e.g., a sensor module malfunctions during a specific time period, causing a batch of data to simultaneously lack multiple relevant features), or a flaw in the data entry process. When such samples are discovered, in addition to removing them according to the rules, feedback should be provided to the data acquisition team for root cause analysis to improve data quality from the source.

[0132] Coordination with other rules: A sample may have multiple problems simultaneously. For example, it may have one outlier that clearly exceeds physical limits, while its missing value is >10%. In this case, the principle of overall removal is still followed, without needing to correct the outlier first and then determine the missing value.

[0133] In summary, "removing all data with a missing rate > 10%" is a key principle based on data quality management and model reliability considerations. The core idea is: better to have fewer but better data, avoiding low-quality data from contaminating the entire training set, thereby ensuring the accuracy and robustness of the final model.

[0134] Example 3

[0135] KFR-35GW Air Conditioner Production Line Intelligent Quality Inspection

[0136] Suppose there is a production line that specializes in producing KFR-35GW model air conditioners. This solution uses a method to automatically detect defective units coming off the production line.

[0137] Step 1: Define the objective and prepare the data. The objective is to generate a prediction label Yi for each air conditioner. Yi = 0: indicates that the air conditioner is judged as normal by the model. Yi = 1: indicates that the air conditioner is judged as an abnormal suspected machine by the model.

[0138] Dataset Construction: Historical data from 10,000 air conditioners of this model were collected, and a data profile Xi (integrating genetic, physical examination, and strain data) containing 20 features was constructed for each air conditioner. The dataset was divided as follows: Training set (6,000 units): All air conditioners have been manually confirmed to be normal in the past. This set is used to teach the model "what a normal air conditioner should look like." Validation set (2,000 units): Contains 1,500 normal units and 500 known abnormal units of various types (e.g., refrigerant leaks, fan failures, capacitor degradation, etc.). This set is used to find the optimal anomaly threshold. Test set (2,000 units): Contains 1,800 normal units and 200 abnormal units (simulating newly produced products awaiting inspection; their labels are unknown to us, used for final model evaluation).

[0139] Step 2: Model Training and Threshold Setting. Model Selection and Parameter Initialization: The model is a Gradient Boosting Regression Tree (GBDT) model. Objective: mse (the goal is to minimize the reconstruction error); Learning_Rate: 0.1; Max_Depth: 6 (controlling tree depth to prevent overfitting); N_Estimators: 200 (building 200 trees); Subsample: 0.8 (each tree randomly uses 80% of the samples); Colsample: 0.8 (each tree randomly uses 80% of the features).

[0140] Training and Fitting. Training: The GBDT model was trained using 6000 normal training machines. The model's task was to learn a function f such that f(Xi)≈Xi. That is, given the features of a normal air conditioner, the model could almost perfectly reconstruct its features.

[0141] Fitting: The trained model is used to predict the performance of 2000 air conditioners in the validation set. For each air conditioner, the mean squared error (MSE) is calculated. Normal air conditioners: Because the model learned from these, the predicted values ​​are very close to the true values, and the MSE values ​​are small. For example, most are between 0.01 and 0.5. Abnormal air conditioners: The model has never seen their patterns before, and the model cannot accurately reconstruct them. The predicted values ​​deviate significantly from the true values, and the MSE values ​​are large. For example, they may be as high as 5.0 or 10.0.

[0142] Set an error threshold TH. Plot the MSE distribution of all 2000 samples in the validation set and calculate their mean (μ_mse) and standard deviation (σ_mse). Assume μ_mse = 0.3, σ_mse = 0.2. It is found that the MSE of most normal units is concentrated in the low-value region, and the distribution is approximately normal. According to the 3σ principle, set the anomaly detection threshold: TH = μ_mse + 3 * σ_mse = 0.3 + 3 * 0.2 = 0.9. This means that any air conditioner with an MSE > 0.9 on the validation set has a 99.7% probability of being considered abnormal.

[0143] Step 3: Anomaly Detection and Diagnosis. Currently, 2000 new air conditioners (test set) have rolled off the simulated production line to simulate real-world testing.

[0144] Anomaly detection. Data from these 2000 air conditioners was input into the trained model to obtain the MSE (Mean Sequence Size) for each air conditioner. Decision rule: IF MSE > 0.9 THEN Yi = 1 ELSE Yi = 0. The model identified 250 suspected anomalies with Yi = 1. The remaining 1750 air conditioners were classified as normal (Yi = 0).

[0145] Feature Importance Analysis. We want to know not only which air conditioner is malfunctioning, but also what's wrong with it. We use SHAP analysis to diagnose an air conditioner (assuming SN: AC20230920001) that was flagged as malfunctioning (high MSE).

[0146] Analysis revealed that the three characteristics contributing most to the MSE of this sample were: "Condenser inlet and outlet temperature difference" (45% contribution): the actual value was much lower than the predicted value. Possible causes: insufficient refrigerant or low condenser heat exchange efficiency; "Compressor operating current" (30% contribution): the actual value was higher than the predicted value. Possible causes: excessive compressor load, possibly due to excessive refrigerant or system blockage; and "Fan speed" (15% contribution): the actual value was lower than the predicted value. Possible causes: fan capacitor failure or motor malfunction.

[0147] Advantages: Maintenance personnel no longer need to blindly check the entire line, but can directly prioritize checking the condenser, compressor and fan modules, which greatly improves the efficiency of troubleshooting.

[0148] Step 4: Closed-loop feedback and continuous optimization. The 250 suspected anomalous machines identified by the model were manually reviewed, confirming that: 220 were indeed genuine anomalous (TP), and 30 were misclassified as normal machines (FP). It was also found that 5 anomalous machines in the test set were missed by the model (FN). Correction: The 30 FPs were removed from the anomalous list, and the 5 FNs were added. This resulted in a more accurate list of 225 anomalous machines.

[0149] Performance evaluation. Model performance is calculated using confirmed data: Precision = TP / (TP + FP) = 220 / (220 + 30) ≈ 88%. This means that 88% of the air conditioners judged as abnormal by the model are actually faulty, which reduces unnecessary rework costs.

[0150] Recall rate = TP / (TP + FN) = 220 / (220 + 5) ≈ 97.8%. This means that 97.8% of all problematic air conditioners were successfully identified, with very few escaping detection.

[0151] F1 score = 2 * (88% * 97.8%) / (88% + 97.8%) ≈ 92.6%. This is a very high overall score, indicating that the model achieves an excellent balance between precision and recall.

[0152] Model optimization iteration. Incremental training: 30 misclassified normal machines (FPs) were added to the training set as new normal samples. Simultaneously, five newly discovered anomalous sample types were saved as references. Recalibration: The model was retrained using the expanded training set, and the threshold TH was recalculated based on the new validation set results. Root cause analysis (RCA): Analysis of the 30 FPs revealed that they all originated from a new batch of capacitors from supplier B. Although the capacitor parameters were within acceptable limits, their electrical characteristics caused the model to misclassify them. This conclusion was fed back to the purchasing and quality departments for supplier quality verification.

[0153] As mentioned above, this solution provides a complete closed-loop quality control system that integrates automatic detection, intelligent diagnosis, and continuous optimization, which can truly empower air conditioning manufacturers.

[0154] Example 4

[0155] The following explains the SHAP analysis: The intelligent quality inspection model on the air conditioning production line detected an air conditioner with serial number SN: AC20230920001. Its reconstruction error MSE was as high as 8.5, far exceeding the anomaly judgment threshold TH (0.9), and therefore it was judged as an "anomaly suspect" (Yi=1). The task is: without disassembling the machine, quickly locate the most likely root cause of the problem.

[0156] Step 1: Model Prediction and Data Preparation. Model Input: Extract all raw feature data of the air conditioner to form a feature vector. Assuming 20 features are collected, this example focuses on 5 key features: compressor operating current (A), condenser inlet / outlet temperature difference (°C), evaporator inlet / outlet temperature difference (°C), fan operating speed (RPM), and system operating pressure (MPa). Model output: The input is fed into a pre-trained GBDT regression model, which outputs a predicted value for each feature. The model is trained on a large number of normal air conditioners, so the predicted values ​​represent "what these feature values ​​should be under normal circumstances".

[0157] Calculate residuals: Calculate the difference between the true value and the predicted value of each feature (residuals), which is the source of MSE.

[0158] Step 2. Perform SHAP analysis. Use a SHAP library (e.g., shap.Explainer) to analyze this anomalous machine. SHAP calculates a value for each feature, representing the feature's contribution to the final "high MSE" result. Assume the obtained SHAP values ​​are shown in the table below.

[0159]

[0160] A positive SHAP value indicates that the true value of that feature, compared to the predicted value, increases the model's output (in this scenario, it increases the MSE error). The larger the value, the greater the contribution.

[0161] Step 3: Interpreting Results and Root Cause Analysis. Based on the SHAP analysis results, maintenance engineers can immediately pinpoint the problem area, rather than blindly troubleshooting.

[0162] Primary suspect - Condenser module (45% contribution): Symptom: The actual temperature difference (12°C) is significantly lower than the model-predicted normal value (22°C). An excessively small temperature difference indicates extremely low condenser heat exchange efficiency. Possible causes: Refrigerant leakage: Insufficient refrigerant in the system leads to low condensing pressure and poor heat exchange. Condenser blockage: Dust and oil clogging the heat sink fins results in insufficient airflow and ineffective heat dissipation. Condenser fan malfunction: Although the fan speed is normal, the fan blades may be damaged or improperly installed, leading to insufficient actual airflow.

[0163] Secondary Suspicious Target - Compressor Module (Contribution 30%): Phenomenon: Operating current (8.5A) is significantly higher than the predicted value (7.0A). High current indicates excessive compressor load. Possible Causes: System blockage: This could be due to weld blockage or dirt blockage, leading to excessively high discharge pressure and increased compressor work. Compressor-related problems: Mechanical issues such as short circuits between motor windings or poor lubrication can cause increased operating current. Correlation Analysis: Notably, poor condenser heat exchange (the primary issue) can lead to increased condensing pressure, which can indirectly cause increased compressor current. Therefore, these two problems are likely rooted in the same underlying cause.

[0164] Based on the above analysis, the quality inspection system can automatically generate an intelligent diagnostic report. Abnormal Machine Diagnostic Report - SN: AC20230920001: Abnormal Score: MSE = 8.5. Primary Investigation Point: Condenser System (Confidence: Very High). Recommended Checks: Refrigerant pressure and leakage; condenser cleanliness; condenser fan operating status.

[0165] Secondary troubleshooting point: Compressor load (confidence level: high). Recommended check: Are there any blockages in the system? Related tip: A condenser malfunction may be the cause of high compressor current; please check this first.

[0166] like Figure 2 As shown, the air conditioner quality inspection model training device 200 includes:

[0167] The acquisition module 210 is configured to acquire production process parameters, operating performance indicators, and environmental adaptability data;

[0168] The preprocessing module 220 is configured to preprocess production process parameters, operating performance indicators, and environmental adaptability data to obtain high-quality feature data. The high-quality feature data is divided into a training set, a validation set, and a test set. The training set consists of quality inspection data of normal air conditioners, the validation set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners, and the test set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners.

[0169] Training module 230 is configured to train a gradient boosting regression tree model using the training set;

[0170] Prediction module 240 is configured to use a trained gradient boosting regression tree model to predict the validation set and determine the abnormal air conditioning judgment threshold.

[0171] Input module 250 is configured to input the test set into the trained gradient boosting regression tree model to obtain the mean squared error of each air conditioner;

[0172] The comparison module 260 is configured to compare the mean square error of each air conditioner with the abnormal air conditioner judgment threshold to determine whether the current air conditioner is normal.

[0173] The calculation module 270 is configured to calculate the precision, recall, and F1 score of the regression tree model based on the number of truly abnormal air conditioners after manual re-inspection, the number of normally misjudged air conditioners, and the number of abnormal air conditioners missed in the test set, thereby completing the training of the air conditioner quality inspection model.

[0174] like Figure 3 As shown, the air conditioning quality inspection model training apparatus 300 includes a memory 310 and a processor 320 coupled to the memory 310. The memory 310 is used to store instructions for executing embodiments of the air conditioning quality inspection model training method. The processor 320 is configured to execute the air conditioning quality inspection model training method in any of the embodiments of this disclosure based on the instructions stored in the memory 310.

[0175] Figure 4 This is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure.

[0176] like Figure 4 As shown, the computer system 400 can be represented in the form of a general computing device. The computer system 400 includes a memory 410, a processor 420, and a bus 430 connecting different system components.

[0177] The memory 410 may include, for example, system memory, non-volatile storage media, etc. The system memory may store, for example, an operating system, application programs, a boot loader, and other programs. The system memory may include volatile storage media, such as random access memory (RAM) and / or cache memory. The non-volatile storage media may store, for example, instructions for executing at least one of the corresponding embodiments of the air conditioning quality inspection model training method. Non-volatile storage media include, but are not limited to, disk storage, optical storage, flash memory, etc.

[0178] The processor 320 can be implemented using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete hardware components such as discrete gates or transistors. Accordingly, each module, such as the acquisition module, preprocessing module, training module, prediction module, input module, comparison module, and calculation module, can be implemented by executing instructions in the central processing unit (CPU) running memory to perform the corresponding steps, or by implementing dedicated circuitry to perform the corresponding steps.

[0179] Bus 330 can use any of the various bus architectures. For example, bus architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, and Peripheral Component Interconnect (PCI) bus.

[0180] The computer system 300 may also include an input / output interface 340, a network interface 350, and a storage interface 360. These interfaces 340, 350, and 360, as well as the memory 310 and processor 320, can be connected via a bus 330. The input / output interface 340 provides a connection interface for input / output devices such as a monitor, mouse, and keyboard. The network interface 350 provides a connection interface for various networked devices. The storage interface 360 ​​provides a connection interface for external storage devices such as floppy disks, USB flash drives, and SD cards.

[0181] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations thereof, can be implemented by computer-readable program instructions.

[0182] These computer-readable program instructions are provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable device to produce a machine, such that execution of the instructions by the processor produces means for implementing the functions specified in one or more boxes of the flowchart and / or block diagram.

[0183] These computer-readable program instructions may also be stored in a computer-readable storage medium. These instructions cause a computer to work in a particular manner to produce an article of manufacture, including instructions that implement the functions specified in one or more boxes in a flowchart and / or block diagram.

[0184] This disclosure may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0185] By applying deep learning to product quality inspection datasets, the system gains the ability to automatically intercept suspected defective products. Through early warnings, closed-loop feedback, and adjustment of production parameters, it transforms "post-event processing" into "process prevention," significantly reducing the defect rate. It also reveals complex and implicit correlations between parameters, uncovering deep-seated patterns to quickly pinpoint the root cause of batch defects. Furthermore, the model continuously iterates and adapts to new processes and material changes by learning and optimizing from new data, ensuring that the detection accuracy remains within a stable range.

[0186] The training method, apparatus, and medium for the air conditioning quality inspection model according to this disclosure have been described in detail. To avoid obscuring the concept of this disclosure, some details known in the art have not been described. Those skilled in the art can fully understand how to implement the technical solutions disclosed herein based on the above description.

[0187] While specific embodiments of this disclosure have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of this disclosure. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this disclosure. The scope of this disclosure is defined by the appended claims.

Claims

1. A training method for an air conditioner quality inspection model, characterized in that, The method includes: Obtain production process parameters, operating performance indicators, and environmental adaptability data for air conditioners; The production process parameters, operating performance indicators, and environmental adaptability data are preprocessed to obtain standard feature data. The standard feature data is divided into training set, validation set, and test set. The training set consists of quality inspection data of normal air conditioners, the validation set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners, and the test set includes quality inspection data of normal air conditioners and quality inspection data of abnormal air conditioners. Use the training set to train a gradient boosting regression tree model; The trained gradient boosting regression tree model is used to predict the validation set to determine the threshold for abnormal air conditioning. Input the test set into the trained gradient boosting regression tree model to obtain the mean squared error of each air conditioner; The mean square error of each air conditioner is compared with the threshold for judging abnormal air conditioners to determine whether the current air conditioner is normal. Based on the number of truly abnormal air conditioners after manual re-inspection, the number of normally normal air conditioners misjudged, and the number of abnormal air conditioners missed in the test set, the precision, recall, and F1 score of the regression tree model are calculated to complete the training of the air conditioner quality inspection model.

2. The air conditioning quality inspection model training method according to claim 1, characterized in that, The process of preprocessing production process parameters, operational performance indicators, and environmental adaptability data to obtain standard feature data includes: By associating production process parameters, operational performance indicators, and environmental adaptability data through the unique serial number of the air conditioner, an air conditioner data archive sample is obtained. The air conditioner data archive sample includes the unique serial number of the air conditioner and multiple feature variables. When the missing rate of a feature variable in an air conditioner data archive sample is less than or equal to the missing threshold, the mean or mode of that feature from other complete samples of the same model and under the same operating conditions is used to fill the gap. If the missing rate of a feature variable in an air conditioning data archive sample is greater than the missing threshold, the entire air conditioning data archive sample will be removed.

3. The air conditioning quality inspection model training method according to claim 2, characterized in that, If a certain characteristic of an air conditioning data archive sample exceeds the normal range or physical limits, it is corrected using the median of that characteristic from other valid samples in the same batch.

4. The air conditioning quality inspection model training method according to claim 2 or 3, characterized in that, If the value of the feature variable is numerical, convert the value of the feature variable into a distribution with a mean of 0 and a standard deviation of 1.

5. The air conditioning quality inspection model training method according to claim 4, characterized in that, Derived features are constructed based on the feature variables.

6. The air conditioning quality inspection model training method according to claim 5, characterized in that, When the feature variable is a category, the category value is one-hot encoded.

7. The air conditioning quality inspection model training method according to claim 1, characterized in that, The determination of the abnormal air conditioner judgment threshold includes: Determine the mean square error for each air conditioner; The mean and standard deviation are calculated based on the mean square error. The threshold for identifying abnormal air conditioners is determined based on the mean and standard deviation.

8. The air conditioning quality inspection model training method according to claim 7, characterized in that, The step of determining the abnormal air conditioner judgment threshold based on the mean and standard deviation includes: According to the formula: The threshold for determining abnormal air conditioning was calculated. ; in, This represents the mean. It represents the standard deviation.

9. The air conditioning quality inspection model training method according to claim 1, characterized in that, After determining whether the air conditioner is functioning properly, the process also includes: SHAP analysis is used to diagnose abnormal air conditioners and determine the root cause of the malfunction.

10. The air conditioning quality inspection model training method according to claim 1, characterized in that, Before completing the training of the air conditioner quality inspection model, the following steps are also included: Add the quality inspection data of the misjudged normal air conditioners to the training set, and use the new training set to train the gradient boosting regression tree model; The trained gradient boosting regression tree model is used to predict new validation sets, the threshold for abnormal air conditioning is redefined, and the model optimization iteration is continuously performed.

11. An air conditioner quality inspection model training device, characterized in that, include: The acquisition module is used to acquire production process parameters, operating performance indicators, and environmental adaptability data of the air conditioner. The preprocessing module is used to preprocess production process parameters, operating performance indicators, and environmental adaptability data to obtain standard feature data. The standard feature data is divided into training set, validation set, and test set. The training set is the quality inspection data of normal air conditioners, the validation set includes the quality inspection data of normal air conditioners and the quality inspection data of abnormal air conditioners, and the test set includes the quality inspection data of normal air conditioners and the quality inspection data of abnormal air conditioners. The training module is used to train gradient boosting regression tree models using the training set; The prediction module is used to use the trained gradient boosting regression tree model to predict the validation set and determine the threshold for abnormal air conditioning judgment. The input module is used to input the test set into the trained gradient boosting regression tree model to obtain the mean squared error of each air conditioner. The comparison module is used to compare the mean square error of each air conditioner with the abnormal air conditioner judgment threshold to determine whether the current air conditioner is normal. The calculation module is used to calculate the precision, recall, and F1 score of the regression tree model based on the number of truly abnormal air conditioners after manual re-inspection, the number of normally misjudged air conditioners, and the number of abnormal air conditioners missed in the test set, and to complete the training of the air conditioner quality inspection model.

12. An air conditioner quality inspection model training device, characterized in that, include: Memory; as well as A processor coupled to the memory, the processor being configured to execute the air conditioning quality inspection model training method as described in any one of claims 1 to 10, based on instructions stored in the memory.

13. A computer-readable storage medium, characterized in that, It stores computer program instructions, which, when executed by a processor, implement the air conditioning quality inspection model training method as described in any one of claims 1 to 10.

Images