A method and system for predicting frosting of an evaporator of a central air conditioning system of a cold storage

Abstract

The application provides a cold storage central air conditioner evaporator operation state frosting prediction method and system. Air side pressure difference, fan current and refrigerant superheat degree are synchronously collected by a sensor array and pretreated; based on the physical evolution law of frosting, the superheat degree offset, pressure difference rising rate and current fluctuation variance complementary to the monitoring blind area are extracted to construct a multi-dimensional feature vector; historical data is labeled, and a linear weighted equation is constructed by using the characteristics of three features decoupling by stages through multiple linear regression to output a comprehensive feature value CFV in real time; a mapping relationship between CFV and physical frost layer thickness is established, the judgment threshold is calibrated to the CFV value corresponding to the physical boundary when the frost layer thickness is set; when the CFV reaches the threshold, the defrosting strategy is issued, otherwise the refrigeration is maintained. The application eliminates the stage blind area of a single parameter, anchors the prediction value to the physical boundary with thermodynamic basis, has small calculation amount, strong interpretability and is easy to deploy in PLC or low-power single-chip microcomputer.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for refrigeration systems, and in particular to a method and system for predicting the frost formation on the evaporator of a central air conditioning unit in a cold storage facility. Background Technology

[0002] During long-term operation, the low temperature and high humidity environment inside cold storage makes it easy for frost to accumulate on the fin surface of the evaporator. Frost has an extremely low thermal conductivity (approximately 0.02~0.1 W / (m·K)), which significantly increases the evaporator's thermal resistance, leading to decreased refrigeration efficiency, increased energy consumption, and in severe cases, even system shutdown or liquid slugging risk. Accurately detecting the evaporator's frosting status and promptly initiating defrosting is crucial for energy conservation and equipment protection in cold storage facilities.

[0003] Existing indirect frosting detection methods typically rely on a single operating parameter and can be mainly categorized as follows: (1) Pressure difference method: The degree of frosting is determined by monitoring the change in air pressure difference on both sides of the evaporator. This method is effective when the wind resistance increases significantly in the middle and late stages, but it has almost no response when the frost layer is very thin in the early stage of frosting, and there is a significant monitoring blind zone; (2) Current method: The frosting state is indirectly determined by monitoring changes in the fan current. Similar to the pressure difference method, the current method is sensitive when the wind resistance increases sharply in the middle and later stages, causing the fan to stall, but it is almost insensitive in the early stages. (3) Temperature difference method or superheat method: Frosting is determined by monitoring the temperature difference between the inlet and outlet of the evaporator or the change in the superheat of the refrigerant. This method is sensitive to changes in thermal resistance in the early stage of frosting, but its sensitivity decreases in the middle and late stages when the superheat tends to saturate.

[0004] This demonstrates that the sensitivity of a single parameter varies significantly across different stages of frosting, making it difficult to cover the entire frosting cycle. Furthermore, while some existing methods employ multi-parameter fusion, their comprehensive feature values ​​lack an explicit mapping relationship with real physical quantities (such as frost thickness), and the judgment thresholds are often based on empirical settings, lacking clear physical basis. Other methods use nonlinear machine learning models such as neural networks for fusion, which can improve accuracy, but their black-box nature leads to poor interpretability. On-site maintenance personnel find it difficult to ascertain the contribution of each physical parameter to frosting, and complex models place excessive demands on the computing power of the underlying controller (such as a PLC or lightweight microcontroller), making large-scale deployment at the industrial edge difficult.

[0005] Therefore, there is an urgent need for a frost prediction solution that can cover the entire frost cycle, has physical interpretability, and has the advantage of low computing power deployment. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for predicting frost formation on the evaporator of a cold storage central air conditioning system. The aim is to eliminate the blind spots in the phased monitoring of a single parameter, anchor the prediction results to the frost thickness with clear physical meaning, and achieve high-precision full-cycle prediction with white-box interpretability under extremely low computing power.

[0007] Specifically, it includes: S1: During operation, the air-side pressure difference of the evaporator, the fan current and the refrigerant superheat are synchronously collected by the sensor array at a set period, and a standard dataset is obtained after noise reduction and normalization preprocessing. S2: Based on the physical evolution law of the entire frost cycle, three core feature variables that complement each other in the monitoring blind zone are extracted from the standard dataset: the overheating offset, which is dominantly sensitive in the early stage of frost, and the differential pressure rise rate and current fluctuation variance, which are dominantly sensitive in the middle and late stages of frost. A multidimensional feature state vector is constructed from the three variables. S3: The historical operation data is labeled using the automatic back-induction labeling method. Taking advantage of the decoupling characteristics of the three core feature variables according to the frosting stage, the multidimensional feature state vector is used as input and the labeling result is used as the supervision signal. The weight coefficients of each dimension are solved by multiple linear regression and a linear fusion weighted equation is constructed to output the comprehensive feature value CFV in real time. S4: Establish the comprehensive characteristic value CFV and the physical frost thickness on the evaporator surface. The function mapping relationship is determined, and the forced defrosting threshold T is calibrated to... Take the CFV boundary value corresponding to the set physical boundary value; S5: Compare the real-time CFV with the threshold T. If CFV≥T, it is determined to be severe frost and a defrosting control strategy is issued. Otherwise, the current refrigeration cycle is maintained.

[0008] Furthermore, in S1, the cycle is set to 1 minute; the air-side pressure difference is collected by a micro differential pressure transmitter arranged on the windward and leeward sides of the evaporator; the fan current is collected by a Hall current sensor that is non-contactly connected to the fan power supply line; the refrigerant superheat is obtained by a combined pressure and temperature measurement method, which is the difference between the saturated evaporation temperature calculated from the suction pressure and the measured suction temperature.

[0009] Furthermore, the preprocessing of S1 includes: using the Laida criterion to remove outliers and using linear interpolation to fill in missing values; using moving average filtering or Kalman filtering to denoise and smooth the data; and using Min-Max normalization to map each dimension of the data to the [0,1] interval.

[0010] Furthermore, the three core feature variables in S2 are highly coupled with the physical evolution of the entire frosting cycle: in the early stage of frosting, the overheating offset is the dominant sensitive variable, and in the middle and late stages of frosting, the pressure difference rise rate and current fluctuation variance are the dominant sensitive variables, so that the multidimensional feature state vector always has at least one feature component in the sensitive interval throughout the entire cycle.

[0011] Furthermore, in S3, the automatic back-calculation labeling method uses the actual time of each defrosting trigger as a benchmark, labeling sampling points within a preset time window (preferably 30 minutes) as frosted and the rest as non-frosted; and uses the least squares method to perform multiple linear regression to obtain a weighted equation. The weighting coefficients vary depending on the temperature and humidity load conditions of the cold storage. Their relative magnitudes reflect the physical contribution of each characteristic parameter to the frost determination, enabling the model to have white-box interpretability and adaptability to different operating conditions.

[0012] Furthermore, in step S4, the thickness of the frost layer is simultaneously acquired by a laser displacement thickness sensor on the surface of the evaporator. The mapping relationship between the CFV sample pairs and the CFV sample pairs is obtained by binomial regression fitting. The set physical boundary value (preferably 0.1 mm) is then substituted into the inverse solution to obtain the judgment threshold T.

[0013] Furthermore, S5 also includes a frosting trend warning mechanism: a warning interval is set below the threshold T. When CFV falls into the interval but does not reach T, the CFV rise slope is calculated and the remaining time to reach T is output. When CFV≥T, the defrosting strategy is converted into a standard control signal of the underlying controller to drive the hot gas bypass solenoid valve or electric heater to defrost.

[0014] Furthermore, the method also includes step S6: generating and storing white-box log information each time the judgment and defrosting control are executed, the log containing raw operating data, standardization results, feature vector values, weight coefficients of each dimension, CFV, mapping calibration parameters and defrosting action instructions.

[0015] Furthermore, the method also includes at least one of the following: judgment shielding of compressor start-stop transients, extending the feature calculation sliding period when the environment changes abruptly, judgment hysteresis during the recovery period after defrosting, and guiding with an initial weight template when new deployments are made without historical records and switching to automatic back-inference annotation training after sufficient samples.

[0016] The present invention also provides a frosting prediction system for the evaporator operation status of a cold storage central air conditioning system, for implementing the above method, including: an operation data acquisition module, a feature state vector construction module, a multi-dimensional parameter fusion prediction module, a physical mapping and judgment module, a strategy execution module and a log management module; Furthermore, the system includes differential pressure transmitters arranged on the windward and leeward sides, Hall current sensors snapped into the power supply line of the wind turbine, and pressure transmitters and platinum resistance temperature sensors installed in the intake pipe.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention provides a method and system for predicting the frost formation of the evaporator in a cold storage central air conditioning system. It captures the initial thermal resistance change by superheat offset, the pressure difference rise rate and the current fluctuation variance to capture the airflow blockage in the middle and later stages. The integration of these three factors eliminates the blind spots in the phased monitoring of a single parameter.

[0018] 2. This invention provides a method and system for predicting frost formation in the evaporator of a cold storage central air conditioning system. It uses multiple linear regression to visually display the contribution of each parameter and anchors the comprehensive feature value to the physical boundary of the frost layer thickness, so that each defrosting determination has a thermodynamic basis.

[0019] 3. This invention provides a method and system for predicting the frost formation of the evaporator in a cold storage central air conditioning system. It reduces the detection dimension to linear algebraic operations, which can be directly programmed into a regular PLC or a low-power microcontroller to perform high-speed calculations, significantly reducing computational overhead compared to the neural network black-box method.

[0020] 4. This invention provides a method and system for predicting the frost formation of the evaporator in a cold storage central air conditioning system. It generates a complete white-box log and outputs the remaining time by the rising slope, thus realizing the transformation from post-processing to predictive maintenance. Compared with the traditional timed defrosting scheme, it has a significant energy-saving effect. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of the full-cycle frost prediction method of the present invention; Figure 2 This is a schematic diagram of the coupling of multidimensional parameter frosting physical evolution and the weighted process of multiple linear regression; Figure 3 This is a schematic diagram of the process for mapping and calibrating the comprehensive characteristic value CFV with the physical frost layer thickness and setting the threshold. Detailed Implementation

[0023] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0024] CFV: Composite Feature Value, a custom-defined continuous prediction value in this invention that characterizes the severity of frost formation.

[0025] SH: Superheat, the difference between the actual suction temperature and the saturated evaporation temperature of the refrigerant.

[0026] FS: Full scale, "±0.5%FS" means the error is ±0.5% of the full scale.

[0027] PT100: Standard designation for platinum resistance temperature sensor, referring to a platinum resistance thermometer with a resistance of 100Ω at 0℃.

[0028] As a specific implementation method, this embodiment takes a large cold storage as the implementation scenario, with the storage temperature set at -18℃, relative humidity at 85%, storage capacity at 10000m³, equipped with 6 parallel air coolers, and the sampling period set at 1 minute.

[0029] like Figure 1 As shown, in this embodiment, steps S1 to S5 are executed cyclically according to the sampling cycle, namely, data acquisition and preprocessing, construction of feature vectors, linear regression weighting to obtain CFV, mapping and calibration threshold T, comparison of CFV with T, and issuance of defrosting control or maintenance of cooling cycle based on the comparison result. At the same time, a white box log is generated for each determination.

[0030] The data acquisition module consists of a sensor array comprising a differential pressure transmitter, a Hall current sensor, and a pressure / temperature sensor, which synchronously senses the physical environment in real time while the evaporator is in operation.

[0031] Feature state vector construction module and multidimensional parameter fusion prediction module: Using a low-power microcontroller or PLC embedded at the edge of the cold storage as a carrier, microsecond-level feature extraction and calculation are performed using a pre-programmed multiple linear regression weighted equation.

[0032] The physical mapping and judgment module and the strategy execution module compare the comprehensive feature value CFV with the physical threshold and convert the decision into a 4~20mA standard analog quantity or relay switching quantity signal and send it to the underlying actuator (such as a hot gas bypass solenoid valve).

[0033] Log management module: Relies on host computer or cloud database to record white-box decision-making links.

[0034] Within each sampling period, the sensor array synchronously acquires three signals aligned with a unified timestamp: the air-side differential pressure uses a micro differential pressure transmitter with a range of 0~500Pa and an accuracy of ±0.5%FS, with pressure taps arranged on the windward and leeward sides; the fan current uses a Hall current sensor with a range of 0~20A and an accuracy of ±1%, non-contactly connected to the power supply line; the refrigerant superheat is determined using a pressure-temperature method, where the suction pressure is obtained from a pressure transmitter with a range of 0~2MPa and converted to the saturated evaporation temperature Tsat according to the refrigerant saturation pressure-temperature relationship, and the actual suction temperature is measured by a PT100 platinum resistance temperature sensor with an accuracy of ±0.2℃. The difference between the two is .

[0035] Raw data enters the preprocessing pipeline: Outlier removal and missing value handling are performed using... The criteria are used to remove outliers, and linear interpolation is used to fill in the missing samples if there are no more than 3 consecutive missing samples. For noise reduction and smoothing, a moving average filter is used to take the average value of the current sampling point and its previous N historical sampling points as the effective output. In this embodiment, N=5. Since the transient airflow disturbances caused by changes in the location of goods or the opening and closing of the door curtain in the cold storage usually disappear smoothly within 3 to 5 minutes, N=5 can filter out high-frequency spikes without masking the slow evolution trend of frost. The filtered pressure difference, current, and superheat are uniformly standardized and mapped to the [0,1] interval using Min-Max normalization.

[0036] See Figure 2 This system extracts features based on deep coupling analysis of the entire frosting cycle: In stage one (early stage of frosting, with a frost layer thickness of 0~0.1mm), the thermal resistance of the frost layer is much higher than that of the fins, and the superheat offset ΔSH reacts significantly first, while the pressure difference and current remain almost unchanged. In stage two (mid-frosting, 0.1~0.5mm), the cross-section of the airflow channel shrinks, the rate of increase in pressure difference ΔP′ increases linearly, and the variance of current fluctuation gradually increases. In stage three (late stage of frosting, >0.5mm), the airflow is nearly blocked, the current fluctuation variance reaches its peak, and the superheat tends to saturate.

[0037] The three monitoring blind spots complement each other, and after integration, a sensitive component is always present throughout the entire cycle.

[0038] The quantitative derivation of superheat offset and frost thermal resistance is as follows: from evaporator energy balance And the effective heat transfer coefficient satisfies When the thermal resistance of the frost layer When it increases, As the temperature drops, the refrigerant absorbs less heat, causing the actual evaporator outlet temperature to deviate from the design value. The superheat deviation then changes in the opposite direction. .

[0039] Numerical estimation: When the frost layer thickness increases from 0 to 0.1 mm, the median thermal conductivity of 0.06 W / (m·K) is taken. The overheating offset increases to approximately 0.00167 m²·K / W, corresponding to a decrease in ΔSH of approximately 2.5~4.5℃, far exceeding the ±0.2℃ accuracy of the platinum resistance temperature sensor. Therefore, the overheating offset is significantly physically observable in the early stages of frosting and is the most sensitive predictive feature in the initial phase. This derivation supports the non-obviousness of the feature selection in this invention.

[0040] Based on this, the characteristic sliding period m = 15 minutes is selected (taking into account the compressor's 10-20 minute start-stop cooling capacity adjustment cycle, which smooths out single flow pulsations and includes a micro-thermodynamic stability cycle), and the pressure difference rise rate is calculated. Current fluctuation variance A three-dimensional feature state vector is constructed using a superheat offset ΔSH based on a baseline superheat of 0.5. .

[0041] like Figure 2 As shown, during the model training phase: historical data from the past 12 months (approximately 525,600 sampling points, including 98 defrosting events) was retrieved. An automatic back-labeling method was used, with each defrosting trigger time as the baseline. Sampling points within the first 30 minutes were labeled as frosted (label 1), and the rest as unfrosted (label 0), automatically generating a supervised training dataset. The 30-minute backtracking duration was chosen because: a typical defrosting cycle in a cold storage facility is 4-6 hours, and 30 minutes represents approximately the last 10%-12% of the operation time before defrosting. During this stage, the frost layer generally exceeds the 0.1mm critical point, and heat exchange efficiency drops sharply, making the label's physical representation strongest. Too short a duration would result in insufficient positive samples, while too long a duration would introduce slight frost interference, contaminating the labels.

[0042] Using feature vectors as input and labels as supervision signals, a set of example weights is obtained by performing multiple linear regression using the least squares method: Among them, differential pressure has the greatest impact on the middle and late stages, and is given the highest weight of 0.48; current fluctuation is the second highest at 0.32; although superheat has a relatively low weight of 0.18, it plays a key role in breaking through the blind spots of initial monitoring.

[0043] In terms of statistical validation, the regression model showed a goodness of fit of R² = 0.918, a root mean square error (RMSE) of 0.087, and residuals that were approximately normal (Shapiro-Wilk test p = 0.34 > 0.05). The weighting coefficients were all significant in the t-test (p < 0.001), and the variance inflation factor (VIF) was all < 2.5 with no significant collinearity. This indicates that the linear regression model has sufficient statistical validity in this feature space, and there is no need to use a higher-order or nonlinear model.

[0044] Since the weighting coefficients depend on the specific cold storage conditions, for new deployment scenarios without historical records, an initial weight template can be used to guide the process by looking up values ​​in a table according to the storage temperature range. Once the accumulated number of defrosting events reaches 20, the system can switch to automatic back-calculation annotation training to achieve adaptive operation based on the conditions. An example of the initial weight template is shown in Table 1. Table 1

[0045] like Figure 3 As shown, laser displacement thickness sensors with an accuracy of ±0.01 mm are arranged on the surface of the evaporator fins to simultaneously collect 100 sets of data at different frosting stages. The sample pairs were fitted using binomial regression. (Goodness of fit R² = 0.962, RMSE = 0.018 mm; it was verified that the improvement of R² by the third and fourth regressions was less than 0.005, so the binomial regression was the optimal order to avoid overfitting).

[0046] The selection of the 0.1mm physical boundary is based on the following: when the thermal conductivity of the frost layer is 0.02~0.1W / (m·K), the additional thermal resistance generated by a 0.1mm frost layer is approximately 0.001~0.005m²·K / W, which is more than 20% of the normal heat exchange thermal resistance of the evaporator. Furthermore, at this point, the additional energy consumption increase due to the decrease in heat exchange efficiency reaches the critical value required for one defrosting operation. Substituting 0.1 into the mapping function and solving the inverse problem... The threshold value is T≈0.65.

[0047] The process is repeated every 1 minute: data acquisition → preprocessing → feature extraction → linear weighted CFV calculation → comparison with T=0.65. If CFV < 0.65, it is considered no frost or light frost, and cooling is maintained; if CFV ≥ 0.65, it is considered severe frost (corresponding to...). The control command is converted into a 4~20mA signal and sent to the PLC to open the hot gas bypass solenoid valve for defrosting. After the defrosting is completed, the cooling is automatically restored.

[0048] Early warning mechanism: An early warning interval is set between [0.85T, T) and [0.553, 0.650). When the CFV falls into this interval, the upward slope k over the past 30 minutes is calculated using first-order linear extrapolation, and the estimated remaining time is... When the CFV curve shows a decline or jitter, or other non-monotonic conditions, extrapolation should be paused to avoid false alarms.

[0049] To ensure robustness in complex operating environments, the logic decision-making process incorporates nested adaptive and fault-tolerant mechanisms. (1) Transient shielding of system start-up and shutdown: When the compressor start-up and shutdown signal is detected, the system enters the judgment shielding period within 5 minutes before and after the state switch and locks the CFV output; (2) External disturbance removal: When the temperature and humidity change caused by frequent opening and closing of the warehouse door or large-scale goods entering the warehouse is detected, the feature calculation sliding period is automatically extended to deepen the smoothing filter depth. (3) Post-defrost recovery period hysteresis: The hysteresis period is entered within 15 minutes after defrosting and cooling is restored. During this period, the severity of frost will not be reassessed. (4) Cold start initialization: New deployment scenarios without historical records are guided by the initial weight template. After the accumulated number of defrost events reaches 20, the system seamlessly switches to automatic reverse annotation training.

[0050] Sensor fault tolerance degradation: Real-time monitoring of three types of sensor health status (data unchanged for a long time, over-range, communication interruption, and abnormal noise). If any sensor fails, the system will degrade to the following level as shown in Table 2, and a maintenance alarm will be sent to the host computer: Table 2

[0051] Multiple evaporators in parallel: Each evaporator is equipped with an independent sensor array and an independent algorithm instance, calculating its own CFV and threshold. They share a cold start template, but their weight fine-tuning is independent. Defrosting scheduling supports both individual defrosting and alternating defrosting modes, with alternating defrosting being preferred to avoid significant temperature fluctuations. Defrosting ends under a combination of conditions (fin surface temperature rises to +5°C and remains there for 2 minutes, or defrosting accumulates to the maximum for 15 minutes, or the differential pressure drops to within ±5% of the clean baseline and remains there for 60 seconds), with the first condition met taking priority.

[0052] Storage temperature compensation: A storage temperature compensation factor α is introduced to make a slight adjustment to CFV. ,in , , The compensation factor table is generated by factory calibration and interpolated in real time according to the storage temperature.

[0053] Each judgment is automatically written to the log, including: sampling timestamp, original ΔP / I / SH values, standardized values, feature vector values, real-time weight distribution (w1=0.48, w2=0.32, w3=0.18, b=0.02), comprehensive feature value CFV, threshold T and corresponding frost layer boundary, mapping function coefficients, judgment result and defrosting action instructions; Logs are retained for 2 years and can be retrieved and replayed by time range and device number, enabling full traceability of the control process.

[0054] During three months of continuous operation in the aforementioned cold storage facility, this invention triggered 98 defrosting detections, with 91 true positives, 3 false positives, and 4 false negatives, resulting in an overall accuracy of 95.9%, sensitivity of 95.8%, and specificity of 88.2%. A horizontal comparison with existing methods in the same cold storage facility during the same time period is shown in Table 3. Table 3

[0055] The comparison shows that the present invention reduces the execution time per execution while maintaining the same accuracy as the neural network black box method, thus meeting the real-time control requirements of PLC or low-power microcontroller. Compared to timed defrosting, it saves approximately 13.8% of energy annually, and compared to the single-parameter method, it saves approximately 5% to 7%.

[0056] Regarding the sensitivity of key parameters, the feature calculation sliding period m has the highest accuracy (approximately 96%) at 15 minutes. Shorter periods (5-10 minutes) are susceptible to pulsation interference and lack thermodynamic stability, while longer periods (30 minutes) result in significant lag. The moving average window N is optimal at 5 (approximately 96%). The historical backtracking annotation duration is optimal at 30 minutes (approximately 96%), with shorter durations resulting in insufficient positive samples and longer durations contaminating labels. The above analysis shows that the parameters selected in this invention (m=15, N=5, and 30-minute backtracking) are all within the optimal performance range and exhibit strong robustness to small parameter perturbations.

[0057] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. A method for predicting frost formation on the evaporator of a cold storage central air conditioning system, characterized in that, include: S1: During operation, the air-side pressure difference of the evaporator, the fan current and the refrigerant superheat are synchronously collected by the sensor array at a set period, and a standard dataset is obtained after noise reduction and normalization preprocessing. S2: Based on the physical evolution law of the entire frost cycle, three core feature variables that complement each other in the monitoring blind zone are extracted from the standard dataset: the overheating offset, which is dominantly sensitive in the early stage of frost, and the differential pressure rise rate and current fluctuation variance, which are dominantly sensitive in the middle and late stages of frost. A multidimensional feature state vector is constructed from the three variables. The coupling relationship between the three core characteristic variables and the physical evolution of the entire frost cycle is as follows: In the early stage of frosting, the increased thermal resistance of the frost layer leads to deterioration of heat transfer, and the superheat offset is used as the dominant sensitive variable in this stage. In the later stages of frosting, the airflow channel narrows and the wind resistance increases. The rate of increase of the pressure difference and the variance of the current fluctuation are used as the dominant sensitive variables in this stage. The three core feature variables are combined to form the multidimensional feature state vector, so that the multidimensional feature state vector always has at least one feature component in the sensitive range throughout the entire frosting cycle, so as to avoid misjudgment caused by a single feature mutation. S3: The automatic back-inference annotation method is used to annotate historical operation data. The three core feature variables are decoupled from each other according to the frost stage. The multidimensional feature state vector is used as input and the annotation result is used as supervision signal. The weight coefficients of each dimension are solved by multiple linear regression and a linear fusion weighted equation is constructed to output the comprehensive feature value CFV in real time. S4: Establish the comprehensive characteristic value CFV and the physical frost thickness on the evaporator surface. The function mapping relationship is determined, and the forced defrosting threshold T is calibrated to... Take the CFV boundary value corresponding to the physical boundary value set; specifically including: Sample pairs are obtained by synchronously acquiring the actual frost layer thickness δ and the corresponding comprehensive feature value CFV using a laser displacement thickness sensor set on the surface of the evaporator; The sample pairs are fitted using binomial regression to obtain the function mapping relationship as follows: , where a, b, and c are fitting coefficients; Substituting the set physical boundary value into the function mapping relationship and solving inversely, the determination threshold T is obtained; S5: Compare the real-time CFV with the threshold T. If CFV≥T, it is determined to be severe frost and a defrosting control strategy is issued. Otherwise, the current refrigeration cycle is maintained.

2. The method for predicting frost formation on the evaporator of a cold storage central air conditioning system according to claim 1, characterized in that, In S1, the period is set to 1 minute. The three raw signals—air-side pressure difference, fan current, and refrigerant superheat—are synchronously collected in real time by the following sensors constituting the sensor array, aligned with a uniform timestamp, within each sampling period: The air-side pressure difference is acquired using a micro differential pressure transmitter with a range of 0~500Pa and an accuracy of ±0.5%FS. Its two pressure taps are respectively arranged on the windward and leeward sides of the evaporator to measure the inlet and outlet air pressure difference caused by the thickening of the frost layer and the narrowing of the airflow channel. The fan current is collected by a Hall current sensor with a range of 0~20A and an accuracy of ±1% in a non-contact manner by being clipped onto the power supply line of the evaporator fan. This is used to monitor the fluctuation of the operating current caused by the increase in wind resistance of the fan. The refrigerant superheat is obtained using a combined pressure and temperature measurement method. Specifically, a pressure transmitter installed on the suction line at the evaporator outlet obtains the actual suction pressure and converts it to the saturated evaporation temperature Tsat according to the saturated pressure-temperature correspondence of the refrigerant used. At the same time, a platinum resistance temperature sensor is installed in a patch manner on the pipe wall near the pressure transmitter to measure the actual suction temperature Tactual. The refrigerant superheat is calculated in real time according to the formula SH=Tactual−Tsat. The three raw signals are synchronously acquired and then sent to the noise reduction and normalization preprocessing pipeline described in S1.

3. The method for predicting frost formation on the evaporator of a cold storage central air conditioning system according to claim 1, characterized in that, The denoising and normalization preprocessing in S1 includes: The Laida criterion was used to identify and remove outliers from the original data, and linear interpolation was used to fill in missing data. The data is denoised and smoothed using moving average filtering or Kalman filtering, and the average value of the current sampling point and several previous historical sampling points is taken as the effective output. The Min-Max normalization method is used to linearly map the denoised and smoothed data of each dimension to the dimensionless interval [0,1].

4. The method for predicting frost formation in the evaporator of a cold storage central air conditioning system according to claim 1, characterized in that, In S3: The automatic back-inference annotation method is as follows: retrieve historical defrosting action records, take the actual time of each defrosting trigger as the benchmark, mark the sampling points within a preset time window before that time as the frosted state, and mark the remaining sampling points as the non-frosted state, thereby generating a supervised training dataset. Using a multidimensional feature state vector as input, the least squares method is used for multivariate linear regression training to find the weight coefficients that minimize the sum of squared residuals. The constructed multidimensional parameter fusion weighted equation is as follows: ； in The rate of increase of pressure differential, Let ΔSH be the variance of current fluctuation, and ΔSH be the superheat offset. , , Here, m represents the corresponding weighting coefficients, and m is the bias term. The preset time window is set to 30 minutes, that is, the sampling points within 30 minutes before each defrost trigger are marked as the frosted state; The weighting coefficient , , The relative magnitude of the model weights varies depending on the temperature and humidity load conditions of the cold storage. It reflects the physical contribution of each characteristic parameter to the frost determination under the condition. Based on this, the model weights can be adapted to different cold storage conditions and traced in a white-box manner.

5. The method for predicting frost formation in the evaporator of a cold storage central air conditioning system according to claim 1, characterized in that, The physical boundary value is set to a frost layer thickness of 0.1 mm; the selection criteria for the frost layer thickness of 0.1 mm as the physical reference boundary are as follows: When the thermal conductivity of the frost layer is 0.02~0.1 W / (m·K), the additional thermal resistance generated by the 0.1mm frost layer reaches more than 20% of the normal heat exchange thermal resistance of the evaporator. At this time, the additional energy consumption increase caused by the decrease in heat exchange efficiency reaches the critical value of the energy consumption required to perform one defrost. The S5 also includes a frosting trend early warning mechanism: A warning threshold range is set below the judgment threshold T. When the comprehensive feature value CFV falls into the warning threshold range but does not reach the judgment threshold T, physical defrosting is not triggered. Instead, the rising slope of the comprehensive feature value CFV is calculated, and the remaining time expected to reach the judgment threshold T is output accordingly. When the comprehensive characteristic value CFV≥T, the defrosting control strategy is converted into a standard control signal sent by the underlying controller to drive the hot gas bypass solenoid valve or electric heater to perform physical defrosting.

6. The method for predicting frost formation in the evaporator of a cold storage central air conditioning system according to claim 1, characterized in that, The method further includes step S6: White-box log information is generated and stored each time a judgment and defrost control is executed. The white-box log information includes the original running data at the time of triggering the judgment, the standardized results, the multi-dimensional feature state vector values, the weight coefficients of each dimension, the comprehensive feature value CFV, the mapping calibration parameters, and the defrost action command finally executed.

7. The method for predicting frost formation in the evaporator of a cold storage central air conditioning system according to claim 1, characterized in that, The method also includes at least one of the following: adaptive operating conditions and fault-tolerant mechanisms: For compressor start-stop transients, a judgment shielding period is entered during a set time period before and after the state switch, and the output of the comprehensive characteristic value CFV is locked. In response to sudden environmental changes caused by the opening and closing of warehouse doors or the entry of goods into the warehouse, the feature calculation sliding period used to extract the three core feature variables is automatically extended to deepen the smoothing filter depth. After defrosting is completed and cooling is restored, a hysteresis period is entered during which the severity of frost is not reassessed. For new deployment scenarios with no historical defrosting records, a preset initial weight template is used as the initial guide, and the training mode of the automatic reverse annotation method is switched after the cumulative number of running samples reaches a preset value.

8. A frosting prediction system for the evaporator operating status of a cold storage central air conditioning system, used to implement the frosting prediction method for the evaporator operating status of a cold storage central air conditioning system as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect multi-dimensional operating data through a sensor array and perform noise reduction and normalization preprocessing when the evaporator is in operation. The feature state vector construction module is used to extract core feature variables and construct multi-dimensional feature state vectors based on the physical evolution law of the entire frost cycle. The multidimensional parameter fusion prediction module is used to construct a weighted equation using a multiple linear regression algorithm and weighting coefficients and output the comprehensive feature value CFV in real time. The physical mapping and judgment module is used to establish the mapping relationship between the comprehensive feature value CFV and the physical frost layer thickness δ, and to compare the real-time comprehensive feature value CFV with the judgment threshold T. The strategy execution module is used to issue defrosting control strategies to the underlying controller based on the judgment results; The log management module is used to generate and store white-box log information; The sensor array includes: a differential pressure transmitter arranged on the windward and leeward sides of the evaporator, a Hall current sensor that is non-contactly connected to the fan power supply line, and a pressure transmitter and a platinum resistance temperature sensor installed on the evaporator outlet suction pipe.

Images