Refrigerator control method and device, electronic equipment and storage medium
By extracting features from refrigerator operating data at different time scales and processing them using an attention mechanism, the problem of insufficient early fault diagnosis capability in existing refrigerator fault prediction methods is solved, enabling early identification and accurate prediction of faults.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-04-10
Smart Images
Figure CN121829022A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of refrigerator technology, and particularly relates to a refrigerator control method, control device, electronic device and storage medium. Background Technology
[0002] Some refrigerators currently use convolutional neural networks (CNNs) for fault prediction. However, this method treats all input features and all time points equally, lacking the ability to selectively focus on signals most relevant to the fault. Therefore, in the early stages of a fault, when the data is noisy and the fault-related features are relatively weak, the existing methods usually have poor fault prediction capabilities. Summary of the Invention
[0003] This application provides a refrigerator control method, control device, electronic device, and storage medium to address the problem that existing refrigerator fault prediction methods have poor diagnostic capabilities for early-stage faults.
[0004] This application provides a refrigerator control method, the method comprising: Obtain the operating data of the refrigerator; At least two different time scales are used to extract features from the running data to obtain multiple target feature information; By fusing the multiple target feature information, fused feature information is obtained; The fused feature information is adaptively weighted based on an attention mechanism to obtain weighted feature information; Based on a preset classification network, the weighted feature information is used to identify faults, and the fault information includes fault type and fault probability.
[0005] This application embodiment also provides a refrigerator control device, the device comprising: The operation data acquisition module is configured to acquire the operation data of the refrigerator; The feature extraction module is configured to perform feature extraction on the running data at at least two different time scales to obtain multiple target feature information; The fusion module fuses the multiple target feature information to obtain fused feature information; The weighting module is configured to perform adaptive weighting processing on the fused feature information based on an attention mechanism, and obtain weighted feature information; The analysis module is configured to perform fault identification on the weighted feature information based on a preset classification network to obtain fault information, which includes fault type and fault probability.
[0006] Optionally, the feature extraction of the running data at at least two different time scales includes: The running data is feature extracted using three one-dimensional dilated convolutional layers respectively; The expansion factor or convolution kernel of the three one-dimensional hollow convolutional layers increases sequentially.
[0007] Optionally, the adaptive weighting processing of the fused feature information based on the attention mechanism includes: The fused feature information is adaptively weighted using a channel attention module to obtain fault channel feature information; The fused feature information is adaptively weighted using a time attention module to obtain fault time feature information; The fault channel feature information and the fault time feature information are concatenated to obtain the weighted feature information.
[0008] Optionally, the fault identification based on the weighted feature information using a preset classification network to obtain fault information includes: The weighted feature information is flattened and globally pooled to obtain the feature vector; The feature vector is input into at least one fully connected layer to obtain the integrated vector; The integrated vector is input into the Softmax function to obtain the probability distribution vector; The fault information is obtained based on the probability distribution vector.
[0009] Optionally, it also includes: If the maximum value of the fault probability is greater than the preset probability, then the fault type corresponding to the maximum fault probability is determined as the target fault.
[0010] Optionally, it also includes: The target fault generates an early warning message and sends it to the user's mobile device.
[0011] Optionally, before performing feature extraction on the running data at at least two different time scales, the method further includes: Determine whether the refrigerator has met the diagnostic trigger conditions; The diagnostic triggering conditions include one or more of the following: the refrigerator's running time reaches a preset duration, the running data reaches a preset quality, and some of the running data exceeds a preset threshold.
[0012] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the refrigerator control method described above.
[0013] This application embodiment also provides a storage medium storing control instructions, which, when executed by a processor, implement the refrigerator control method described above.
[0014] The refrigerator control method provided in this application captures multi-granular information of the time series by extracting features from the refrigerator's operating data at at least two different time scales. This avoids misjudgment or omission caused by data from a single time scale, ensuring the comprehensiveness of data capture. Simultaneously, an attention mechanism is introduced, which helps to achieve automatic learning and assign higher weights to fault-related features, amplifying subtle fault signals and suppressing noise interference. In other words, by extracting features from the operating data at different time scales and introducing an attention mechanism, subtle fault signals can be captured and amplified in the early stages of a fault, enabling observation of fault trends and thus achieving early fault diagnosis. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] To gain a more complete understanding of this application and its beneficial effects, the following description will be provided in conjunction with the accompanying drawings. In the following description, the same reference numerals denote the same parts.
[0017] Figure 1 This is a first flowchart illustrating the refrigerator control method provided in an embodiment of this application.
[0018] Figure 2 This is a second flowchart illustrating the refrigerator control method provided in an embodiment of this application.
[0019] Figure 3 This is a schematic diagram of the control device for a refrigerator provided in an embodiment of this application.
[0020] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.
[0021] Figure 5 This is a schematic diagram of the structure of a refrigerator provided in an embodiment of this application.
[0022] Explanation of reference numerals in the attached figures: 201. Data Acquisition Module; 202. Feature Extraction Module; 203. Fusion Module; 204. Weighting Module; 205. Analysis Module; 300. Electronic device; 301. Memory; 3011. Computer program; 302. Processor. Detailed Implementation
[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0024] In the description of the embodiments of this application, "module" and "processor" can include hardware, software, or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, and memory, and may also include software components, such as program code, or a combination of software and hardware. A processor can be a central processing unit, a microprocessor, a digital signal processor, or any other suitable processor. The processor has data and / or signal processing capabilities. The processor can be implemented in software, in hardware, or a combination of both. Non-transitory computer-readable storage media includes any suitable medium capable of storing program code, such as magnetic disks, hard disks, optical disks, flash memory, read-only memory, random access memory, etc.
[0025] This application provides a refrigerator control method, control device, electronic device, and storage medium to solve the problem that existing refrigerator fault prediction methods have poor ability to diagnose early faults. The following description is in conjunction with the accompanying drawings.
[0026] The refrigerator control method provided in this application embodiment is described in detail below. Figure 1 The method includes the following steps: S101: Obtain the refrigerator's operating data.
[0027] The refrigerator's operating data includes temperature data (refrigerator compartment temperature, freezer compartment temperature, or ambient temperature), current data (operating current of the compressor, fan, and defrost heater), vibration data (vibration frequency and amplitude of the compressor and fan bearings), pressure data (pressure values at the high and low ends of the refrigeration system), and refrigerator door opening data (number of openings and opening duration). For example, temperature data can be collected using a temperature sensor, current data using a current sensor, vibration data using a vibration sensor, pressure data using a pressure sensor, and door opening data using a door switch sensor. The acquisition parameters of different sensors can be configured differently based on their signal characteristics. For example, vibration data can use a 100Hz sampling frequency to capture mechanical vibration details, while temperature data can use a 1Hz sampling frequency. All sensor data are strictly synchronized through a real-time clock module to ensure data consistency.
[0028] S102: Perform feature extraction on the running data at at least two different time scales to obtain multiple target feature information.
[0029] The time scale refers to the length of the period used to observe and analyze the data, or the size of the selected "analysis window." Target feature information refers to information that is more representative and discriminative, obtained from the raw operational data after calculation and refinement for a specific analytical purpose (i.e., the "target").
[0030] For example, assuming continuous temperature data is obtained from the refrigerator compressor, the short-term target feature information (analysis window: 10 minutes) extracted from the continuous temperature data includes: Feature 1: the maximum temperature rise slope in the past 10 minutes (this information can be used to identify faults such as "abnormal heat dissipation" or "sudden load increase"); Feature 2: the standard deviation of temperature fluctuations in the past 10 minutes (this information can be used to identify faults such as "unstable control system"). The long-term target feature information (analysis window: 24 hours) extracted from the continuous temperature data includes: Feature 1: the average operating temperature of the compressor in the past 24 hours (this information can be used to identify faults such as "overall system efficiency decline" or "insufficient refrigerant"); Feature 2: the trend slope of the daily average operating temperature in the past week (this information can be used to identify faults such as "slow performance degradation").
[0031] The "maximum rate of ascent," "standard deviation," "average operating temperature," and "trend slope" calculated above constitute the target characteristic information. They are more concise and directly point to specific equipment health conditions or failure modes than the raw, one-per-second temperature readings.
[0032] S103: Fuse multiple target feature information to obtain fused feature information.
[0033] Among them, fused feature information is a new and comprehensive feature representation obtained after multiple target feature information are fused.
[0034] For example, multiple target feature information can be directly concatenated to obtain fused feature information. The concatenation retains all the original information from different time scales without any information loss, providing the richest raw materials for subsequent attention mechanisms.
[0035] S104: Adaptive weighting of the fused feature information is performed based on the attention mechanism to obtain the weighted feature information.
[0036] It should be understood that the number of fused feature information extracted and fused from different time scales is large (for example, there may be dozens or even hundreds of features), but their importance for identifying different faults is completely different. Therefore, by introducing an attention mechanism, an "attention score" is calculated for each individual feature based on the current overall fused feature information. This score represents the importance of the feature to the current diagnostic task. Then, the "attention scores" of all features are normalized through functions such as Softmax and converted into a set of weights that sum to 1 (i.e., importance ratios). Then, this weight is used to multiply the corresponding original feature value to achieve the effect of making high-weight features larger and low-weight features smaller.
[0037] It should be noted that this weight is not a fixed value set in advance. The model will dynamically and automatically calculate the most appropriate weight based on the specific pattern and combination of the "fusion feature information" input each time. This is the meaning of "adaptive" - it can flexibly cope with the complex situation of the refrigerator under different operating conditions and different fault symptoms.
[0038] S105: Based on a preset classification network, the weighted feature information is used to identify faults and obtain fault information, which includes fault type and fault probability.
[0039] Among them, classification network refers to a machine learning model that can automatically determine and output the fault category (and probability) based on input features. It plays a decision-making role in the fault diagnosis process and is usually used to realize the complex mapping relationship from feature space to fault category space.
[0040] The refrigerator control method provided in this application captures multi-granular information of the time series by extracting features from the refrigerator's operating data at at least two different time scales. This avoids misjudgment or omission caused by data from a single time scale and ensures the comprehensiveness of data capture. Simultaneously, an attention mechanism is introduced to automatically learn and assign higher weights to fault-related features, amplifying fault signals and suppressing noise interference. In other words, by extracting features from the operating data at different time scales and introducing an attention mechanism, the method captures and amplifies minute fault signals in the early stages of a fault, enabling observation of fault trends and thus achieving early fault diagnosis.
[0041] Optionally, after acquiring the refrigerator's operating data, the process also includes: preprocessing the operating data. This preprocessing may include six sequential processing steps: outlier handling, using the interquartile range (IQR) method to identify and process outlier data points; missing value imputation, using linear interpolation to fill data gaps; signal filtering, processing according to sensor type: applying a 50Hz power frequency notch filter to the current signal to eliminate power grid interference, applying a 5Hz to 1000Hz bandpass filter to the vibration signal to retain the characteristic frequency band of mechanical faults, and applying a moving average filter to the temperature signal to smooth random fluctuations; data standardization, normalizing the data from each sensor to the same scale; sequence segmentation, dividing continuous data into analysis segments according to a preset length; and multi-sensor data alignment, ensuring accurate matching of data from different sensors on the time axis.
[0042] Optionally, feature extraction is performed on the running data at at least two different time scales, including: extracting features from the running data using three one-dimensional dilated convolutional layers respectively; wherein the dilation factor or convolution kernel of the three one-dimensional dilated convolutional layers increases sequentially. It should be noted that the dilation factor and convolution kernel are two core parameters for constructing dilated convolution (Dilated Convolution / Atrous Convolution). The larger the dilation factor or convolution kernel, the larger the receptive field of the dilated convolution, and the longer the time scale for feature extraction.
[0043] For example, please refer to Figure 2 The three one-dimensional dilated convolutional layers can be short-term mode, medium-term mode, and long-term mode, respectively.
[0044] The parameters of the short-term convolutional layer include: kernel size: Kernel_size = 3 (in a single calculation, it will focus on 3 consecutive time points in the input sequence), dilation factor: Dilation_rate = 1 (no skipping, focusing on the local relationship of adjacent time points, capturing high-frequency details. The dilation factor defines the interval between internal parameters of the convolutional kernel when processing input data. It greatly expands the receptive field of the convolutional layer without increasing the number of parameters), number of convolutional layers: 2 layers (each branch consists of two consecutive one-dimensional convolutional layers with the same dilation factor. The first convolutional layer extracts primary features from the original input sequence X and outputs feature map F1. The second convolutional layer takes F1 as input and further extracts more abstract and higher-level composite features based on the primary features, outputting feature map F2), activation function: ReLU (Rectified Linear Unit), whose mathematical expression is f(x) = max(0, x). It sets all negative input values to 0, while outputting positive input values as is. The output feature map shape is: [Batch Size, L, C_out], where Batch Size refers to the number of samples input into the model for processing at one time (e.g., 32, 64), parallel processing to improve training and inference efficiency, L (sequence length): the length of the output feature map in the time dimension (by "padding" before and after each convolutional layer (Padding='same'), the output sequence length L can be kept consistent with the input sequence length, ensuring that the three branches and the attention module are aligned in the time dimension, so that subsequent splicing operations can be performed), C_out (number of output channels): this is determined by the number of filters in the convolutional layer (each filter is designed to extract a specific feature pattern from the input data, for example, one filter may be specifically responsible for detecting "rapid temperature rise", and another may be responsible for detecting "stable current". C_out is a hyperparameter that defines how many different feature patterns this branch should learn. The C_out of the three branches can be the same or different, but for the sake of design simplicity, it is usually set to the same value).
[0045] The short-term convolutional layer has the smallest receptive field and focuses on capturing high-frequency, instantaneous features, such as the instantaneous current spike when the compressor starts or the instantaneous temperature fluctuation caused by the door opening and closing. Its dilation factor is 1, which is the standard convolution, and it can analyze the changes at each time point and its neighboring points in detail.
[0046] The parameters of the mid-mode convolutional layer include: kernel size: Kernel_size = 3, dilation factor: Dilation_rate = 4 (a moderate jump, allowing it to overcome short-term noise and jitter, and see the relationship between several key nodes within a defrost cycle (which may last for several hours), number of convolutional layers: 2, activation function: ReLU, and output feature map shape: [BatchSize, L, C_out].
[0047] Mid-term mode convolutional layers have a moderate receptive field, enabling them to skip short-term fluctuations and capture more periodic patterns. For example, the waveform of temperature rising and falling during a complete defrost cycle, or the complete cycle of a compressor running and stopping. The dilation factor of 4 allows the convolutional kernel to have a 3-point interval in the middle during sampling, thus covering a wider time range.
[0048] The parameters for the long-term mode convolutional layer include: kernel size: Kernel_size = 3, dilation factor: Dilation_rate = 16 (a large jump, enabling it to connect points in time that are far apart (e.g., temperature at the same time every day) to capture long-term trend changes in days or weeks), number of convolutional layers: 2, activation function: ReLU, and output feature map shape: [Batch Size, L, C_out].
[0049] Long-term mode convolutional layers have the largest receptive field, used to identify slowly developing long-term trends. For example, a slow decline in overall cooling efficiency over days or weeks due to a minor refrigerant leak, or the slow accumulation of vibration energy during fan bearing lubrication failure.
[0050] Correspondingly, multiple target feature information is fused to obtain fused feature information. This involves concatenating the feature maps output from the three convolutional layers (all with shapes [Batch Size, L, C_out]) in the last dimension (channel dimension). This concatenation preserves all original information from the three branches without any loss of information, providing the richest raw material for subsequent attention mechanisms. For example, the fusion operation can be `Concat([Feature_short,Feature_mid, Feature_long], axis=-1)`, resulting in a fused feature map with shape [Batch Size, L, 3 * C_out]. This tensor fuses comprehensive temporal information from instantaneous to long-term, providing a rich feature foundation for subsequent diagnosis.
[0051] Optionally, the feature extraction process is carried out in three dimensions: time-domain feature extraction includes the calculation of mean, variance, peak value, and peak-to-peak value; frequency-domain feature extraction analyzes the energy proportion of the main frequency components through fast Fourier transform; and time-series feature extraction includes the calculation of autocorrelation coefficient and slope of the trend. Finally, data standardization is performed, and the z-score normalization method is used to eliminate the influence of data from different sensors.
[0052] Optionally, the fused feature information is adaptively weighted based on an attention mechanism, including: adaptively weighting the fused feature information through a channel attention module to obtain fault channel feature information; adaptively weighting the fused feature information through a time attention module to obtain fault time feature information; and concatenating the fault channel feature information and the fault time feature information to obtain weighted feature information.
[0053] The channel attention module is used to evaluate the importance of different feature channels. For example, among features extracted from multiple scales, it determines whether the current vibration peak feature or the average temperature trend feature is more important. It automatically suppresses feature channels that are irrelevant to the current fault or have high noise levels, while enhancing feature channels with strong discriminative power. For instance, when diagnosing electrical faults, the weight of current-related feature channels may be automatically increased, while the weight of vibration features may be decreased, making feature selection more intelligent. When a sensor fails or the data quality is poor, the weight of its corresponding feature channel is automatically reduced, and the system relies on other reliable channels to make judgments, enhancing fault tolerance and improving model robustness.
[0054] The time attention module is used to assess the importance of different time segments. For example, when analyzing operational data, it determines whether the data at the moment of compressor startup is more critical or the data during stable operation. It can automatically focus on the critical time points or periods when a fault occurs. For example, for intermittent faults, it can "notice" the data during the few seconds when the fault occurs, while ignoring long-term normal operating data, accurately locating critical events. Simultaneously, it can determine whether short-term high-frequency characteristics or long-term trend characteristics are more important for the current fault, and accordingly assign different attention weights to features at different time scales.
[0055] By concatenating channel-weighted and time-weighted feature information to form the final weighted feature information, information complementarity and enhancement are achieved. Channel attention emphasizes "which features are useful," while time attention emphasizes "when features are useful." This concatenation achieves a synergistic effect of "focusing on the right features at the right time," resulting in an information enhancement effect of "1+1>2." The final features simultaneously contain importance information from both the feature dimension and the time dimension, providing a more comprehensive and multi-dimensional decision-making basis for the classification network.
[0056] For example, obtaining faulty channel feature information by adaptively weighting the fused feature information through a channel attention module can include the following steps: First, perform global average pooling on the feature map X in the time dimension L, compressing the feature sequence of each channel into a global statistic, with an output shape of [Batch Size, 1, 3 * C_out]. Second, perform dimensionality reduction through a fully connected layer (or 1x1 convolution) to reduce computation; the dimensionality reduction coefficient is typically 16 (i.e., the number of output channels is 3 * C_out / 16), followed by a ReLU activation function. Then, restore the original number of channels to 3 * C_out through another fully connected layer (or 1x1 convolution). Finally, use the Sigmoid function to map the weights of each channel to between 0 and 1. The final output is a channel attention weight vector W_c with a shape of [Batch Size, 1, 3 * C_out]. Weighting: Multiply the original feature map X by the channel weights W_c: X_c = X * W_c.
[0057] For example, obtaining fault time feature information by adaptively weighting the fused feature information through a temporal attention module can include the following steps: First, the feature map X is compressed into channels using a 1x1 convolution to integrate the information. Then, a one-dimensional convolution (kernel size can be 3) and the Softmax function are used to calculate the combined importance of all channels at each time point, obtaining the weight for each time point. Finally, the temporal attention weight vector W_t is output, with shape [Batch Size, L, 1]. Weighting: The original feature map X is multiplied by the temporal weight W_t: X_t = X * W_t.
[0058] For example, concatenating fault channel feature information and fault time feature information to obtain weighted feature information may include the following steps: adding or averaging the channel weighted feature X_c and the time weighted feature X_t to obtain the final feature representation X_attented after attention weighting, with the same shape as X [Batch Size, L, 3 * C_out].
[0059] Optionally, fault identification is performed on the weighted feature information based on a preset classification network, including: flattening the weighted feature information and global pooling to obtain a feature vector; inputting the feature vector into at least one fully connected layer to obtain an integrated vector; inputting the integrated vector into a Softmax function to obtain a probability distribution vector; and obtaining fault information based on the probability distribution vector.
[0060] By flattening features and performing global pooling, the "weighted feature information" output by the attention mechanism, which may have multi-dimensional structures (such as [batch, channel, time]), is transformed into a fixed-length one-dimensional feature vector. This simplifies the design of subsequent fully connected layers, allowing them to avoid handling variable-sized inputs. Global average pooling averages each feature channel across the entire time dimension. This forces the network to compress key fault information and reflect it in the global statistics of each channel during the learning process, helping to extract more robust and representative features and avoiding overfitting to noise at a local time point. Compared to directly inputting the flattened high-dimensional feature vector into the fully connected layer (with huge parameters), pooling first greatly reduces the dimensionality of the feature vector, thereby significantly reducing the number of parameters in subsequent fully connected layers, improving computational efficiency, and reducing the risk of overfitting. At the same time, the fully connected layer acts as an "advanced inference engine," responsible for learning the complex nonlinear combination relationships between different global features, performing high-order feature integration for the final classification. The Softmax function transforms the "score" output by the fully connected layer into a probability distribution, and then parses the specific fault type and probability from the probability distribution.
[0061] For example, the weighted feature information is flattened and globally pooled to obtain a feature vector; the feature vector is input into at least one fully connected layer to obtain an integrated vector; the integrated vector is input into a Softmax function to obtain a probability distribution vector; fault information is obtained based on the probability distribution vector, which may include the following steps: flattening the feature map X_attented into a one-dimensional vector. Specifically, the operation is Flatten(), and the output shape becomes [Batch Size, L * 3 * C_out]. The flattened vector is input into one or more fully connected layers for nonlinear transformation and feature integration. The number of neurons in the last fully connected layer is equal to the number of fault categories N_classes (e.g., normal, refrigerant leak, fan jam, compressor failure, then N_classes=4); the Softmax function is applied to the output of the fully connected layer to normalize the output value into a probability distribution vector, in which the value of each element is between 0 and 1, and the sum of all elements is 1. The example output is [0.02, 0.85, 0.10, 0.03], which means that the model judges the input sequence to have a 2% probability of being "normal", an 85% probability of being "fan stuck", a 10% probability of being "refrigerant leak", and a 3% probability of being "compressor malfunction".
[0062] Optionally, it also includes: if the maximum value of the failure probability is greater than the preset probability, then the failure type corresponding to the maximum failure probability is determined as the target failure.
[0063] That is, by comparing the obtained multiple fault probabilities, the highest fault probability is obtained. If the highest fault probability is greater than the preset probability, it means that the fault with the highest predicted probability is very likely to occur, so the fault is directly identified as the target fault.
[0064] Of course, as an alternative implementation, all fault types corresponding to fault probabilities greater than the preset probability can be identified as target faults. In this case, there may be multiple target faults, covering more possibilities.
[0065] The preset probability can be an empirical value obtained from multiple experiments. Its specific value can be adjusted according to the specific test conditions or usage environment. For example, when the system's fault tolerance is low, the preset probability can be relatively small, and when the system's fault tolerance is high, the preset probability can be relatively large.
[0066] Optionally, it also includes sending a target fault warning to the user's mobile device. The mobile device may include a smartphone, tablet, computer, or smart speaker, etc. The warning information may include the fault type, fault probability and cause analysis, and suggested solutions.
[0067] By sending early warning information to the user's terminal, users can take corresponding measures in advance to suppress the development of the fault, move the response node forward, and avoid the need for downtime maintenance and delays in use.
[0068] Optionally, before performing feature extraction on the operating data at at least two different time scales, the method further includes: determining whether the refrigerator meets diagnostic trigger conditions; diagnostic trigger conditions include one or more of the following: the refrigerator's operating time reaches a preset duration, the operating data reaches a preset quality, and some operating data exceeds a preset threshold. The preset duration can be 3 hours, 4 hours, 5 hours, etc., in routine testing; it automatically increases to once every 30 minutes when abnormal signs are detected in high-frequency monitoring; and once every hour within 72 hours after maintenance. The preset quality can be that the acquired operating data reaches a preset time length (i.e., the data segment is complete); and some operating data exceeding the preset threshold can be detected when a single parameter exceeds a safety threshold at the edge, or when an abnormal usage pattern is detected (including frequent door opening and extreme temperature settings, etc.).
[0069] To save system resources, fault diagnosis can be performed only when the refrigerator meets the diagnostic trigger conditions, thereby reducing energy consumption and computing costs.
[0070] Furthermore, diagnostic triggering conditions can also include manual triggering. Manual triggering can include two types of operation interfaces. One is management interface triggering, which allows operation and maintenance personnel to manually initiate instant diagnosis through the web interface. It supports backtracking analysis for specified historical time periods and allows adjustment of diagnostic sensitivity parameters, providing three modes: conservative, standard, and sensitive. The other is API interface triggering, which supports integration calls from third-party systems through the interface. It adopts a request-response model, synchronously returning diagnostic identifiers and asynchronously pushing detailed results.
[0071] Optionally, it also includes: model loading and initialization. Specifically, the system first loads the pre-trained model configuration and weight parameters corresponding to the device model from the model library. The model library supports multiple versions coexisting and A / B testing, and supports hot reloading updates without restarting the service. Depending on the refrigerator model, the system will load the corresponding optimized model, instantiate the model structure, load the weight parameters, and set the model to inference mode.
[0072] Optionally, it also includes: confidence assessment and early warning decision-making. Specifically, it can be based on a multi-level threshold system. The system sets differentiated threshold levels for different fault types, including critical threshold, high threshold, medium threshold, and low threshold. Based on the threshold range in which the confidence of the diagnostic result falls, the system determines the corresponding early warning level: critical early warning, high priority early warning, medium early warning, or only recording without alarming. The decision-making process also considers the historical status of the equipment, and appropriately increases the early warning level for equipment that frequently alarms.
[0073] Optionally, this also includes: health score calculation. Specifically, a weighted comprehensive evaluation method can be used. First, different weight coefficients are assigned according to the severity of the fault type, with severe faults such as refrigerant leaks given higher weights. The base health score is 100 minus the sum of the products of each fault probability and its weight. Then, the historical trend of health is considered, and a trend penalty is applied to cases of rapid decline. Finally, adjustments are made in conjunction with equipment usage intensity factors to generate the final health score, and the health level is determined based on the score range, generating corresponding maintenance recommendations.
[0074] For example, the entire probability vector P = [p_normal, p_fault1, p_fault2, ...] is input into the health score function. A simplified calculation method is: Health_Score = (1 - Σ(p_fault_i * weight_i)) * 100. Here, weight_i represents the severity weight of different faults. This score can intuitively display the overall health status of the refrigerator.
[0075] Optionally, it also includes storing fault information and operational data. Specifically, a triple storage system can be established. A real-time database stores the latest diagnostic results (fault information), supporting fast querying and real-time monitoring; a time-series database records complete historical diagnostic data, supporting trend analysis and statistical queries; and a relational database stores business-related structured data, including equipment information, user data, and work order records.
[0076] Optionally, warning information can be sent in tiers, specifically, different notification mechanisms can be activated based on the warning level. Critical warnings trigger notifications across all channels, including SMS, push notifications, email, and telephone; high-priority warnings enable push notifications and email notifications; medium-priority warnings only send email notifications; low-level warnings are only recorded and do not trigger proactive notifications. The notification content includes device information, fault type, confidence level, health score, and recommended handling measures.
[0077] Furthermore, the system can automatically generate maintenance work orders for medium- to high-level warnings. It determines work order priority based on fault type, assigns corresponding maintenance teams based on equipment location, estimates repair time and necessary spare parts, and attaches detailed diagnostic data to the work order. Simultaneously, it updates the equipment status to "requires maintenance" to ensure subsequent tracking and processing.
[0078] Furthermore, a complete API interface system can be provided, supporting integration with existing operation and maintenance work order systems, customer relationship management systems, and enterprise resource planning systems. The web management interface offers rich visualization functions, including real-time status display, historical trend analysis, health status change curves, and diagnostic report downloads. The mobile application supports real-time alarm push and remote monitoring functions, forming a complete equipment management ecosystem.
[0079] Optionally, data transmission can employ a dual-channel transmission mechanism: the primary channel uses the MQTT protocol to adapt to low-bandwidth, high-latency network environments; the backup channel uses an HTTP REST interface for transmitting large data blocks. All transmitted data is protected by TLS / SSL encryption and undergoes device-level authentication. A three-tiered storage system is established in the cloud: a Redis database caches high-frequency data from the past 7 days, an InfluxDB time-series database stores long-term historical data, and a MySQL relational database stores device metadata, diagnostic results, and user information.
[0080] This application also provides a refrigerator control device; please refer to [link / reference]. Figure 3The device includes an operation data acquisition module 201, a feature extraction module 202, a fusion module 203, a weighting module 204, and an analysis module 205. The operation data acquisition module 201 is configured to acquire the refrigerator's operation data; the feature extraction module 202 is configured to extract features from the operation data at at least two different time scales to obtain multiple target feature information; the fusion module 203 is configured to fuse multiple target feature information to obtain fused feature information; the weighting module 204 is configured to perform adaptive weighting processing on the fused feature information based on an attention mechanism to obtain weighted feature information; and the analysis module 205 is configured to perform fault identification on the weighted feature information based on a preset classification network to obtain fault information, which includes fault type and fault probability.
[0081] By implementing the above method, multi-granularity information of time series data is captured, avoiding misjudgments or omissions caused by data from a single time scale and ensuring the comprehensiveness of data capture. Simultaneously, an attention mechanism is introduced to achieve automatic learning and assign higher weights to fault-related features, amplifying fault signals and suppressing noise interference. In other words, by extracting features from operational data at different time scales and introducing an attention mechanism, subtle fault signals are captured and amplified in the early stages of a fault, enabling observation of fault trends and thus achieving early fault diagnosis.
[0082] This application also provides an electronic device 300, please refer to... Figure 4 The system includes a memory 301, a processor 302, and a computer program 3011 stored in the memory 301 and executable on the processor 302. When the processor 302 executes the computer program 3011, it implements the refrigerator control method described above. The method includes the following steps: S101: Acquire refrigerator operating data. S102: Extract features from the operating data at at least two different time scales to obtain multiple target feature information. S103: Fuse the multiple target feature information to obtain fused feature information. S104: Perform adaptive weighting processing on the fused feature information based on an attention mechanism to obtain weighted feature information. S105: Perform fault identification on the weighted feature information based on a preset classification network to obtain fault information, which includes fault type and fault probability.
[0083] This application embodiment also provides a storage medium storing control instructions. When the control instructions are executed by a processor, they implement the refrigerator control method described above. The method includes the following steps: S101: Acquire refrigerator operating data. S102: Extract features from the operating data at at least two different time scales to obtain multiple target feature information. S103: Fuse the multiple target feature information to obtain fused feature information. S104: Perform adaptive weighting processing on the fused feature information based on an attention mechanism to obtain weighted feature information. S105: Perform fault identification on the weighted feature information based on a preset classification network to obtain fault information, which includes fault type and fault probability.
[0084] This application embodiment also provides a refrigerator, which includes a controller configured to execute the refrigerator control method described above. The method includes the following steps: S101: Acquiring refrigerator operating data. S102: Extracting features from the operating data at at least two different time scales to obtain multiple target feature information. S103: Fusing the multiple target feature information to obtain fused feature information. S104: Performing adaptive weighting processing on the fused feature information based on an attention mechanism to obtain weighted feature information. S105: Performing fault identification on the weighted feature information based on a preset classification network to obtain fault information, which includes fault type and fault probability. The refrigerator structure can be found in the attached document. Figure 5 .
[0085] For example, a computer program can be divided into one or more modules / units, which are stored in memory and executed by a processor to perform the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device.
[0086] Electronic devices can be desktop computers, laptops, handheld computers, and cloud servers, among other electronic devices. Electronic devices may include, but are not limited to, processors and memory. For example, electronic devices may also include input / output devices, network access devices, buses, etc.
[0087] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.
[0088] In the embodiments provided by this invention, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.
[0089] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.
[0090] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0091] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0092] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more features.
[0093] The control method, control device, electronic device, and storage medium for a refrigerator provided in the embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for controlling a refrigerator, characterized in that, The method includes: Obtain the operating data of the refrigerator; At least two different time scales are used to extract features from the running data to obtain multiple target feature information; By fusing the multiple target feature information, fused feature information is obtained; The fused feature information is adaptively weighted based on an attention mechanism to obtain weighted feature information; Based on a preset classification network, the weighted feature information is used to identify faults, and the fault information includes fault type and fault probability.
2. The refrigerator control method according to claim 1, characterized in that, The feature extraction of the running data at at least two different time scales includes: The running data is feature extracted using three one-dimensional dilated convolutional layers respectively; The expansion factor or convolution kernel of the three one-dimensional hollow convolutional layers increases sequentially.
3. The refrigerator control method according to claim 1, characterized in that, The adaptive weighting processing of the fused feature information based on the attention mechanism includes: The fused feature information is adaptively weighted using a channel attention module to obtain fault channel feature information; The fused feature information is adaptively weighted using a time attention module to obtain fault time feature information; The fault channel feature information and the fault time feature information are concatenated to obtain the weighted feature information.
4. The refrigerator control method according to claim 1, characterized in that, The fault identification is performed on the weighted feature information based on the preset classification network to obtain fault information, including: The weighted feature information is flattened and globally pooled to obtain the feature vector; The feature vector is input into at least one fully connected layer to obtain the integrated vector; The integrated vector is input into the Softmax function to obtain the probability distribution vector; The fault information is obtained based on the probability distribution vector.
5. The refrigerator control method according to claim 1, characterized in that, Also includes: If the maximum value of the fault probability is greater than the preset probability, then the fault type corresponding to the maximum fault probability is determined as the target fault.
6. The refrigerator control method according to claim 5, characterized in that, Also includes: The target fault generates an early warning message and sends it to the user's mobile device.
7. The refrigerator control method according to claim 1, characterized in that, Before performing feature extraction on the running data at at least two different time scales, the method further includes: Determine whether the refrigerator has met the diagnostic trigger conditions; The diagnostic triggering conditions include one or more of the following: the refrigerator's running time reaches a preset duration, the running data reaches a preset quality, and some of the running data exceeds a preset threshold.
8. A control device for a refrigerator, characterized in that, The device includes: The operation data acquisition module is configured to acquire the operation data of the refrigerator; The feature extraction module is configured to perform feature extraction on the running data at at least two different time scales to obtain multiple target feature information; The fusion module is configured to fuse the multiple target feature information to obtain fused feature information; The weighting module is configured to perform adaptive weighting processing on the fused feature information based on an attention mechanism, and obtain weighted feature information; The analysis module is configured to perform fault identification on the weighted feature information based on a preset classification network to obtain fault information, which includes fault type and fault probability.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the refrigerator control method as described in any one of claims 1-7.
10. A storage medium, characterized in that, The storage medium stores control instructions, which, when executed by a processor, implement the refrigerator control method as described in any one of claims 1-7.