LSTM-based temperature anomaly trend early warning method, device, equipment and storage medium

Abstract

This application relates to the field of LSTM-based temperature anomaly trend early warning technology, disclosing a method, device, equipment, and storage medium for LSTM-based temperature anomaly trend early warning. The method includes: constructing an input vector that integrates multi-dimensional temporal features and using an LSTM prediction model for nonlinear inference, which can more accurately characterize the dynamics of complex thermal processes, and issuing an alarm through a preset alarm method. The beneficial technical effects of this invention are: achieving a shift from delayed alarm to pre-warning, significantly improving the timeliness and reliability of the warning; and the lightweight model, specially optimized and quantized, is small in size and highly computationally efficient, enabling real-time deployment and low-power operation in resource-constrained temperature control equipment, providing a feasible technical solution for high-precision, low-cost temperature anomaly prediction at the edge, such as in smart home appliances and the Industrial Internet of Things.

Description

Technical Field

[0001] This invention relates to the field of temperature anomaly trend early warning technology based on LSTM, and in particular to a temperature anomaly trend early warning method, apparatus, device and storage medium based on LSTM. Background Technology

[0002] Traditional temperature early warning methods, mostly based on fixed thresholds or simple linear regression models, have certain limitations. First, alarm mechanisms based on fixed thresholds are inherently lagging. Due to the thermal inertia of temperature transmission, by the time the system detects that the temperature at a certain point exceeds the preset safety threshold, the core target area (such as the center of food or the interior of a chemical reactor) has often already undergone irreversible overheating, and damage has already occurred. This "post-event alarm" mode cannot achieve effective preventative intervention in temperature-sensitive scenarios such as cooking, fermentation, and fine chemical processes, thus losing its core value as an early warning system. Second, while prediction methods using simple models such as moving averages or linear regression can reflect trends to some extent, they are difficult to accurately characterize and predict nonlinear thermodynamic processes in complex environments. Actual temperature changes are often influenced by multiple factors, including environmental disturbances, changes in material properties, and fluctuations in equipment power, exhibiting significant nonlinear and time-varying characteristics. Linear models are insufficient in fitting such complex dynamics, leading to large prediction biases, frequent false alarms and missed alarms, and low reliability. Summary of the Invention

[0003] Therefore, it is necessary to propose a method, device, equipment, and storage medium for early warning of temperature anomalies based on LSTM to address the existing problem of early warning of temperature anomalies based on LSTM.

[0004] A method for early warning of temperature anomalies based on LSTM, the method comprising: Periodically collect temperature data from designated temperature control devices to form a temperature dataset; Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; The dimensional values ​​of each preset dimension are arranged in a preset dimensional order to obtain a multidimensional feature vector. The multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered using a preset alarm method.

[0005] Furthermore, before the step of inputting the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk, the method further includes: Acquire multiple sets of raw sample data; wherein each set of raw sample data includes raw temperature time series data and a corresponding risk label, wherein the risk label is normal or abnormal; Extract the dimensional values ​​of multiple preset dimensions of each original temperature time series and arrange them in the order of the dimensions to generate a historical multidimensional feature vector corresponding to each original temperature time series. Combine each historical multidimensional feature vector with the corresponding risk label to obtain the target sample set. An initial long short-term memory network model is constructed, and the target sample set is divided into a training set and a validation set. Using the training set data, with historical multidimensional feature vectors as input and corresponding risk labels as supervision targets, the backpropagation algorithm and Adam optimizer are used to iteratively optimize the model parameters. The training process is monitored through the validation set to prevent overfitting until the initial long short-term memory network model converges, thus obtaining the LSTM prediction model.

[0006] Further, after the steps of constructing an initial long short-term memory network model, dividing the target sample set into a training set and a validation set, using the training set data, taking historical multi-dimensional feature vectors as input and corresponding risk labels as supervision targets, iteratively optimizing the model parameters using the backpropagation algorithm and the Adam optimizer, and monitoring the training process through the validation set to prevent overfitting, until the initial long short-term memory network model converges to obtain the LSTM prediction model, the method further includes: The model weights and activation values ​​of the LSTM prediction model are optimized using post-training static quantization techniques to obtain the optimized LSTM prediction model.

[0007] Furthermore, after the step of determining whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold, the method further includes: If the probability value of the abnormal temperature risk is less than or equal to a preset first risk threshold, then it is determined whether the probability value of the abnormal temperature risk is greater than a preset second risk threshold; wherein, the second risk threshold is less than the first risk threshold. If the probability value of the abnormal temperature risk is greater than the preset second risk threshold, then the temperature change rate is calculated based on the temperature data. Determine whether the temperature change rate is greater than a preset change rate; If the temperature change rate is greater than the preset change rate, an alarm will be triggered using the preset alarm method.

[0008] Furthermore, the multidimensional feature vector includes the average temperature, temperature variance, temperature change slope, current ambient temperature, and target object type encoding within the historical time window.

[0009] Further, the step of inputting the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk includes: The multidimensional feature vector is quantized from floating-point format to fixed-point format to obtain the quantized feature vector; The quantized feature vector is input into the LSTM prediction model for fixed-point inference to obtain the quantized output result. The quantization output result is dequantized into a floating-point number format of the temperature anomaly risk probability value.

[0010] Furthermore, after the step of triggering an alarm via a preset alarm method if the probability value of the abnormal temperature risk is greater than a preset first risk threshold, the method further includes: Based on the alarm, obtain the corresponding alarm feedback information and collect it into the preset feedback database; Determine whether the alarm feedback information in the feedback database has reached a preset quantity; If the number of alarm feedback messages in the feedback database reaches a preset number, the parameters in the LSTM prediction model are adjusted based on the alarm feedback messages in the feedback database to obtain an optimized LSTM prediction model.

[0011] A temperature anomaly trend early warning device based on LSTM, the device comprising: The data acquisition module is used to periodically acquire temperature data from a specified temperature control device to form a temperature dataset. The calculation module is used to calculate the dimension values ​​of multiple preset dimensions based on the temperature dataset; The arrangement module is used to arrange the dimensional values ​​of each preset dimension according to the preset dimensional order to obtain a multidimensional feature vector. The inference module is used to input the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk. The first judgment module is used to determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold. The first alarm module is used to issue an alarm in a preset alarm mode if the probability value of the abnormal temperature risk is greater than a preset first risk threshold.

[0012] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Periodically collect temperature data from designated temperature control devices to form a temperature dataset; Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; The dimensional values ​​of each preset dimension are arranged in a preset dimensional order to obtain a multidimensional feature vector. The multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered using a preset alarm method.

[0013] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Periodically collect temperature data from designated temperature control devices to form a temperature dataset; Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; The dimensional values ​​of each preset dimension are arranged in a preset dimensional order to obtain a multidimensional feature vector. The multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered using a preset alarm method.

[0014] The beneficial effects of this invention are as follows: By constructing an input vector that integrates multi-dimensional temporal features and using an LSTM prediction model for nonlinear inference, it is possible to more accurately characterize the dynamics of complex thermal processes, realizing the transformation from "delayed alarm" to "early warning," significantly improving the timeliness and reliability of the warning. At the same time, this lightweight model has been specially optimized and quantized, resulting in a small size and high computational efficiency. It can be deployed in real time and run with low power consumption in resource-constrained temperature control equipment, providing a feasible technical solution for achieving high-precision and low-cost temperature anomaly prediction at the edge, such as in smart home appliances and industrial IoT. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] in: Figure 1This is an application environment diagram of a temperature anomaly trend early warning method based on LSTM in one embodiment; Figure 2 This is a flowchart of a temperature anomaly trend early warning method based on LSTM in one embodiment; Figure 3 This is a structural block diagram of a temperature anomaly trend early warning device based on LSTM in one embodiment; Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Figure 1 This is a diagram illustrating an LSTM-based temperature anomaly trend early warning application environment in one embodiment. (Refer to...) Figure 1 This LSTM-based temperature anomaly trend early warning method is applied to an LSTM-based temperature anomaly trend early warning system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to collect temperature data from designated temperature control equipment, and the server 120 is used to calculate the probability value of temperature anomaly risk.

[0019] like Figure 2 As shown, in one embodiment, a temperature anomaly trend early warning method based on LSTM is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The LSTM-based temperature anomaly trend early warning method specifically includes the following steps: S1: Periodically collect temperature data from a specified temperature control device to form a temperature dataset; S2: Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; S3: Arrange the dimensional values ​​of each preset dimension in a preset dimensional order to obtain a multidimensional feature vector; S4: Input the multidimensional feature vector into the preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; S5: Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; S6: If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered through the preset alarm method.

[0020] As described in step S1 above, temperature data from designated temperature control devices is periodically collected to form a temperature dataset. Setting up and periodically collecting temperature data from designated temperature control devices involves accurately acquiring temperature readings at different time intervals (e.g., every minute, every five minutes, etc.). Temperature control devices can be various types of equipment such as smart ovens, refrigerators, and fermentation tanks. Temperature changes directly affect the quality of the final product; therefore, the collected temperature data should consider the usage scenario and characteristics of the equipment to ensure the comprehensiveness and accuracy of the collected data. After collecting data at multiple time points, all data will be aggregated into a single temperature dataset, providing a basis for subsequent analysis. Furthermore, to ensure the effectiveness of data collection, temperature sensors must be calibrated periodically to guarantee the accuracy of their readings. Sensors should be protected from external environmental influences or equipment malfunctions during operation. Through precise temperature monitoring, the system can track temperature changes in real time, providing crucial information for subsequent analysis and early warning.

[0021] As described in step S2 above, multiple preset dimension values ​​are calculated based on the temperature dataset. These dimension values ​​are derived through statistical analysis and are primarily used to describe various features in the temperature dataset, such as the average temperature, variance, slope, maximum value, and minimum value. By calculating these dimensions, the system can effectively extract key features of temperature changes, thus providing rich input information for subsequent steps. Specifically, the average temperature can represent the overall level, while the variance reflects the amplitude and stability of temperature fluctuations; the slope indicates the trend of temperature changes. Furthermore, additional feature dimensions can be specified according to specific scenarios, such as relative humidity, ambient temperature, and item type coding, aiming to comprehensively characterize the dynamic changes in the temperature environment. Through such flexible and diverse feature extraction, the resulting dimension values ​​will lay a solid foundation for efficient inference in the subsequent model.

[0022] As described in step S3 above, the dimensional values ​​of each preset dimension are arranged according to a preset dimensional order to obtain a multidimensional feature vector. The arrangement of the calculated dimensional values ​​according to the preset dimensional order forms a multidimensional feature vector. This specific order is crucial for the model's understanding of the input information, as many machine learning models, especially deep learning models, rely on the order and structure of input features to extract latent information. When determining the feature order, the highest priority dimension can be selected based on domain knowledge and practical application needs, such as inputting the average temperature first, then the rate of change, and finally the ambient temperature. Furthermore, to improve the efficiency of model inference, this feature vector can be organized hierarchically, for example, grouping related features into one category and using principal component analysis (PCA) for dimensionality reduction. This process not only improves computational efficiency but also ensures that the model can effectively capture the most critical temperature change features, laying the foundation for the final anomaly risk prediction.

[0023] As described in step S4 above, the multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk. The constructed multidimensional feature vector is input into the preset LSTM prediction model to infer the risk of temperature anomalies. LSTM prediction models can run efficiently in resource-constrained environments, and their structure typically has fewer parameters and computational complexity, making them suitable for rapid inference execution on embedded devices. The model can utilize historical information from time-series data for prediction. After inputting the feature vector, the model will calculate the probability value of temperature anomaly risk through a series of complex matrix operations. The probability value is typically between 0 and 1; the higher the value, the greater the likelihood of a temperature anomaly. The effectiveness and accuracy of this process directly affect the final warning result. Therefore, during the model training phase, it is necessary to ensure that the model is validated with a large amount of data to improve its generalization ability and prediction reliability.

[0024] As described in step S5 above, it is determined whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold. If the risk probability value exceeds the preset first risk threshold, the system considers there to be a potential temperature anomaly risk. When setting the risk threshold, the actual needs of the application scenario and the risk level that the user may tolerate should be comprehensively considered. For example, in food processing, the threshold should be set relatively strictly to ensure food safety, such as 0.6; while in some industrial processes, higher temperature fluctuations can be tolerated, and it can be set to 0.8. Therefore, relevant personnel should set reasonable thresholds based on historical data analysis, expert opinions, and industry standards to ensure effective risk assessment. At the same time, it is worth noting that timely dynamic adjustment of the threshold is also necessary to adapt to changes in different working environments and equipment characteristics.

[0025] As described in step S6 above, if the probability value of the temperature anomaly risk is greater than a preset first risk threshold, an alarm will be triggered using a preset alarm method. If the judgment result shows that the probability value of the temperature anomaly risk exceeds the first risk threshold, the system will trigger the corresponding alarm mechanism to promptly notify relevant personnel to take necessary measures. This alarm method can be diversified, such as issuing audible and visual alarms, pushing SMS messages or service notifications, etc., and the specific implementation can be flexibly selected according to scenario requirements and user habits. Timeliness of alarms is crucial to preventing damage. By monitoring temperature in real time and issuing timely alarms, the system can identify and notify users before or in the early stages of temperature anomalies, thereby reducing potential losses. When designing the alarm mechanism, different levels of alarms can also be introduced, such as minor alarms and severe alarms, to help personnel take different response measures according to the situation. Furthermore, the effectiveness and reliability of the alarm system should be ensured to avoid false alarms, which could lead to unnecessary panic and waste of resources.

[0026] In one embodiment, before step S4, which involves inputting the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk, the method further includes: S301: Obtain multiple sets of raw sample data; wherein each set of raw sample data includes raw temperature time series data and a corresponding risk label, wherein the risk label is normal or abnormal; S302: Extract the dimension values ​​of multiple preset dimensions of each original temperature time series and arrange them in the order of the dimensions to generate a historical multidimensional feature vector corresponding to each original temperature time series, and combine each historical multidimensional feature vector with the corresponding risk label to obtain the target sample set; S303: Construct an initial long short-term memory network model, and divide the target sample set into a training set and a validation set. Using the training set data, with historical multidimensional feature vectors as input and corresponding risk labels as supervision targets, use the backpropagation algorithm and Adam optimizer to iteratively optimize the model parameters. Monitor the training process through the validation set to prevent overfitting until the initial long short-term memory network model converges, and obtain the LSTM prediction model.

[0027] As described in step S301 above, multiple sets of raw sample data are acquired. In step S301, the system first needs to acquire multiple sets of raw sample data, which are the basis for training the LSTM prediction model. Each set of raw sample data contains two main parts: raw temperature time series data and corresponding risk labels. The temperature time series data is a continuous data stream consisting of temperature readings at various time points, while the risk label for each set of data is used to distinguish whether the time series is normal. The label can be "normal" or "abnormal". A normal label indicates that the temperature did not fluctuate abnormally during the time period and did not lead to any negative results; an abnormal label indicates that the temperature fluctuated significantly during the time period, which may lead to excessively high or low temperatures, thus affecting the processing quality. When acquiring raw sample data, temperature can be monitored through the internal sensors of the temperature control device, or it can be extracted from previous records or experimental datasets. Ensuring the authenticity and reliability of the raw sample data is crucial, as it directly affects the effectiveness of subsequent feature extraction and the accuracy of model training.

[0028] As described in step S302 above, the dimensional values ​​of multiple preset dimensions of each original temperature time series are extracted and arranged in the dimensional order to generate a historical multidimensional feature vector corresponding to each original temperature time series. Each historical multidimensional feature vector is then combined with its corresponding risk label to obtain a target sample set. The acquired original temperature time series data is analyzed, and multiple preset dimensional values ​​are extracted for each time series. These dimensional values ​​may include statistical features such as the average temperature, variance, rate of change, maximum value, and minimum value, which can comprehensively reflect the basic characteristics of temperature dynamics. After extraction, these dimensional values ​​are arranged in a preset order to form a historical multidimensional feature vector. After this operation, the system combines each historical multidimensional feature vector with its corresponding risk label to form a target sample set. The target sample set not only contains historical features of temperature changes but also the risk classifications corresponding to these features, ensuring that the model can both extract information on temperature state changes and effectively classify expected risks during subsequent training.

[0029] As described in step S303 above, an initial Long Short-Term Memory (LSTM) network model is constructed, and the target sample set is divided into a training set and a validation set. Using the training set data, with historical multi-dimensional feature vectors as input and corresponding risk labels as supervision targets, the backpropagation algorithm and Adam optimizer are used to iteratively optimize the model parameters. The training process is monitored through the validation set to prevent overfitting until the initial LSTM network model converges, resulting in the LSTM prediction model. The LSTM network model performs well in processing time series data, and is particularly suitable for dynamic prediction and anomaly detection of temperature data. When constructing the model, the structural parameters of the model are first set, including the number of nodes in the input layer, LSTM layer, and output layer. After preparing the model, the system divides the target sample set into two parts: a training set and a validation set. The training set is used for adjusting the model parameters, while the validation set is used to monitor the model's performance to prevent overfitting. During model training, feature vectors are used as input data, and corresponding risk labels are used as supervision targets. The backpropagation algorithm and Adam optimizer are applied to iteratively optimize the model parameters. The Adam optimizer is a commonly used adaptive learning rate algorithm that intelligently adjusts the learning rate during training, thereby accelerating the convergence process and improving model stability. Continuously monitoring the model's performance on the validation set allows for timely detection of overfitting issues, enabling appropriate adjustments (e.g., adjusting model complexity or introducing regularization strategies). Finally, when the model parameters converge to a stable state, the LSTM prediction model obtained by this method is obtained, which can be used for subsequent inference and early warning of temperature anomaly risks. This step is the core of the entire implementation process, directly determining the model's performance and application effectiveness.

[0030] In one embodiment, after step S303 of constructing an initial long short-term memory network model, dividing the target sample set into a training set and a validation set, using the training set data, taking historical multidimensional feature vectors as input and corresponding risk labels as supervision targets, iteratively optimizing the model parameters using the backpropagation algorithm and the Adam optimizer, and monitoring the training process through the validation set to prevent overfitting, until the initial long short-term memory network model converges and the LSTM prediction model is obtained, the method further includes: S3041: The model weights and activation values ​​of the LSTM prediction model are optimized using post-training static quantization techniques to obtain the optimized LSTM prediction model.

[0031] As described in step S3041 above, post-training static quantization is used to optimize the model weights and activation values ​​of the LSTM prediction model, resulting in an optimized LSTM prediction model. The system implementing this LSTM-based temperature anomaly trend early warning method optimizes the previously trained LSTM prediction model using post-training static quantization. Quantization is a commonly used model optimization technique in deep learning, aiming to reduce the model's memory footprint and computational requirements, making it more suitable for operation on resource-constrained embedded devices. The trained model typically uses 32-bit floating-point numbers to represent weights and activation values, which consumes significant storage space and requires high computational power in actual deployment. Through quantization, the system converts these weights and activation values ​​from 32-bit floating-point numbers to low-precision fixed-point numbers (such as 8-bit fixed-point numbers).

[0032] During quantization, the system uses a calibration dataset to evaluate the model's performance changes before and after quantization. By adjusting quantization parameters, such as the scaling factor and zeros of each layer, it ensures that the quantized output behaves as consistently as possible with the trained model. The model after static quantization not only significantly reduces memory usage but also noticeably improves inference speed, making it ideal for implementing real-time LSTM-based temperature anomaly trend early warning in edge computing and practical applications.

[0033] By employing this static quantization technique, the system can effectively minimize the computational complexity of deep learning models while placing them in smaller embedded devices, ensuring their usability and performance in practical applications. This optimization not only improves the overall system's response speed and operational efficiency but also reduces hardware costs, enabling the widespread adoption and promotion of LSTM-based temperature anomaly trend early warning systems across numerous application fields.

[0034] In one embodiment, after step S5 of determining whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold, the method further includes: S601: If the probability value of the abnormal temperature risk is less than or equal to a preset first risk threshold, then determine whether the probability value of the abnormal temperature risk is greater than a preset second risk threshold; wherein, the second risk threshold is less than the first risk threshold. S602: If the probability value of the abnormal temperature risk is greater than the preset second risk threshold, then calculate the temperature change rate based on the temperature data; S603: Determine whether the temperature change rate is greater than a preset change rate; S604: If the temperature change rate is greater than the preset change rate, an alarm will be triggered using the preset alarm method.

[0035] As described in step S601 above, if the probability value of the temperature anomaly risk is less than or equal to a preset first risk threshold, then it is determined whether the probability value of the temperature anomaly risk is greater than a preset second risk threshold; wherein, the second risk threshold is less than the first risk threshold. Firstly, when the probability value of the temperature anomaly risk obtained through the prediction model is less than or equal to the first risk threshold, it indicates that the current temperature condition is within the normal range and has not reached the safety limit requiring immediate alarm. However, safety levels are often tiered, so in practical applications, it is necessary to further assess whether the risk probability has crossed the second risk threshold. This second risk threshold is set relatively low, mainly for monitoring possible abnormal trends, and has higher sensitivity.

[0036] Comparing the second risk threshold with the first risk threshold involves two judgment steps: if the risk probability is lower than the first threshold, the second threshold judgment is performed, prioritizing potential temperature anomalies. If the risk probability exceeds the second threshold, although it does not meet the severe alarm criteria, it still indicates the possibility of a temperature anomaly. This tiered judgment mechanism makes the system more robust in temperature monitoring, enabling early detection of potential problems and providing time and leeway for effective intervention.

[0037] As described in step S602 above, if the probability value of the temperature anomaly risk is greater than the preset second risk threshold, the temperature change rate is calculated based on the temperature data. If the probability value of the temperature anomaly risk is greater than the preset second risk threshold, the system will continue to the next step, namely, calculating the temperature change rate. The temperature change rate is an important indicator for measuring the speed of temperature change, reflecting the trend and intensity of temperature changes. Generally, the temperature change rate is calculated by differentiating the temperature data over a recent period (calculating the rate of temperature change). In specific implementation, the system can calculate it as follows: First, acquire the temperature readings at the most recent moments, use these data to generate a temperature time series, and then calculate the temperature change rate according to the formula: change rate = ΔTΔt (where ΔT is the difference between the two most recent temperature readings, and Δt is the time interval). The significance of calculating the temperature change rate is that excessively rapid temperature changes often lead to equipment failure or product quality loss; therefore, timely acquisition of this information can help the system to more accurately conduct risk assessment and early warning. If the rate of change exceeds a certain threshold, the system may consider the rapidly changing temperature to be an abnormal situation. Even if the probability model fails to reach the first or second risk threshold, the system can still issue an alarm independently, thereby enhancing the sensitivity and reliability of the alarm system.

[0038] As described in step S603 above, it is determined whether the temperature change rate is greater than a preset change rate. This preset change rate is set based on equipment characteristics and empirical values ​​of the operating environment, and is an important parameter for process safety; for example, it may be set to 3°C / s. For instance, in some cooking or chemical processes, excessively rapid temperature changes can lead to overcooked food or uncontrolled chemical reactions. Therefore, setting a change rate threshold will help the system identify potential risks in a timely manner. If the temperature change rate exceeds this threshold, it may indicate that the system is in a dangerous state, and the monitoring results will have a stronger warning effect in this case.

[0039] As described in step S604 above, if the temperature change rate exceeds a preset change rate, an alarm is triggered using a preset alarm method. When triggering an alarm, the system must select an appropriate alarm method based on the characteristics of the equipment and the usage scenario. Audible and visual alarms are suitable for environments with frequent human-machine interaction, allowing on-site operators to perceive temperature anomalies immediately. In remote monitoring scenarios, the system may prioritize data push notifications to ensure that real-time information about temperature hazards is received regardless of location. Furthermore, more granular alarm levels can be set; for example, different intensities of alarm methods may be used for different levels of risk, thereby achieving a more efficient response mechanism in personalized alarm management.

[0040] In one embodiment, the multidimensional feature vector includes the average temperature, temperature variance, temperature change slope, current ambient temperature, and target object type encoding within a historical time window.

[0041] The average temperature is the arithmetic mean of all temperature readings within a historical time window. Calculating the average temperature provides the system with overall steady-state temperature information, reflecting the overall temperature trend of the equipment or environment during that time period. Specifically, this value is obtained by summing each temperature reading within the time window and then dividing by the number of readings. The average temperature is an important benchmark for temperature control systems and can be used to assess whether the current temperature is within the ideal range.

[0042] Temperature variance describes the fluctuation of temperature readings within a time window. It is calculated by squared the difference between each temperature reading and its average, summing all squared differences, and finally dividing by the total number of readings. Temperature variance reflects not only the degree of temperature fluctuation but also its stability. Excessive variance indicates that the system is facing adverse environmental conditions or operational malfunctions in actual operation.

[0043] The slope of temperature change is the ratio of the temperature change between two points in time to the time interval. The slope provides information on the direction and rate of temperature change, helping the system to detect anomalies promptly and take early warning and intervention measures to prevent problems before they occur, ensuring product safety and stability.

[0044] The current ambient temperature refers to the temperature value of the environment in which the temperature control equipment operates, usually obtained in real time through external environmental sensors. It directly affects the internal temperature balance of the equipment, especially in open systems such as food preservation or chemical reactions. Changes in ambient temperature can affect the equipment's heat load, increasing the burden on heating or cooling systems. When the ambient temperature is higher or lower than the preset safe range, it may cause internal temperature fluctuations, affecting the stability of the equipment and the quality of the final product.

[0045] Target object type coding is a numerical representation that maps different types of items or materials associated with temperature control equipment in a specific way. By creating such identification codes, the system can be provided with important information such as material characteristics, processing requirements, or priorities. For example, food, pharmaceuticals, or chemicals have their own specific requirements for temperature control, and through this coding, the system can intelligently identify the properties of the target object and thus take appropriate temperature control strategies.

[0046] In one embodiment, step S4, which involves inputting the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk, includes: S401: Quantize the multidimensional feature vector from floating-point format to fixed-point format to obtain a quantized feature vector; S402: Input the quantized feature vector into the LSTM prediction model for fixed-point inference to obtain the quantized output result; S403: Dequantize the quantization output result into a floating-point format temperature anomaly risk probability value.

[0047] As described in step S401 above, the multidimensional feature vector is quantized from floating-point format to fixed-point format to obtain a quantized feature vector. The quantization process primarily aims to reduce the storage space and computational resources required by the model during runtime. In deep learning models, especially in low-power and resource-constrained embedded devices, hardware resources are typically limited. Models using floating-point format not only consume a large amount of memory but also have slow computation speed during inference. By converting floating-point numbers to fixed-point numbers, the system can significantly reduce the model size and accelerate inference speed, thus maintaining efficient operation in constrained computational environments. Specifically, the quantization steps include selecting appropriate scaling factors and zeros, which are typically calculated statistically based on training data. The quantization process helps map the range of floating-point numbers to a smaller range of fixed-point numbers. For example, reducing from 32-bit floating-point numbers to 8-bit fixed-point numbers; in this process, the system needs to consider the potential loss of precision after quantization.

[0048] As described in step S402 above, the quantized feature vector is input into the LSTM prediction model for fixed-point inference to obtain a quantized output result. During the inference process, each layer of the model performs a series of calculations based on the quantized feature vector, including weighted summation and activation calculations. Because the input data has been converted to a fixed-point format, fixed-point operations are faster than floating-point operations and require significantly fewer hardware resources. Through this process, the model can effectively identify potential temperature anomaly patterns in the input feature vector and output a quantized result, i.e., a fixed-point representation of the probability of temperature anomaly risk. This quantized output result is important intermediate data required in the subsequent dequantization stage and is crucial for the final operation of the early warning system. By maintaining an efficient inference process, the system can ensure good responsiveness in real-time monitoring, improving the accuracy and timeliness of temperature anomaly monitoring.

[0049] As described in step S403 above, the quantized output result is dequantized into a floating-point format for the temperature anomaly risk probability value. Dequantization is the process of converting a fixed-point number back to a floating-point number so that the user or system can understand the output data. The dequantization process typically involves calculations using the scaling factor and zero point determined during the quantization stage. The formula for this process is: Floating-point output = Quantized output × Scaling factor + Zero point. Through this calculation, the converted output is remapped back to floating-point space, making the temperature anomaly risk probability value more consistent with reality and maintaining readability. The output range of the temperature anomaly risk probability value is typically between 0 and 1, where 0 indicates no temperature anomaly risk and 1 indicates an extremely high temperature anomaly risk.

[0050] In one embodiment, after step S6, which involves triggering an alarm using a preset alarm method if the probability value of the temperature anomaly risk is greater than a preset first risk threshold, the method further includes: S701: Obtain the corresponding alarm feedback information based on the alarm and collect it into the preset feedback database; S702: Determine whether the alarm feedback information in the feedback database has reached a preset quantity; S703: If the number of alarm feedback messages in the feedback database reaches a preset number, the parameters in the LSTM prediction model are adjusted based on the alarm feedback messages in the feedback database to obtain an optimized LSTM prediction model.

[0051] As described in step S701 above, corresponding alarm feedback information is obtained based on the alarm and collected into a preset feedback database. Feedback information typically includes the alarm timestamp, alarm type, alarm level, on-site personnel's response measures, response time, and whether an actual temperature anomaly event occurred. Through this detailed data, the system can not only effectively record the background of each alarm event but also provide a rich source of information for subsequent data analysis. The process of collecting feedback information can be achieved through interaction with on-site personnel, recording by the monitoring system, or automated data acquisition by IoT devices.

[0052] As described in step S702 above, it is determined whether the alarm feedback information in the feedback database has reached a preset quantity. This quantity is set based on specific needs or experience values. It can be a requirement for the accumulation of feedback data within a certain time period, or it can be a quantity related to a specific event; specifically, it can be set to 5000 items. Setting a reasonable feedback quantity threshold ensures that the system collects sufficient data before proceeding with subsequent analysis and model adjustments, avoiding misjudgments or improper optimization due to insufficient data. The purpose of this determination is to ensure the quality and validity of the data. Only when the feedback information reaches the preset quantity can the system perform scientific analysis without being affected by accidental factors. Premature analysis may lead to incorrect conclusions, thus negatively impacting model adjustments.

[0053] As described in step S703 above, if the alarm feedback information in the feedback database reaches a preset quantity, the parameters in the LSTM prediction model are adjusted based on the alarm feedback information in the feedback database to obtain an optimized LSTM prediction model. Parameter adjustment methods may include reconfiguring hyperparameters in the model (such as learning rate, number of network layers, number of hidden nodes, etc.) or adjusting weights to make the model more accurate in processing future input data. The optimization process can employ a backpropagation algorithm, calculating the difference between the feedback information and the prediction results to train the system and reduce future prediction errors. Through this data-feedback-based optimization process, the system not only enhances its sensitivity to current environmental changes but also gradually forms a closed-loop learning mechanism in practical applications, enabling the LSTM-based temperature anomaly trend early warning model to maintain continuous improvement over long-term use, ultimately achieving a more intelligent and accurate early warning function. This not only improves the system's reliability but also effectively promotes user trust, thus laying the foundation for the system's widespread application.

[0054] Reference Figure 3 The present invention also provides a temperature anomaly trend early warning device based on LSTM, the device comprising: The acquisition module 902 is used to periodically acquire temperature data from a specified temperature control device to form a temperature dataset; The calculation module 904 is used to calculate the dimension values ​​of multiple preset dimensions based on the temperature dataset; The arrangement module 906 is used to arrange the dimension values ​​of each preset dimension according to the preset dimension order to obtain a multidimensional feature vector. The inference module 908 is used to input the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk. The first judgment module 910 is used to determine whether the probability value of the temperature abnormality risk is greater than a preset first risk threshold. The first alarm module 912 is used to issue an alarm in a preset alarm mode if the probability value of the abnormal temperature risk is greater than a preset first risk threshold.

[0055] In one embodiment, the LSTM-based temperature anomaly trend early warning device further includes: The raw sample data acquisition module is used to acquire multiple sets of raw sample data; wherein each set of raw sample data includes raw temperature time series data and a corresponding risk label, wherein the risk label is normal or abnormal; The extraction module is used to extract the dimension values ​​of multiple preset dimensions of each original temperature time series and arrange them in the order of the dimensions to generate a historical multidimensional feature vector corresponding to each original temperature time series, and combine each historical multidimensional feature vector with the corresponding risk label to obtain a target sample set. The module is used to construct an initial long short-term memory network model and divide the target sample set into a training set and a validation set. Using the training set data, with historical multidimensional feature vectors as input and corresponding risk labels as supervision targets, the model parameters are iteratively optimized using the backpropagation algorithm and the Adam optimizer. The training process is monitored through the validation set to prevent overfitting until the initial long short-term memory network model converges, thus obtaining the LSTM prediction model.

[0056] In one embodiment, the LSTM-based temperature anomaly trend early warning device further includes: The optimization module is used to optimize the model weights and activation values ​​of the LSTM prediction model using post-training static quantization techniques to obtain the optimized LSTM prediction model.

[0057] In one embodiment, the LSTM-based temperature anomaly trend early warning device further includes: The second judgment module is used to determine whether the temperature anomaly risk probability value is greater than a preset second risk threshold if the temperature anomaly risk probability value is less than or equal to a preset first risk threshold; wherein the second risk threshold is less than the first risk threshold. The temperature change rate calculation module is used to calculate the temperature change rate based on the temperature data if the probability value of the temperature anomaly risk is greater than a preset second risk threshold. The third judgment module is used to determine whether the temperature change rate is greater than a preset change rate; The second alarm module is used to trigger an alarm using a preset alarm method if the temperature change rate is greater than a preset change rate.

[0058] In one embodiment, the multidimensional feature vector includes the average temperature, temperature variance, temperature change slope, current ambient temperature, and target object type encoding within a historical time window.

[0059] In one embodiment, the inference module 908 includes: The quantization feature vector acquisition submodule is used to quantize the multidimensional feature vector from floating-point format to fixed-point format to obtain the quantized feature vector; The quantization output result acquisition submodule is used to input the quantized feature vector into the LSTM prediction model for fixed-point inference to obtain the quantized output result. The dequantization submodule is used to dequantize the quantization output result into the temperature anomaly risk probability value in floating-point format.

[0060] In one embodiment, the LSTM-based temperature anomaly trend early warning device further includes: An alarm feedback information acquisition module is used to acquire corresponding alarm feedback information based on the alarm and collect it into a preset feedback database; The fourth judgment module is used to determine whether the alarm feedback information in the feedback database has reached a preset number; The adjustment module is used to adjust the parameters in the LSTM prediction model based on the alarm feedback information in the feedback database if the number of alarm feedback information in the feedback database reaches a preset number, so as to obtain an optimized LSTM prediction model.

[0061] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an LSTM-based temperature anomaly trend early warning method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the LSTM-based temperature anomaly trend early warning method. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0062] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Periodically collect temperature data from designated temperature control devices to form a temperature dataset; Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; The dimensional values ​​of each preset dimension are arranged in a preset dimensional order to obtain a multidimensional feature vector. The multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered using a preset alarm method.

[0063] By constructing an input vector that integrates multi-dimensional temporal features and using an LSTM prediction model for nonlinear inference, the dynamics of complex thermal processes can be more accurately characterized, realizing the transformation from "delayed alarm" to "early warning," significantly improving the timeliness and reliability of the warning. At the same time, this lightweight model has been specially optimized and quantized, resulting in a small size and high computational efficiency. It can be deployed in real time and run with low power consumption in resource-constrained temperature control equipment, providing a feasible technical solution for high-precision and low-cost temperature anomaly prediction at the edge, such as in smart home appliances and industrial IoT.

[0064] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps: Periodically collect temperature data from designated temperature control devices to form a temperature dataset; Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; The dimensional values ​​of each preset dimension are arranged in a preset dimensional order to obtain a multidimensional feature vector. The multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered using a preset alarm method.

[0065] By constructing an input vector that integrates multi-dimensional temporal features and using an LSTM prediction model for nonlinear inference, the dynamics of complex thermal processes can be more accurately characterized, realizing the transformation from "delayed alarm" to "early warning," significantly improving the timeliness and reliability of the warning. At the same time, this lightweight model has been specially optimized and quantized, resulting in a small size and high computational efficiency. It can be deployed in real time and run with low power consumption in resource-constrained temperature control equipment, providing a feasible technical solution for high-precision and low-cost temperature anomaly prediction at the edge, such as in smart home appliances and industrial IoT.

[0066] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A temperature anomaly trend early warning method based on LSTM, characterized in that, The method includes: Periodically collect temperature data from designated temperature control devices to form a temperature dataset; Calculate the dimensional values ​​of multiple preset dimensions based on the temperature dataset; The dimensional values ​​of each preset dimension are arranged in a preset dimensional order to obtain a multidimensional feature vector. The multidimensional feature vector is input into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk; Determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold; If the probability value of the abnormal temperature risk is greater than the preset first risk threshold, an alarm will be triggered using a preset alarm method. 2.The LSTM-based temperature anomaly trend early warning method of claim 1, wherein, Before the step of inputting the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk, the method further includes: Acquire multiple sets of raw sample data; wherein each set of raw sample data includes raw temperature time series data and a corresponding risk label, wherein the risk label is normal or abnormal; Extract the dimensional values ​​of multiple preset dimensions of each original temperature time series and arrange them in the order of the dimensions to generate a historical multidimensional feature vector corresponding to each original temperature time series. Combine each historical multidimensional feature vector with the corresponding risk label to obtain the target sample set. An initial long short-term memory network model is constructed, and the target sample set is divided into a training set and a validation set. Using the training set data, with historical multidimensional feature vectors as input and corresponding risk labels as supervision targets, the backpropagation algorithm and Adam optimizer are used to iteratively optimize the model parameters. The training process is monitored through the validation set to prevent overfitting until the initial long short-term memory network model converges, thus obtaining the LSTM prediction model. 3.The LSTM-based temperature anomaly trend early warning method of claim 2, wherein, The steps of constructing an initial long short-term memory network model, dividing the target sample set into a training set and a validation set, using the training set data, taking historical multi-dimensional feature vectors as input and corresponding risk labels as supervision targets, employing the backpropagation algorithm and Adam optimizer to iteratively optimize the model parameters, and monitoring the training process through the validation set to prevent overfitting, until the initial long short-term memory network model converges to obtain the LSTM prediction model, further include: The model weights and activation values ​​of the LSTM prediction model are optimized using post-training static quantization techniques to obtain the optimized LSTM prediction model. 4.The LSTM-based temperature anomaly trend early warning method of claim 1, wherein, After the step of determining whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold, the method further includes: If the probability value of the abnormal temperature risk is less than or equal to a preset first risk threshold, then it is determined whether the probability value of the abnormal temperature risk is greater than a preset second risk threshold; wherein, the second risk threshold is less than the first risk threshold. If the probability value of the abnormal temperature risk is greater than the preset second risk threshold, then the temperature change rate is calculated based on the temperature data. Determine whether the temperature change rate is greater than a preset change rate; If the temperature change rate is greater than the preset change rate, an alarm will be triggered using the preset alarm method. 5.The LSTM-based temperature anomaly trend early warning method of claim 1, wherein, The multidimensional feature vector includes the average temperature, temperature variance, temperature change slope, current ambient temperature, and target object type encoding within the historical time window. 6.The LSTM-based temperature anomaly trend early warning method of claim 1, wherein, The step of inputting the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk includes: The multidimensional feature vector is quantized from floating-point format to fixed-point format to obtain the quantized feature vector; The quantized feature vector is input into the LSTM prediction model for fixed-point inference to obtain the quantized output result. The quantization output result is dequantized into a floating-point number format of the temperature anomaly risk probability value. 7.The LSTM-based temperature anomaly trend early warning method of claim 1, wherein, After the step of triggering an alarm using a preset alarm method if the probability value of the temperature anomaly risk is greater than a preset first risk threshold, the method further includes: Based on the alarm, obtain the corresponding alarm feedback information and collect it into the preset feedback database; Determine whether the alarm feedback information in the feedback database has reached a preset quantity; If the number of alarm feedback messages in the feedback database reaches a preset number, the parameters in the LSTM prediction model are adjusted based on the alarm feedback messages in the feedback database to obtain an optimized LSTM prediction model. 8.A temperature anomaly trend early warning device based on LSTM, characterized in that, The device includes: The data acquisition module is used to periodically acquire temperature data from a specified temperature control device to form a temperature dataset. The calculation module is used to calculate the dimension values ​​of multiple preset dimensions based on the temperature dataset; The arrangement module is used to arrange the dimensional values ​​of each preset dimension according to the preset dimensional order to obtain a multidimensional feature vector. The inference module is used to input the multidimensional feature vector into a preset LSTM prediction model for inference to obtain the probability value of temperature anomaly risk. The first judgment module is used to determine whether the probability value of the temperature anomaly risk is greater than a preset first risk threshold. The first alarm module is used to issue an alarm in a preset alarm mode if the probability value of the abnormal temperature risk is greater than a preset first risk threshold.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the LSTM-based temperature anomaly trend early warning method as described in any one of claims 1 to 7.

10. An electronic device, comprising: The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the LSTM-based temperature anomaly trend early warning method as described in any one of claims 1 to 7.

Images