A device temperature prediction and early warning method, device, apparatus and medium

Abstract

This application discloses a method, apparatus, device, and medium for equipment temperature prediction and early warning, relating to the field of artificial intelligence technology. The method includes: extracting multi-dimensional features from the collected time-series operating data of the target equipment to obtain a multi-dimensional feature tensor; inputting the multi-dimensional feature tensor into a trained cost-sensitive temporal convolutional network for temperature prediction to obtain a preliminary temperature prediction value and a target hidden feature vector; generating noise parameters for a Kalman filter adapted to the current operating condition of the target equipment based on the target hidden feature vector; fusing the preliminary temperature prediction value and the temperature state prediction value using the Kalman filter based on the noise parameters to obtain an optimal temperature state estimate and the corresponding estimation error variance; and calculating the comprehensive risk probability of the equipment temperature exceeding a preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance, thereby providing an early warning based on the comprehensive risk probability. This application achieves accurate prediction and risk warning of equipment temperature.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for predicting and warning equipment temperature. Background Technology

[0002] In industrial production, overheating of high-voltage equipment is one of the main causes of failures, making accurate prediction and early warning of equipment temperature crucial.

[0003] Existing equipment temperature prediction and early warning methods are difficult to meet the actual application needs of industrial sites. For example, traditional threshold methods require manual setting of fixed thresholds; while pure data-driven machine learning or deep learning prediction methods are insufficient in capturing high temperature extremes and rapid temperature rise inflection points; and traditional physical model filtering correction methods based on fixed parameters cannot dynamically adjust their noise parameters according to the real-time operating conditions of the equipment.

[0004] In summary, how to achieve accurate prediction of equipment temperature and risk warning is a problem that needs to be solved. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for predicting and warning of equipment temperature, capable of achieving accurate prediction and risk warning of equipment temperature. The specific solution is as follows: In a first aspect, this application discloses a method for predicting and warning equipment temperature, including: Multi-dimensional feature extraction is performed on the time-series operation data of the collected target device to obtain a multi-dimensional feature tensor; The multidimensional feature tensor is input into a trained cost-sensitive temporal convolutional network for temperature prediction, resulting in a preliminary temperature prediction value and a target hidden feature vector. Based on the target hidden feature vector, noise parameters of a Kalman filter adapted to the current operating condition of the target device are generated; The Kalman filter is used to fuse the preliminary temperature prediction value and the temperature state prediction value based on the noise parameters to obtain the optimal temperature state estimate value and the corresponding estimation error variance; the temperature state prediction value is obtained by a preset temperature state transition model based on the optimal temperature state estimate value at the previous moment and the equipment load input at the current moment. Based on the optimal temperature state estimate and the corresponding estimation error variance, the comprehensive risk probability of the equipment temperature exceeding the preset safety threshold is calculated, and an early warning is issued based on the comprehensive risk probability.

[0006] Optionally, the training process of the cost-sensitive temporal convolutional network includes: Multiple training samples are determined; each training sample includes historical time-series running data and a corresponding future real temperature sequence, wherein the historical time-series running data includes a historical temperature data sequence corresponding to the temperature dimension; A comprehensive cost-sensitive weight matrix is ​​constructed based on the historical temperature data sequence in each training sample; wherein, the comprehensive cost-sensitive weight matrix is ​​used to amplify the loss contribution of the preset high temperature range and preset heating stage during the training process; For each training sample, a multi-dimensional feature tensor extracted from the historical time-series running data of the training sample is input into the time-series convolutional network to be trained, so as to output the preliminary predicted temperature sequence corresponding to the training sample. The target loss value is obtained by weighting the error loss value between the preliminary predicted temperature sequence and the future actual temperature sequence using the comprehensive cost-sensitive weight matrix. The network parameters of the temporal convolutional network are updated based on the target loss value until the preset training termination condition is met, thus obtaining a trained cost-sensitive temporal convolutional network.

[0007] Optionally, constructing a comprehensive cost-sensitive weight matrix based on the historical temperature data sequence in each training sample includes: Obtain the historical temperature data sequence for each training sample; The absolute deviation between the historical temperature value at each moment in the historical temperature data sequence and the preset high temperature alarm threshold is converted into a basic weight value, and the basic weight value is trimmed to obtain the absolute temperature level cost weight. The first temperature trend sequence is obtained by calculating the first difference along the time dimension of the historical temperature data sequence, and the first temperature trend sequence is converted into a warming trend cost weight. The absolute temperature level cost weight and the warming trend cost weight are fused to obtain a comprehensive cost-sensitive weight matrix.

[0008] Optionally, the step of using the comprehensive cost-sensitive weight matrix to weight the error loss value between the preliminary predicted temperature sequence and the future true temperature sequence to obtain the target loss value includes: The first-order difference between the preliminary predicted temperature sequence and the future actual temperature sequence is calculated along the time dimension to obtain the corresponding second temperature trend sequence and third temperature trend sequence. A first mask is constructed based on the non-zero elements in the third temperature trend sequence; The second temperature trend sequence and the third temperature trend sequence are multiplied element by element, and a second mask is constructed based on the elements less than zero in the product result. The first mask and the second mask are multiplied element-wise to obtain the target mask. Calculate the absolute error loss value between the preliminary predicted temperature sequence and the future actual temperature sequence, and generate an enhanced error loss value based on the absolute error loss value and the target mask; The enhancement error is weighted using the comprehensive cost-sensitive weight matrix to obtain the target loss value.

[0009] Optionally, the noise parameters include process noise covariance parameters and observation noise covariance parameters; The generation of noise parameters for a Kalman filter adapted to the current operating condition of the target device based on the target hidden feature vector includes: The target hidden feature vector is reduced in dimensionality and aggregated in terms of time dimension to obtain aggregated features; The aggregated features are subjected to a nonlinear transformation to obtain the transformed features; The transformation features are respectively input to the first generation channel and the second generation channel of the trained parameter generation network, so as to output the original process noise covariance of the Kalman filter through the first generation channel and the original observation noise covariance of the Kalman filter through the second generation channel. Positive definite constraints are applied to the original process noise covariance and the original observation noise covariance, and the parameters after positive definite constraints are mapped to a numerical range that conforms to the thermal evolution law of the equipment, so as to obtain process noise covariance parameters and observation noise covariance parameters that are adapted to the current operating conditions.

[0010] Optionally, the step of fusing the preliminary temperature prediction value and the temperature state prediction value of the preset temperature state transition model based on the noise parameters using the Kalman filter to obtain the optimal temperature state estimate and the corresponding estimation error variance includes: The preliminary temperature prediction is used as the virtual observation value of the Kalman filter; based on the optimal temperature state estimate of the previous moment, the preset temperature state transition model and the equipment load input at the current moment, the prior temperature estimate and prior estimate error covariance at the current moment are calculated. The Kalman gain is calculated based on the noise parameters and the prior estimation error covariance. The Kalman gain is then used to fuse the virtual observations and the prior temperature estimates to obtain the optimal temperature state estimate and the corresponding estimation error variance.

[0011] Optionally, calculating the comprehensive risk probability of the equipment temperature exceeding a preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance includes: Determine the peak point among all optimal temperature state estimates within the prediction time window, and use the time corresponding to the peak point as the risk characteristic time. Calculate the standardized deviation of the optimal temperature state estimate corresponding to the risk characteristic moment from the preset safety threshold, and perform probability distribution mapping based on the standardized deviation and the estimation error variance corresponding to the risk characteristic moment to obtain the extreme risk probability of the actual temperature exceeding the preset safety threshold at the risk characteristic moment; The extreme value risk probability is calibrated using preset model performance parameters to obtain a comprehensive risk probability; wherein, the preset model performance parameters include the model baseline confidence level and the time-dependent confidence weight.

[0012] Optionally, the early warning based on the comprehensive risk probability includes: The target threshold range in which the comprehensive risk probability is located is determined, and the risk level corresponding to the target threshold range is taken as the current risk level; wherein, different threshold ranges correspond to different risk levels, and multiple threshold ranges are divided based on at least two preset risk level thresholds, and the risk level includes at least a safe state, a warning state, and an alarm state. The frequency of alarm events of different levels within a historical time window is statistically analyzed, and a corresponding decision action is determined based on the current risk level and the frequency of occurrence, so as to execute the corresponding early warning action based on the decision action.

[0013] Secondly, this application discloses a device for predicting and warning equipment temperature, comprising: The feature extraction module is used to extract multi-dimensional features from the collected time-series operation data of the target device to obtain a multi-dimensional feature tensor. The temperature prediction module is used to input the multidimensional feature tensor into a trained cost-sensitive temporal convolutional network to predict the temperature and obtain a preliminary temperature prediction value and a target hidden feature vector. The parameter generation module is used to generate noise parameters of a Kalman filter that are adapted to the current operating conditions of the target device based on the target hidden feature vector; The fusion module is used to fuse the preliminary temperature prediction value and the temperature state prediction value based on the noise parameters using the Kalman filter to obtain the optimal temperature state estimate value and the corresponding estimation error variance; the temperature state prediction value is obtained by a preset temperature state transition model based on the optimal temperature state estimate value at the previous moment and the equipment load input at the current moment. The early warning module is used to calculate the comprehensive risk probability of the equipment temperature exceeding the preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance, and to issue an early warning based on the comprehensive risk probability.

[0014] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the aforementioned disclosed device temperature prediction and early warning method.

[0015] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed device temperature prediction and early warning method.

[0016] As can be seen, this application, by extracting multi-dimensional features from time-series operational data and inputting them into a pre-trained cost-sensitive temporal convolutional network, outputs preliminary temperature predictions and target hidden feature vectors. This enables the accurate mining of temperature evolution-related features from multi-dimensional equipment operational data, improving the targeting and accuracy of preliminary temperature predictions. Furthermore, by generating Kalman filter noise parameters adapted to the current operating conditions of the equipment based on the target hidden feature vectors, the noise parameters can be dynamically adjusted according to the equipment's operating conditions, ensuring that the Kalman filter's operating parameters match the real-time operating status of the equipment. Moreover, by using a pre-defined temperature state transition model, a temperature state prediction is obtained based on the previous optimal temperature state estimate and the current equipment load input, ensuring that the temperature prediction aligns with physical laws such as the equipment's thermal inertia, avoiding the problem of unreasonable trends in purely data-driven predictions. Combining the noise parameters with the Kalman filter to fuse the data-driven preliminary temperature predictions and the temperature state predictions based on physical mechanisms, the advantages of data features and physical models are complemented, further improving the accuracy and reliability of the optimal temperature state estimate. Finally, based on the optimal temperature state estimate and the variance of the estimation error, the comprehensive risk probability of the equipment temperature exceeding the safety threshold is calculated, thereby achieving a quantitative assessment of temperature risk, making the early warning results more objective and forward-looking, and effectively reducing false alarms and missed alarms. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of a device temperature prediction and early warning method disclosed in this application; Figure 2 This is a diagram illustrating the training process architecture of a cost-sensitive temporal convolutional network disclosed in this application. Figure 3This is a flowchart of a specific equipment temperature prediction and early warning method disclosed in this application; Figure 4 This is a schematic diagram of the structure of a parameter generation network disclosed in this application; Figure 5 This is a schematic diagram of the structure of a device for predicting and warning equipment temperature, as disclosed in this application. Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0019] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Existing equipment temperature prediction and early warning methods still fall short of meeting the practical application needs of industrial sites: First, traditional threshold alarms or shallow statistical model methods can only set fixed thresholds based on historical experience for binary judgment, which cannot adapt to dynamic operating conditions such as sudden load changes and environmental interference. Their response to temperature anomalies is severely delayed, and they lack forward-looking prediction of temperature evolution trends, resulting in insufficient timeliness and specificity of early warnings. Second, while pure data-driven machine learning or deep learning prediction methods can improve overall prediction accuracy, they are prone to problems such as peak compression and trend smoothing. They lack the ability to capture high-temperature extremes and rapid temperature rise inflection points, and the predicted trajectory may contradict the equipment's specifications. Inherent physical laws such as thermal inertia lead to a lack of physical rationality in the prediction results. Furthermore, these methods only output a single predicted value, failing to quantify the uncertainty of the prediction results and introducing blindness into early warning decisions. Thirdly, traditional physical model filtering correction methods based on fixed parameters have preset fixed noise parameters that cannot be dynamically adjusted according to the real-time operating conditions of the equipment. During periods of stable operating conditions, excessive filtering can easily cause unnecessary prediction fluctuations, while during periods of sudden changes in operating conditions, overly strong filtering can lead to response lags. It is difficult to balance the smoothness of prediction with the agility of dynamic tracking, and it fails to achieve efficient integration of data-driven prediction and physical model deduction, limiting the accuracy and reliability of the corrected prediction results. Therefore, this application discloses a method, device, equipment, and medium for equipment temperature prediction and early warning, capable of achieving accurate prediction and risk warning of equipment temperature.

[0021] See Figure 1 As shown in the figure, this application discloses a method for predicting and warning equipment temperature, the method comprising: Step S11: Perform multi-dimensional feature extraction on the collected time-series operation data of the target device to obtain a multi-dimensional feature tensor.

[0022] In this embodiment, the target equipment specifically refers to high-voltage power equipment such as high-voltage frequency converters, transformers, and high-voltage motors. At the industrial IoT sensing layer, multiple types of sensors deployed on the target equipment can continuously collect data on key measurement points during equipment operation, forming multivariate time-series data. The collected time-series operational data includes, but is not limited to, multi-dimensional monitoring data such as workshop ambient temperature, instantaneous current and voltage, equipment load rate, operating frequency, cooling fan operating power, unit temperature, and transformer temperature, thereby comprehensively depicting the equipment's operating status and its thermal evolution process.

[0023] In a specific implementation, the step of extracting multi-dimensional features from the collected time-series operating data of the target device to obtain a multi-dimensional feature tensor includes: preprocessing the collected time-series operating data of the target device to obtain processed data; extracting preset statistical derived features from the processed data on the target dimension; wherein the target dimension includes a temperature dimension and a load dimension, and the statistical derived features include at least one of first-order difference, second-order difference, moving average, standard deviation, and transformation slope; decomposing the data corresponding to the temperature dimension to be predicted using a time-series decomposition method to obtain a decomposition result; wherein the decomposition result includes at least one of trend term, periodic term, and seasonal term; extracting spectral features and thermodynamic prior features from the processed data; the thermodynamic prior features include at least one of heat generation to heat dissipation per unit time, transient thermal resistance, and steady-state thermal resistance; and obtaining a multi-dimensional feature tensor based on the statistical derived features, the decomposition result, the spectral features, and the thermodynamic prior features.

[0024] Specifically, assuming time-series data is synchronously collected from M sensors (temperature, current, voltage, etc.), after preprocessing such as timestamp alignment, missing value completion, noise reduction, and combining detrending analysis and smoothing filtering to reduce high-frequency interference, the processed data is obtained. The processed data is specifically the original sequence matrix with a time step of T. : ; in, Let T represent the real number space, T be the time step, and M be the number of dimensions of the device operation data.

[0025] Furthermore, this application requires exploratory data analysis (EDA) to extract pre-defined statistical derived features from the processed data in the target dimensions at the time domain level. The target dimensions include temperature and load dimensions, and the statistical derived features include at least one of first-order difference, second-order difference, moving average, standard deviation, and slope of change. These statistical derived features can reflect the dynamic changing trends of the equipment's operating status; for example, first-order difference represents the rate of change, second-order difference represents acceleration, moving average reflects local trends, standard deviation measures the degree of fluctuation, and slope of change describes the direction and intensity of change. At the time-series structure level, time-series decomposition methods are needed to decompose the data corresponding to the temperature dimension to be predicted, obtaining decomposition results including trend terms, periodic terms, and seasonal terms. Specifically, time-series decomposition methods such as SFT (Sparse Fourier Transform) can be used to explicitly decompose the temperature sequence into long-term evolution trends, periodic fluctuations, and random seasonal components, providing structured time-series prior knowledge for the model. At the frequency domain level, Fast Fourier Transform (FFT) can be used to extract spectral features from the processed data, which can be used to capture periodic patterns and potential anomalous frequency components in time-series data. Furthermore, thermodynamic prior features can be introduced to model the thermal inertia of the equipment. These thermodynamic prior features include the heat generation to heat dissipation ratio per unit time, transient thermal resistance, and steady-state thermal resistance, etc., which can enhance the model's ability to express temperature rise trends from a physical perspective. Finally, by normalizing the scales of these different features—statistically derived features, decomposition results, spectral features, and thermodynamic prior features—a multidimensional feature tensor is formed. : ; Where B represents the sample training batch size (a batch of independent, equally long segments of device operating status); L represents the input sequence length (the time step of the input sequence window); and C represents the number of feature channels, which is the integration of multi-source heterogeneous features in the channel dimension. In particular, the feature channel is the mathematical representation of the explicit fusion cost prior, which directly encodes the physical constraints and knowledge of the device, such as thermal inertia and thermal resistance, into numerical features that the model can read, injecting key domain causal logic into the prediction model.

[0026] Step S12: Input the multidimensional feature tensor into the trained cost-sensitive temporal convolutional network to perform temperature prediction, and obtain the preliminary temperature prediction value and the target hidden feature vector.

[0027] In this embodiment, the extracted multi-dimensional features are extracted and then input into a pre-trained cost-sensitive temporal convolutional network to output preliminary temperature prediction values ​​and target hidden feature vectors. This enables the accurate mining of temperature evolution-related features from multi-dimensional equipment operation data, providing a feature basis that fits the actual operating state of the equipment for subsequent temperature prediction, and improving the pertinence and accuracy of preliminary temperature prediction.

[0028] Understandably, the input to a cost-sensitive temporal convolutional neural network is a multidimensional feature tensor. The output consists of two parts: preliminary temperature prediction values. The target hidden feature vector, specifically referring to the deep hidden features. Where T is the length of the output sequence (i.e. the prediction window length) and D is the number of deep feature channels.

[0029] It should be noted that the Risk-Aware Temporal Convolutional Network (RA-TCN) is composed of multiple stacked dilated causal convolutional residual blocks. Each residual block contains dilated causal convolutional layers and residual connections, thereby ensuring: (1) strict temporal causality, that is, the output at the current moment depends only on historical and current information; (2) an exponentially growing receptive field, enabling the network to capture multi-scale dependencies from short-term fluctuations to long-term trends.

[0030] The core feature extraction mechanism of cost-sensitive temporal convolutional neural networks is dilated causal convolution, which is used for a given input sequence. and a one-dimensional convolution kernel The dilated causal convolution operation at time t and output channel c can be precisely defined as: ; Where k is the kernel size; d is the dilation factor, which defines the field of view of the kernel when processing input data. As the network depth n increases, d typically increases by... It grows exponentially, causing the effective receptive field R to expand rapidly: This means that only a small number of layers are needed to capture extremely long historical dependencies. (Subscript) This ensures the causality of the computation, meaning that the output at time t depends only on the historical inputs up to and before time t. It strictly meets the constraints of autoregressive prediction.

[0031] In addition, to ensure consistent input and output sequence lengths, zero-padding is performed on the left side of the input sequence before each convolutional layer, with a padding amount of 1 / 2. This left padding is key to achieving causality. The network employs a dual-output structure at the end, which is fundamental for enabling subsequent adaptive correction and collaborative optimization of the physical model to adapt to temperature change trends. The network output can be divided into two paths: Predicted output path: This path represents the final deep hidden features. Through a linear projection layer, it is mapped to a preliminary prediction sequence for a single channel. : ; in, Let be the projection weight matrix. For projection bias term; Deep hidden feature output path: Before the linear projection layer, the same hidden feature is output... Directly extracted as a deep hidden feature : ; Deep hidden features The abstract patterns extracted from the raw data and closely related to the dynamics of temperature evolution will be provided to the subsequent Kalman filter parameter learning network for dynamically adjusting the state estimation parameters.

[0032] The training process of the cost-sensitive temporal convolutional network includes: determining multiple training samples; each training sample includes historical temporal data and a corresponding future real temperature sequence, wherein the historical temporal data includes a historical temperature data sequence corresponding to the temperature dimension; constructing a comprehensive cost-sensitive weight matrix based on the historical temperature data sequence in each training sample; wherein the comprehensive cost-sensitive weight matrix is ​​used to amplify the loss contribution of a preset high-temperature range and a preset heating stage during the training process; for each training sample, inputting a multi-dimensional feature tensor extracted from the historical temporal data in the training sample into the temporal convolutional network to be trained to output a preliminary predicted temperature sequence corresponding to the training sample; using the comprehensive cost-sensitive weight matrix to weight the error loss value between the preliminary predicted temperature sequence and the future real temperature sequence to obtain a target loss value; updating the network parameters of the temporal convolutional network based on the target loss value until a preset training termination condition is met, thereby obtaining a trained cost-sensitive temporal convolutional network.

[0033] That is, such as Figure 2As shown, this application first determines multiple training samples, each of which includes historical time-series data and a corresponding future true temperature sequence. The historical time-series data includes historical temperature data sequences corresponding to the temperature dimension. That is, each training sample contains historical multi-source time-series data of input sequence length L and a future true temperature sequence of output sequence length T. Furthermore, to address the learning deficiencies and warning lags caused by uneven sample distribution in traditional prediction models during high-temperature critical ranges and rapid temperature rise stages, this embodiment innovatively introduces a multi-dimensional cost-sensitive learning mechanism during the training sample construction stage. This mechanism constructs a dynamic, comprehensive cost-sensitive weight matrix based on the historical temperature data sequence in each training sample. This matrix amplifies the loss contribution of preset high-temperature ranges and preset temperature rise stages during model training, explicitly strengthening the contribution of high-risk samples and forcing the model to focus on decision boundaries crucial for safety warnings. Further, for each training sample, a multi-dimensional feature tensor extracted from the historical time-series data in the training sample is input into the temporal convolutional network to be trained, outputting a corresponding preliminary predicted temperature sequence. Then, the aforementioned comprehensive cost-sensitive weight matrix is ​​used to weight the error loss value between the preliminary predicted temperature sequence and the future actual temperature sequence to obtain the target loss value. That is, during the model training phase, this application applies the aforementioned comprehensive cost-sensitive weight matrix to the construction of the loss function (such as mean squared error, MSE, etc.), and dynamically adjusts the learning gradient of the model parameters through backpropagation of weighted loss, so that the prediction accuracy and sensitivity of the model in high-temperature and high-risk areas are systematically improved, providing a key data and learning paradigm foundation for subsequent high-temperature trend early warning.

[0034] Specifically, constructing a comprehensive cost-sensitive weight matrix based on historical temperature data sequences in each training sample includes: acquiring historical temperature data sequences in each training sample; converting the absolute deviation between the historical temperature value at each moment in the historical temperature data sequence and a preset high-temperature alarm threshold into a basic weight value, and trimming the basic weight value to obtain an absolute temperature level cost weight; calculating the first-order difference of the historical temperature data sequence along the time dimension to obtain a first temperature trend sequence, and converting the first temperature trend sequence into a rising temperature trend cost weight; and fusing the absolute temperature level cost weight and the rising temperature trend cost weight to obtain a comprehensive cost-sensitive weight matrix. It can be understood that the construction of the comprehensive cost-sensitive weight matrix is ​​based on two core dimensions: one is the absolute temperature level dimension, used to measure the degree of deviation of the sample temperature value from the safety threshold, with the corresponding weight being the absolute temperature level cost weight; the other is the rising temperature trend cost dimension, used to measure the rate and magnitude of temperature increase of the sample per unit time, with the corresponding weight being the rising temperature trend cost weight.

[0035] In terms of absolute temperature level, for a given batch of historical temperature data sequences It is necessary to calculate the absolute deviation between the historical temperature value at each moment in the sequence and the preset high-temperature alarm threshold. : ; in, It is a very small positive constant used to ensure gradient fluidity for samples close to the threshold; A preset high-temperature alarm threshold is set. Furthermore, this is achieved through a configurable first preset mapping function. Convert the deviation into a base weight value. : ; Wherein, mapping function It can be adapted to linear or nonlinear (such as exponential) amplification strategies, where α is the amplification factor: ,or ; Finally, the base weights are pruned to ensure training stability, resulting in the absolute temperature level cost weights. : ; in These are the preset minimum and maximum weight boundaries, respectively.

[0036] In terms of the cost of warming trends, to more precisely capture and enhance the high-risk transient process of rapid warming, this embodiment directly analyzes historical temperature data sequences. The discrete gradient (i.e., the first difference) is calculated along the time dimension to obtain the first temperature trend sequence G, which serves as a measure of the warming trend: ; Here, `diff` represents the backward difference operation along the time dimension; furthermore, to obtain a smoother weight distribution and introduce hyperparameter controllability, the first temperature trend sequence is converted into a warming trend cost weight using a second preset mapping function. Specifically, a calculation method based on the Sigmoid function can be used: ; in, To pad the difference sequence G with a zero at the beginning, so that the length is restored to T; This is the trend reinforcement coefficient, which controls the amplification strength of the trend weight; This is the threshold for the significance of temperature increase; the weights increase significantly only when the gradient exceeds this threshold. The Sigmoid function maps the gradient to smooth weight increments. This is a sensitivity adjustment parameter used to control the steepness of the change in the weight of the heating trend.

[0037] Finally, this embodiment integrates the absolute temperature level cost weights and the warming trend cost weights, which are measured from the perspectives of state and trend respectively, to form the final comprehensive cost-sensitive weight matrix. : ; in This indicates element-wise multiplication. This fusion strategy can simultaneously highlight both sustained high-temperature states and dangerous warming processes, achieving a panoramic and forward-looking sample enhancement for high-temperature risks.

[0038] Furthermore, the step of weighting the error loss value between the preliminary predicted temperature sequence and the future true temperature sequence using the comprehensive cost-sensitive weight matrix to obtain the target loss value includes: calculating the first-order difference of the preliminary predicted temperature sequence and the future true temperature sequence along the time dimension to obtain the corresponding second temperature trend sequence and third temperature trend sequence; constructing a first mask based on the non-zero elements in the third temperature trend sequence; multiplying the second temperature trend sequence and the third temperature trend sequence element by element, and constructing a second mask based on the elements less than zero in the product result; multiplying the first mask and the second mask element by element to obtain the target mask; calculating the absolute error loss value between the preliminary predicted temperature sequence and the future true temperature sequence, and generating an enhanced error loss value based on the absolute error loss value and the target mask; and weighting the enhanced error using the comprehensive cost-sensitive weight matrix to obtain the target loss value.

[0039] Understandably, to address the issues of "peak compression" in high-temperature range predictions and "trend misjudgment" during rapid temperature rise, this embodiment discloses a method for constructing a target loss function. This target loss function is essentially a Trend-Misjudgment Penalty Loss (TMP Loss). This loss function not only penalizes the absolute error between the predicted and actual values ​​but also explicitly and forcefully penalizes misjudgments where the predicted trend is opposite to the actual trend. Specifically, given a model that initially predicts a temperature sequence... With future real temperature sequence And the comprehensive cost-sensitive weight matrix generated by constructing high-temperature cost samples. The objective loss function is calculated as follows: First, the preliminary temperature sequence will be predicted. and future true temperature sequence The first-order difference is calculated along the time dimension to obtain the corresponding second temperature trend sequence. and the third temperature trend sequence : ; ; In addition, the difference sequence needs to be zero-padded at the start time step to maintain the same shape B×T×1 as the original sequence.

[0040] Secondly, a first mask is constructed based on the non-zero elements in the third temperature trend sequence. This is used to identify the moments when the actual temperature changes. The second and third temperature trend sequences are then multiplied element-wise, and a second mask is constructed based on the elements less than zero in the product. This is used to identify the moment when the predicted direction of change is opposite to the actual direction of change. Then, the first mask and the second mask are multiplied element-wise to obtain the target mask. This mask identifies the locations of all sample points where the actual temperature changes and the predicted change direction is opposite to the actual change direction.

[0041] Furthermore, the absolute error loss between the preliminary predicted temperature series and the future actual temperature series is calculated: ; in, For smoothing coefficients; Then, an enhanced error loss value is generated based on the absolute error loss value and the target mask; ; in, This is the trend misjudgment penalty coefficient. This formula significantly amplifies the calculation of the error loss value for trend misjudgment points (such as predicting a decline when the trend is actually rising).

[0042] Finally, the enhancement error is weighted using the comprehensive cost-sensitive weight matrix to obtain the target loss value of the trend misjudgment penalty loss function: .

[0043] Therefore, this application constructs a joint optimization system using a cost-sensitive temporal convolutional network and a trend misjudgment penalty loss function: Forward propagation: Network-generated prediction and deep hidden features ; Loss Calculation: Compared with the true value Substitute the trend misjudgment penalty loss function At the same time, it incorporates a comprehensive cost-sensitive weight matrix. ; Backpropagation: Loss Function The generated gradients include: the absolute error gradient, which drives the model's output predictions to continuously approach the true values; the trend misjudgment penalty gradient, which strongly corrects directional errors and is the key driver for solving "trend misjudgment"; and the cost weighted gradient, which amplifies the gradient signal in high-temperature and high-temperature-trend regions, forcing the model to invest more modeling capacity in such high-risk regions.

[0044] Through this joint design, the model is taught and learns two things during training: first, to accurately predict numerical values; and second, especially for high-temperature and high-temperature-rise regions, to ensure the correctness of the direction of change. This directly counteracts the model's global conservative tendency (peak compression) at the optimization objective level, enabling it to make more aggressive and timely predictions of high-temperature inflection points.

[0045] Step S12: Generate noise parameters of a Kalman filter adapted to the current operating conditions of the target device based on the target hidden feature vector.

[0046] In this embodiment, by inputting the target hidden feature vector into a pre-trained parameter generation network, Kalman filter noise parameters adapted to the current operating conditions of the equipment are generated. This enables the noise parameters to be dynamically adjusted according to the operating conditions of the equipment, allowing the operating parameters of the Kalman filter to match the real-time operating status of the equipment. This overcomes the limitation of poor adaptability of fixed noise parameters and improves the operating condition adaptability of filter fusion.

[0047] It should be noted that the parameter generation network includes a first generation channel and a second generation channel. The noise parameters include the process noise covariance parameter of the Kalman filter output by the first generation channel and the observation noise covariance parameter of the Kalman filter output by the second generation channel. That is, the parameter generation network of this application is used to map the target hidden feature vector to the noise parameters of the Kalman filter. The parameter generation network includes two independent parameter generation channels, which is actually a Bi-Channel Independent Parameter Generation Network (BCI-PGN). The first generation channel is used to output the process noise covariance parameter of the Kalman filter, characterizing the uncertainty of the preset temperature state transition model itself; the second generation channel is used to output the observation noise covariance parameter of the Kalman filter, characterizing the uncertainty of the initial temperature prediction value as an observation value.

[0048] In a specific implementation, the training process of the parameter generation network includes: after the cost-sensitive temporal convolutional network is trained, freezing the parameters of the cost-sensitive temporal convolutional network, and inputting historical time-series running data into the cost-sensitive temporal convolutional network after freezing the parameters, outputting historical preliminary temperature prediction values ​​and historical target hidden feature vectors; inputting the historical target hidden feature vectors into the parameter generation network to be trained to generate historical noise parameters of the Kalman filter; using the Kalman filter to fuse the historical preliminary temperature prediction values ​​and the historical temperature state prediction values ​​of the preset temperature state transition model based on the historical noise parameters to obtain the historical optimal temperature state estimate, and constructing a joint loss function based on the error between the historical optimal temperature state estimate and the actual temperature value; performing backpropagation based on the joint loss function to update the network parameters of the parameter generation network until a preset training termination condition is met, thus obtaining the trained parameter generation network.

[0049] In other words, after the cost-sensitive temporal convolutional network is trained, its parameters are frozen, and the parameter generation network and Kalman filter are jointly trained end-to-end. First, historical time-series running data is input into the cost-sensitive temporal convolutional network with frozen parameters, outputting historical preliminary temperature predictions and historical target hidden feature vectors. Then, the historical target hidden feature vectors are input into the parameter generation network to be trained, generating historical noise parameters for the Kalman filter (including historical process noise covariance and historical observation noise covariance). Next, the Kalman filter is used to fuse the historical preliminary temperature predictions and the historical temperature state predictions from the preset temperature state transition model based on the historical noise parameters, obtaining the historical optimal temperature state estimate. The preset temperature state transition model is a constant-rate state transition model oriented towards the thermal characteristics of the target equipment, using the equipment load as an external control input, and its state vector includes temperature and its rate of change. Finally, a joint loss function is constructed based on the error between the historical optimal temperature state estimate and the actual temperature value. In this embodiment, the joint loss function includes a prediction error loss term and a regularization constraint term. The prediction error loss term uses the mean square error between the historical best temperature state estimate and the actual temperature value. The regularization constraint term is used to impose norm constraints on the historical noise parameters output by the parameter generation network to prevent overfitting. The expression of the joint loss function is as follows: ; in, For the prediction error loss term, λ is the regularization constraint term, and λ is the regularization coefficient.

[0050] Finally, backpropagation is performed based on the joint loss function value to update the network parameters of the parameter generation network until the preset training termination condition is met, resulting in a trained parameter generation network. In other words, the gradient is backpropagated to the parameter generation network through the Kalman filter update equation, enabling it to learn to predict the optimal noise parameters that minimize the final prediction error of the entire system based on the input features, thus achieving adaptation.

[0051] Step S14: The Kalman filter is used to fuse the preliminary temperature prediction value and the temperature state prediction value based on the noise parameters to obtain the optimal temperature state estimate value and the corresponding estimation error variance; the temperature state prediction value is obtained by the preset temperature state transition model based on the optimal temperature state estimate value of the previous moment and the equipment load input at the current moment.

[0052] This embodiment discloses a preset temperature state transition model. This model can obtain a predicted temperature state value based on the optimal temperature state estimate from the previous moment and the predicted equipment load input at the current moment. Then, by using a Kalman filter based on dynamic noise parameters, the preliminary temperature prediction value and the predicted temperature state value are fused to obtain the optimal temperature state estimate and the corresponding estimation error variance. This model can achieve an organic fusion of data-driven prediction results and physical mechanism inference results. It retains the sensitivity of data-driven prediction to changes in equipment operation, while ensuring that the temperature prediction trend conforms to the inherent characteristics of the equipment through physical model constraints. At the same time, it quantifies the uncertainty of the prediction results, making the temperature prediction results more reliable and more referential.

[0053] Step S15: Based on the estimated optimal temperature state value and the corresponding estimation error variance, calculate the comprehensive risk probability that the equipment temperature exceeds the preset safety threshold, and issue an early warning based on the comprehensive risk probability.

[0054] In this embodiment, the comprehensive risk probability of the equipment temperature exceeding the safety threshold is calculated based on the optimal temperature state estimate and the estimation error variance. Then, an early warning is issued based on the comprehensive risk probability, which can realize a quantitative assessment of the risk of the equipment temperature exceeding the limit, making the early warning results more objective and forward-looking, and effectively reducing false alarms and missed alarms.

[0055] As can be seen, this application, by extracting multi-dimensional features from time-series operational data and inputting them into a pre-trained cost-sensitive temporal convolutional network, outputs preliminary temperature predictions and target hidden feature vectors. This enables the accurate mining of temperature evolution-related features from multi-dimensional equipment operational data, improving the targeting and accuracy of preliminary temperature predictions. Furthermore, by generating Kalman filter noise parameters adapted to the current operating conditions of the equipment based on the target hidden feature vectors, the noise parameters can be dynamically adjusted according to the equipment's operating conditions, ensuring that the Kalman filter's operating parameters match the real-time operating status of the equipment. Moreover, by using a pre-defined temperature state transition model, a temperature state prediction is obtained based on the previous optimal temperature state estimate and the current equipment load input, ensuring that the temperature prediction aligns with physical laws such as the equipment's thermal inertia, avoiding the problem of unreasonable trends in purely data-driven predictions. Combining the noise parameters with the Kalman filter to fuse the data-driven preliminary temperature predictions and the temperature state predictions based on physical mechanisms, the advantages of data features and physical models are complemented, further improving the accuracy and reliability of the optimal temperature state estimate. Finally, based on the optimal temperature state estimate and the variance of the estimation error, the comprehensive risk probability of the equipment temperature exceeding the safety threshold is calculated, thereby achieving a quantitative assessment of temperature risk, making the early warning results more objective and forward-looking, and effectively reducing false alarms and missed alarms.

[0056] See Figure 3 As shown, this application discloses a specific method for predicting and warning equipment temperature. Compared to the previous embodiment, this embodiment further explains and optimizes the technical solution. Specifically, it includes: Step S21: Extract multi-dimensional features from the collected time-series operation data of the target device to obtain a multi-dimensional feature tensor.

[0057] Step S22: Input the multidimensional feature tensor into the trained cost-sensitive temporal convolutional network for temperature prediction to obtain a preliminary temperature prediction value and a target hidden feature vector; wherein, the cost-sensitive temporal convolutional network is constructed based on several dilated causal convolutional residual blocks, and each dilated causal convolutional residual block includes a dilated causal convolutional layer and a residual connection.

[0058] Step S23: Perform dimensionality reduction and time dimension aggregation on the target hidden feature vector to obtain aggregated features.

[0059] In this embodiment, as Figure 4 As shown, the parameter generation network's input is directly taken from a pre-trained cost-sensitive temporal convolutional network model, and the extracted target hidden feature vector... This feature has already been processed by a deep network, containing multi-scale time-series patterns ranging from short-term fluctuations to long-term trends, and it inherently contains high-order statistical properties (such as volatility implicitly captured through convolution operations). Therefore, explicit sliding window statistical pooling is unnecessary.

[0060] First, this embodiment uses a lightweight feature projection layer to reduce the dimensionality of the target hidden feature vector, thus reducing the high-dimensional features... Mapped to a feature space with appropriate dimensionality that is more suitable for parameter generation: ; in, For projection weights, For bias terms, The projected feature dimension is used to preserve key information while reducing the complexity of subsequent networks and the risk of overfitting.

[0061] Furthermore, considering that the Kalman filter needs to utilize the noise parameters corresponding to each time step of the entire prediction sequence at the output, the dimensionality-reduced features are averaged and aggregated along the time dimension to obtain aggregated features: ; Where L is the length of the input sequence. For indexed slices, it means taking the first slice. l All samples and all channels at each time step.

[0062] Step S24: Perform a nonlinear transformation on the aggregated features to obtain transformed features, and input the transformed features into the first generation channel and the second generation channel of the trained parameter generation network, respectively, so as to output the original process noise covariance of the Kalman filter through the first generation channel and the original observation noise covariance of the Kalman filter through the second generation channel.

[0063] In this embodiment, the aggregation features are further subjected to a nonlinear transformation to obtain the transformed features: ; in, This refers to transformed features (or shared features). This is the weight matrix. For bias terms, As an aggregation feature, Using this as the activation function avoids repeatedly extracting similar features for noisy parameters, thus improving the model's parameter efficiency.

[0064] Furthermore, such as Figure 4 As shown, the transformation features will be The parameters are input to the first and second generation channels of the trained parameter generation network, respectively, to output the original process noise covariance of the Kalman filter through the first generation channel and the original observation noise covariance of the Kalman filter through the second generation channel.

[0065] Specifically, the first generation channel corresponding to the process noise covariance focuses on estimating the uncertainty of the state prediction model itself. For the constant temperature state transition model, its state vector is... Process noise covariance It is usually modeled as a diagonal matrix. ;in, Indicates temperature. This represents the rate of temperature change, with the upper right corner T indicating transpose. This represents the variance of noise during the temperature process. The variance of the noise in the rate of change process is represented by the superscript TT, which indicates the covariance between temperature and temperature. This represents the covariance of the rate of temperature change; therefore, it generates the original process noise covariance. The structure of this channel is as follows: ; ; ; in, To hide process noise features, This is the weight matrix of the first fully connected layer. For the bias term of the first-layer fully connected network, This is the weight matrix of the second fully connected layer. For the bias term of the second-layer fully connected network, This is the output layer weight matrix. H is the activation function, and H is the prediction step size. This embodiment uses the ELU activation function to alleviate the gradient vanishing problem and allow negative activation values, providing a wider input range for subsequent transformations.

[0066] The second generation channel corresponding to the observation noise covariance focuses on estimating the uncertainty of data-driven predictions (as virtual observations), thereby generating the original observation noise covariance. The structure of the channel is as follows: ; ; ; in, To observe the hidden features of noise, The weight matrix of the first fully connected layer network. For the bias term of the first-layer fully connected network, This is the weight matrix of the second fully connected layer. For the bias term of the second-layer fully connected network, This is the output layer weight matrix. H is the activation function, and H is the prediction step size.

[0067] Step S25: Apply positive definite constraints to the original process noise covariance and the original observation noise covariance, and map the parameters after positive definite constraints to a numerical range that conforms to the thermal evolution law of the equipment, so as to obtain process noise covariance parameters and observation noise covariance parameters that are adapted to the current operating conditions.

[0068] In this embodiment, to ensure that the generated parameters meet the mathematical requirements of the Kalman filter (such as positive definiteness), a reasonable physical prior needs to be applied to the parameter range. Specifically, the Softplus activation function is first used to ensure that the output is strictly positive: ; ; in, It should be a very small positive number to prevent the output from being zero.

[0069] Then, learnable preset mapping parameters are introduced, specifically including scaling factor (s) and bias term (b), to map the parameters after positive definiteness constraint to a numerical range that conforms to physical meaning, in accordance with the thermal evolution law of the equipment, so as to obtain process noise covariance parameters and observation noise covariance parameters adapted to the current operating conditions: ; ; ; in, , , All are scaling factors, and all are >0; , , All are bias factors, and all are ≥0; k Representing the k Each prediction time step. The initial values ​​of these parameters can be set based on historical data or physical knowledge, and can be further optimized through learning during training.

[0070] As can be seen, this application addresses the problem of excessive fluctuations introduced by fixed noise covariance parameters (Q, R) in traditional Kalman filtering during stable periods and lags in response during abrupt changes by designing an independent dual-channel parameter generation network parallel to the main prediction network. The core innovation of this network lies in combining data-driven feature learning capabilities with the constraints of physical mechanisms to achieve adaptive mapping from deep temporal features to physical model parameters. Through this learnable, condition-adaptive parameter generation network, the internal parameters of the physical state estimator (Kalman filter) can be dynamically adjusted, enabling refined and interpretable corrections to the initial data-driven predictions. This maintains the prediction model's sensitivity to abrupt changes while introducing physical consistency constraints to smooth out unreasonable fluctuations.

[0071] Step S26: Use the preliminary temperature prediction value as the virtual observation value of the Kalman filter; calculate the prior temperature estimate and prior estimation error covariance at the current moment based on the optimal temperature state estimate value of the previous moment, the preset temperature state transition model, and the equipment load input at the current moment; calculate the Kalman gain based on the noise parameter and the prior estimation error covariance, and use the Kalman gain to fuse the virtual observation value and the prior temperature estimate value to obtain the optimal temperature state estimate value and the corresponding estimation error variance.

[0072] In this embodiment, based on the noise parameters dynamically generated by the parameter generation network, a Kalman filter specifically designed for the thermal evolution process of high-voltage equipment is used to refine the preliminary predictions of the data-driven network through physical correction. This system innovatively integrates data-driven preliminary predictions, external equipment load input estimation based on operating conditions, and a targeted temperature state transition model for high-voltage equipment, achieving multi-level modeling of the equipment's temperature rise pattern.

[0073] To accurately characterize the temperature dynamics of high-voltage equipment under short-term stable operating conditions, this application constructs a constant-rate temperature state transition model specifically for describing the thermal process of high-voltage equipment, namely the aforementioned preset temperature state transition model, and uses it as the core physical mechanism of the Kalman filter. This model is based on the physical characteristic of high-voltage equipment having significant thermal inertia due to its large thermal mass. Its core assumption is that within a short time window when external conditions such as load remain relatively stable, the rate of change of the core temperature inside the equipment is approximately constant. Based on this assumption, this embodiment extends the standard constant-rate model by innovatively using load fluctuations as an external control input. A state-space equation is introduced to better reflect the dynamic process of "electric-thermal" coupling in the actual operation of high-voltage equipment. The corresponding state-space model is expressed as follows: ; ; Wherein, the state vector This includes temperature and its rate of change; This is the state transition matrix, which reflects physical continuity; Given the device load input value at time k, and considering that direct and accurate measurement or high-precision prediction of the load is difficult to obtain in practice, this application designs a short-term load estimation algorithm based on moving average, which uses recent historical load data to make a stable estimate of the future short-term load, and uses this as an observable external excitation input. B The input control matrix quantifies the instantaneous impact of a unit change in load on the rate of temperature change. It is one of the key physical parameters that the model learns from the data and is used to establish the dynamic relationship between "electrical load" and "thermal change". This is the observation matrix, indicating that only the temperature state can be directly observed. Process noise covariance and The observation noise covariance is dynamically generated by the parameter generation network based on real-time operating conditions. For process noise, To observe noise.

[0074] The physical state correction process along the prediction time dimension k=1,2,...,H Iterative process: State prediction steps: ; ; in, The temperature is the prior estimate at the current moment. Let F be the optimal temperature state estimate from the previous moment, and let F be the state transition matrix. B is the current device load input, and B is the input control matrix. Let the prior estimate error covariance be the value at the current moment. The error covariance of the previous time step. Let be the process noise covariance.

[0075] This step is based on a constant temperature model for high-voltage equipment, combined with the previous time-lapse estimate of the optimal temperature state. and the current device load input The prior prediction of the current temperature and rate of change is performed to obtain the prior estimate of the current temperature and the covariance of the prior estimate error. (Load term) The introduction of this feature enables the model to respond to changes in operating conditions.

[0076] Observation update steps: ; ; ; in, Preliminary temperature forecast , as virtual observations injected into the filtering process; H is the observation matrix; Kalman gain This acts as a dynamic weight, whose values ​​are generated by the parameter generation network. and This joint decision achieves optimal Bayesian fusion between prior predictions from the physical model and data-driven observations. This is an estimate of the optimal temperature state within the prediction time window; This is the error covariance at the current moment, which includes the estimation error variance corresponding to the optimal temperature state estimate; It is an identity matrix.

[0077] In other words, the above constitutes a complete "sensing-generation-fusion" intelligent closed loop for high-voltage equipment temperature prediction. Specifically, the sensing step involves a pre-trained cost-sensitive temporal convolutional network extracting deep features from the raw data and generating preliminary predictions. Generation: The parameter generation network generates deep hidden features from a cost-sensitive temporal convolutional network. Dynamically generate configurations adapted to the current operating conditions. and Fusion: The Kalman filter utilizes the parameters generated by the parameter generation network to integrate the preliminary predictions of the cost-sensitive temporal convolutional network. The predictions are optimally fused with those of a constant-rate-temperature state transition model for high-voltage equipment, while also considering external load input. The influence of physical constraints is ultimately reflected in the output, resulting in a more accurate and reliable state estimate. .

[0078] As can be seen, this application defines a more practical, temperature-controlled constant-rate state transition model for high-voltage equipment by combining a general constant-rate model with the thermal inertia physical characteristics of high-voltage equipment and the estimated load available in practice. This is achieved by introducing load input. and the control matrix driven by it B The model can explicitly characterize the excitation effect of electrical load disturbance on temperature dynamics. At the same time, it uses a dual-channel independent parameter generation network to achieve adaptive adjustment of noise parameters, thus achieving an excellent balance between the flexibility of data-driven approach and the rigor of physical mechanisms. This significantly improves the prediction accuracy and robustness of equipment temperature rise trend under complex and time-varying operating conditions.

[0079] Step S27: Based on the optimal temperature state estimate and the corresponding estimation error variance, calculate the comprehensive risk probability that the equipment temperature exceeds the preset safety threshold, and issue an early warning based on the comprehensive risk probability.

[0080] In this embodiment, the calculation of the comprehensive risk probability of the equipment temperature exceeding a preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance includes: determining the peak point among all optimal temperature state estimates within the prediction time window, and taking the time corresponding to the peak point as the risk characteristic time; calculating the standardized deviation of the optimal temperature state estimate corresponding to the risk characteristic time relative to the preset safety threshold, and performing probability distribution mapping based on the standardized deviation and the estimation error variance corresponding to the risk characteristic time to obtain the extreme risk probability of the actual temperature exceeding the preset safety threshold at the risk characteristic time; calibrating the extreme risk probability using preset model performance parameters to obtain the comprehensive risk probability; wherein, the preset model performance parameters include model baseline confidence and time-dependent confidence weights.

[0081] First, it should be noted that, to address the core pain points of traditional discrete threshold alarm methods—ignoring prediction uncertainty, high false alarm / false negative rates, and lack of risk prediction capabilities—this application innovatively introduces residual probability distribution mapping theory and designs a risk decision-making model based on time-series probability density integrals. This achieves a paradigm shift in risk decision-making from "discrete threshold judgment" to "continuous probability assessment." This layer, based on the optimal temperature state estimate output from the previous stage and its corresponding estimation error variance, quantifies the cumulative risk density exceeding the threshold at future moments by constructing a dynamic probability distribution model, ultimately outputting a risk quantification index with physical and statistical significance.

[0082] Uncertainty modeling based on prior distributions includes the following: This layer receives the optimal temperature state estimate from the physical state correction layer. and the corresponding estimation error variance Based on prior assumptions, the predicted residuals are identified. Follows a mean of 0 and a standard deviation of follows a normal distribution, where If it is true, This represents the optimal temperature estimate; that is, each predicted location corresponds to a dynamically evolving probability density distribution. ; in, For the actual temperature variable, To estimate the standard deviation of the error, which reflects the degree of divergence of the model's predictions at different future times (uncertainty quantification).

[0083] The point out-of-limit probability quantification based on CDF (Cumulative Distribution Function) includes the following: Risk probability exceeding the preset safety threshold Defined as the probability density function at that moment in the interval Integrating over the standard normal cumulative distribution function (CDF), the probability of this location can be expressed as: ; This indicator quantifies the mathematical probability that fluctuations at a single moment will trigger an apparent alarm.

[0084] Understandably, to accurately assess the maximum potential risk within the prediction window (step size H), this application abandons the approach of averaging risk across all moments and instead adopts the worst-case extreme value assessment method. This method focuses on the critical moment with the highest temperature in the predicted trajectory and directly quantifies the highest probability of failure by calculating the standardized deviation (Sigma distance) of this extreme point relative to the threshold.

[0085] Specifically, firstly, within the prediction time window Within this process, the optimal temperature state estimate is iterated through at all times, and the moment with the largest temperature estimate (i.e., the peak point) is identified and used as the risk characteristic moment. ; This moment represents the time within the future prediction time window when the device is closest to or exceeds the safety boundary, i.e., the most dangerous moment.

[0086] Furthermore, the optimal temperature state estimate corresponding to the risk characteristic moment is calculated. The standardized deviation relative to the preset safety threshold, where the standardized deviation is adopted. calculate: ; in, To preset a safety threshold, These represent the optimal temperature state estimate and the corresponding standard deviation of the estimate error.

[0087] Then, based on the standardized deviation and the estimated error variance corresponding to the risk characteristic moment, the extreme risk probability of the true temperature exceeding the preset safety threshold at the risk characteristic moment is obtained through probability distribution mapping. Specifically, the survival function (SF) of the normal distribution is used to calculate the extreme risk probability of the true temperature exceeding the preset safety threshold at the risk characteristic moment. The actual temperature value always falls within the safety threshold. The probability area on the right side (i.e., the region exceeding the limit) is used to obtain the final risk quantification index, namely the extreme value risk probability. : .

[0088] Finally, the extreme value risk probability is calibrated using preset model performance parameters to obtain the comprehensive risk probability; these preset model performance parameters include the model baseline confidence level and the time-dependent confidence weight. In other words, to ensure the engineering reliability of the risk indicators, model performance parameters are introduced to... Perform multi-dimensional calibration to obtain the final comprehensive risk probability. : ; in, The baseline confidence score is an empirical value obtained from offline evaluation. It represents the overall prediction accuracy of the model on historical datasets and reflects the model's long-term reliability and generalization ability. For time-sensitivity confidence weights, and for prediction step sizes... k The weighting coefficients, which decrease dynamically with increasing distance, reflect the characteristic that uncertainty increases as the prediction progresses further away; ultimately, the output... It is a continuous variable between 0 and 1. The system based on... Mapping to preset warning levels drives refined operation and maintenance intervention decisions.

[0089] In a specific implementation, the early warning based on the comprehensive risk probability includes: determining the target threshold range where the comprehensive risk probability is located, and using the risk level corresponding to the target threshold range as the current risk level; wherein, different threshold ranges correspond to different risk levels, multiple threshold ranges are divided based on at least two preset risk level thresholds, and the risk level includes at least a safe state, an early warning state, and an alarm state; statistically analyzing the occurrence frequency of alarm events of different levels within a historical time window, and determining the corresponding decision action based on the current risk level and the occurrence frequency, so as to execute the corresponding early warning action based on the decision action.

[0090] In other words, to overcome the false alarms and false negatives that may be caused by instantaneous probability fluctuations, this system comprehensively considers risk probability. Building upon this foundation, a more robust multi-level decision-making strategy is implemented by introducing alarm history statistics based on time windows. This strategy comprises two levels of logic: the first level makes a preliminary judgment based on the instantaneous risk probability; the second level makes a final decision based on the recent alarm frequency.

[0091] Specifically, an initial risk classification is performed based on the instantaneous risk probability. In this embodiment, the comprehensive risk probability can be compared with at least two preset risk level thresholds to determine the current target threshold range, thereby using the risk level corresponding to the target threshold range as the current risk classification. Different threshold ranges correspond to different risk levels, and multiple threshold ranges are divided based on at least two preset risk level thresholds. The risk classification includes at least a safe state, a warning state, and an alarm state. For example, three basic risk levels can be defined: Safe state: < (For example Warning status: ≤ < (For example Alarm status: ≥ .

[0092] Secondly, incorporating historical alarm frequency into the final decision-making process, the system maintains a sliding historical time window (e.g., the most recent 24 hours) and counts the frequency of alarm events of different levels within the window in real time. (Frequency of early warning status) and (Alarm status). Finally, based on instantaneous classification and historical frequency, the corresponding decision action is determined, and then the corresponding early warning action is executed based on the decision action. In a specific implementation, the final action is determined according to the content of Table 1: Table 1

[0093] in, , For configurable frequency thresholds, such as =1, =3.

[0094] For more detailed processing procedures of steps S21 and S22, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0095] As can be seen, this application discloses a condition-adaptive Kalman filter framework that optimally integrates data-driven preliminary predictions with the physical evolution laws based on a thermodynamic constant-rate model. By dynamically adjusting the fusion weights through a parameter generation network, it ensures that the final predicted trend conforms to both data patterns and basic physical causal constraints. Furthermore, this application discloses a continuous risk quantification and hierarchical early warning mechanism based on residual distribution mapping. Utilizing the optimal state estimate and estimation error variance as a basis, it focuses on the critical moment with the highest temperature in the predicted trajectory. By calculating the standardized deviation of this extreme point relative to a threshold, it directly quantifies the highest probability of failure. In other words, this application enhances the model's sensitivity to high-risk conditions through a cost-sensitive learning mechanism, ensures the rationality of the predicted trend through physical model constraints, and ultimately outputs continuous risk probabilities to achieve intelligent early warning.

[0096] See Figure 5 As shown in the figure, this application discloses a device for predicting and warning equipment temperature, the device comprising: Feature extraction module 11 is used to extract multi-dimensional features from the collected time-series operation data of the target device to obtain a multi-dimensional feature tensor; Temperature prediction module 12 is used to input the multidimensional feature tensor into a trained cost-sensitive temporal convolutional network for temperature prediction, and obtain a preliminary temperature prediction value and a target hidden feature vector; wherein, the cost-sensitive temporal convolutional network is constructed based on several dilated causal convolutional residual blocks, and each dilated causal convolutional residual block includes a dilated causal convolutional layer and a residual connection. The parameter generation module 13 is used to generate noise parameters of a Kalman filter that are adapted to the current operating conditions of the target device based on the target hidden feature vector; The fusion module 14 is used to fuse the preliminary temperature prediction value and the temperature state prediction value based on the noise parameters using the Kalman filter to obtain the optimal temperature state estimate value and the corresponding estimation error variance; the temperature state prediction value is obtained by a preset temperature state transition model based on the optimal temperature state estimate value at the previous moment and the equipment load input at the current moment. The early warning module 15 is used to calculate the comprehensive risk probability of the equipment temperature exceeding the preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance, and to issue an early warning based on the comprehensive risk probability.

[0097] As can be seen, this application, by extracting multi-dimensional features from time-series operational data and inputting them into a pre-trained cost-sensitive temporal convolutional network, outputs preliminary temperature predictions and target hidden feature vectors. This enables the accurate mining of temperature evolution-related features from multi-dimensional equipment operational data, improving the targeting and accuracy of preliminary temperature predictions. Furthermore, by generating Kalman filter noise parameters adapted to the current operating conditions of the equipment based on the target hidden feature vectors, the noise parameters can be dynamically adjusted according to the equipment's operating conditions, ensuring that the Kalman filter's operating parameters match the real-time operating status of the equipment. Moreover, by using a pre-defined temperature state transition model, a temperature state prediction is obtained based on the previous optimal temperature state estimate and the current equipment load input, ensuring that the temperature prediction aligns with physical laws such as the equipment's thermal inertia, avoiding the problem of unreasonable trends in purely data-driven predictions. Combining the noise parameters with the Kalman filter to fuse the data-driven preliminary temperature predictions and the temperature state predictions based on physical mechanisms, the advantages of data features and physical models are complemented, further improving the accuracy and reliability of the optimal temperature state estimate. Finally, based on the optimal temperature state estimate and the variance of the estimation error, the comprehensive risk probability of the equipment temperature exceeding the safety threshold is calculated, thereby achieving a quantitative assessment of temperature risk, making the early warning results more objective and forward-looking, and effectively reducing false alarms and missed alarms.

[0098] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the device temperature prediction and early warning method performed by the electronic device disclosed in any of the foregoing embodiments.

[0099] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0100] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0101] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0102] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device 20 to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. The operating system 221 can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the device temperature prediction and early warning method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0103] Furthermore, embodiments of this application also disclose a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the device temperature prediction and early warning method disclosed in any of the foregoing embodiments.

[0104] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0105] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0106] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, compact disc read-only memory (CD-ROM), or any other form of storage medium known in the art.

[0107] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0108] The above provides a detailed description of the device temperature prediction and early warning method, apparatus, equipment, and storage medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of the present invention. 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 the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting and warning equipment temperature, characterized in that, include: Multi-dimensional feature extraction is performed on the time-series operation data of the collected target device to obtain a multi-dimensional feature tensor; The multidimensional feature tensor is input into a trained cost-sensitive temporal convolutional network for temperature prediction, resulting in a preliminary temperature prediction value and a target hidden feature vector. Based on the target hidden feature vector, noise parameters of a Kalman filter adapted to the current operating condition of the target device are generated; The Kalman filter is used to fuse the preliminary temperature prediction value and the temperature state prediction value based on the noise parameters to obtain the optimal temperature state estimate value and the corresponding estimation error variance; the temperature state prediction value is obtained by a preset temperature state transition model based on the optimal temperature state estimate value at the previous moment and the equipment load input at the current moment. Based on the optimal temperature state estimate and the corresponding estimation error variance, the comprehensive risk probability of the equipment temperature exceeding the preset safety threshold is calculated, and an early warning is issued based on the comprehensive risk probability.

2. The equipment temperature prediction and early warning method according to claim 1, characterized in that, The training process of the cost-sensitive temporal convolutional network includes: Multiple training samples are determined; each training sample includes historical time-series running data and a corresponding future real temperature sequence, wherein the historical time-series running data includes a historical temperature data sequence corresponding to the temperature dimension; A comprehensive cost-sensitive weight matrix is ​​constructed based on the historical temperature data sequence in each training sample; wherein, the comprehensive cost-sensitive weight matrix is ​​used to amplify the loss contribution of the preset high temperature range and preset heating stage during the training process; For each training sample, a multi-dimensional feature tensor extracted from the historical time-series running data in the training sample is input into the time-series convolutional network to be trained, so as to output the preliminary predicted temperature sequence corresponding to the training sample. The target loss value is obtained by weighting the error loss value between the preliminary predicted temperature sequence and the future actual temperature sequence using the comprehensive cost-sensitive weight matrix. The network parameters of the temporal convolutional network are updated based on the target loss value until the preset training termination condition is met, thus obtaining a trained cost-sensitive temporal convolutional network.

3. The equipment temperature prediction and early warning method according to claim 2, characterized in that, The construction of the comprehensive cost-sensitive weight matrix based on the historical temperature data sequence in each training sample includes: Obtain the historical temperature data sequence for each training sample; The absolute deviation between the historical temperature value at each moment in the historical temperature data sequence and the preset high temperature alarm threshold is converted into a basic weight value, and the basic weight value is trimmed to obtain the absolute temperature level cost weight. The first temperature trend sequence is obtained by calculating the first difference along the time dimension of the historical temperature data sequence, and the first temperature trend sequence is converted into a warming trend cost weight. The absolute temperature level cost weight and the warming trend cost weight are fused to obtain a comprehensive cost-sensitive weight matrix.

4. The equipment temperature prediction and early warning method according to claim 2, characterized in that, The step of using the comprehensive cost-sensitive weight matrix to weight the error loss value between the preliminary predicted temperature sequence and the future actual temperature sequence to obtain the target loss value includes: The first-order difference between the preliminary predicted temperature sequence and the future actual temperature sequence is calculated along the time dimension to obtain the corresponding second temperature trend sequence and third temperature trend sequence. A first mask is constructed based on the non-zero elements in the third temperature trend sequence; The second temperature trend sequence and the third temperature trend sequence are multiplied element by element, and a second mask is constructed based on the elements less than zero in the product result. The first mask and the second mask are multiplied element-wise to obtain the target mask. Calculate the absolute error loss value between the preliminary predicted temperature sequence and the future actual temperature sequence, and generate an enhanced error loss value based on the absolute error loss value and the target mask; The enhancement error is weighted using the comprehensive cost-sensitive weight matrix to obtain the target loss value.

5. The equipment temperature prediction and early warning method according to claim 1, characterized in that, The noise parameters include process noise covariance parameters and observation noise covariance parameters; The generation of noise parameters for a Kalman filter adapted to the current operating condition of the target device based on the target hidden feature vector includes: The target hidden feature vector is reduced in dimensionality and aggregated in terms of time dimension to obtain aggregated features; The aggregated features are subjected to a nonlinear transformation to obtain the transformed features; The transformation features are respectively input to the first generation channel and the second generation channel of the trained parameter generation network, so as to output the original process noise covariance of the Kalman filter through the first generation channel and the original observation noise covariance of the Kalman filter through the second generation channel. Positive definite constraints are applied to the original process noise covariance and the original observation noise covariance, and the parameters after positive definite constraints are mapped to a numerical range that conforms to the thermal evolution law of the equipment, so as to obtain process noise covariance parameters and observation noise covariance parameters that are adapted to the current operating conditions.

6. The equipment temperature prediction and early warning method according to claim 1, characterized in that, The step of fusing the preliminary temperature prediction value and the temperature state prediction value of the preset temperature state transition model based on the noise parameters using the Kalman filter to obtain the optimal temperature state estimate and the corresponding estimation error variance includes: The preliminary temperature prediction is used as the virtual observation value of the Kalman filter; Based on the optimal temperature state estimate of the previous moment, the preset temperature state transition model, and the equipment load input at the current moment, the prior temperature estimate and prior estimate error covariance at the current moment are calculated. The Kalman gain is calculated based on the noise parameters and the prior estimation error covariance. The Kalman gain is then used to fuse the virtual observations and the prior temperature estimates to obtain the optimal temperature state estimate and the corresponding estimation error variance.

7. The equipment temperature prediction and early warning method according to claim 1, characterized in that, The calculation of the comprehensive risk probability of the equipment temperature exceeding the preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance includes: Determine the peak point among all optimal temperature state estimates within the prediction time window, and use the time corresponding to the peak point as the risk characteristic time. Calculate the standardized deviation of the optimal temperature state estimate corresponding to the risk characteristic moment from the preset safety threshold, and perform probability distribution mapping based on the standardized deviation and the estimation error variance corresponding to the risk characteristic moment to obtain the extreme risk probability of the actual temperature exceeding the preset safety threshold at the risk characteristic moment; The extreme value risk probability is calibrated using preset model performance parameters to obtain a comprehensive risk probability; wherein, the preset model performance parameters include the model baseline confidence level and the time-dependent confidence weight.

8. The equipment temperature prediction and early warning method according to any one of claims 1 to 7, characterized in that, The early warning based on the comprehensive risk probability includes: The target threshold range in which the comprehensive risk probability is located is determined, and the risk level corresponding to the target threshold range is taken as the current risk level; wherein, different threshold ranges correspond to different risk levels, and multiple threshold ranges are divided based on at least two preset risk level thresholds, and the risk level includes at least a safe state, a warning state, and an alarm state. The frequency of alarm events of different levels within a historical time window is statistically analyzed, and a corresponding decision action is determined based on the current risk level and the frequency of occurrence, so as to execute the corresponding early warning action based on the decision action.

9. A device for predicting and warning equipment temperature, characterized in that, include: The feature extraction module is used to extract multi-dimensional features from the collected time-series operation data of the target device to obtain a multi-dimensional feature tensor. The temperature prediction module is used to input the multidimensional feature tensor into a trained cost-sensitive temporal convolutional network to predict the temperature and obtain a preliminary temperature prediction value and a target hidden feature vector. The parameter generation module is used to generate noise parameters of a Kalman filter that are adapted to the current operating conditions of the target device based on the target hidden feature vector; The fusion module is used to fuse the preliminary temperature prediction value and the temperature state prediction value based on the noise parameters using the Kalman filter to obtain the optimal temperature state estimate value and the corresponding estimation error variance; the temperature state prediction value is obtained by a preset temperature state transition model based on the optimal temperature state estimate value at the previous moment and the equipment load input at the current moment. The early warning module is used to calculate the comprehensive risk probability of the equipment temperature exceeding the preset safety threshold based on the optimal temperature state estimate and the corresponding estimation error variance, and to issue an early warning based on the comprehensive risk probability.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the device temperature prediction and early warning method as described in any one of claims 1 to 8.

Images