An lstm-based high-temperature sintering furnace temperature anomaly early warning method

Abstract

The application discloses a high-temperature sintering furnace temperature abnormality early warning method based on an LSTM, and comprises the following steps: A1, collecting multi-zone temperature sequences, pressure sequences and current sequences, and performing data cleaning on the sequences respectively; A2, performing time axis alignment, windowing, intercepting and arranging to obtain preprocessed multi-zone temperature sequences, preprocessed pressure sequences and preprocessed current sequences, and constructing a spatial topology matrix; A3, extracting an enhanced multi-zone temperature window hidden state sequence, a spatially enhanced temperature time sequence hidden state and a temperature local trend deviation feature; A4, extracting temperature-pressure cross-modal features and temperature-current cross-modal features; A5, performing interaction enhancement to obtain multi-variable collaborative abnormality features; A6, extracting time sequence abnormality discrimination features and final time sequence abnormality discrimination features; and A7, calculating an abnormality probability score and an early warning trigger flag. The application can solve the problems that the high-temperature sintering furnace alarm mode is strong in hysteresis and early thermal field abnormalities are difficult to identify.

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology for sintering furnaces, and in particular to a method and system for early warning of abnormal temperatures in high-temperature sintering furnaces based on LSTM. Background Technology

[0002] In metal powder metallurgy and metal powder injection molding processes, high-temperature sintering is the core process that determines the density and mechanical properties of the finished product. The metal billet needs to undergo degreasing, heating, high-temperature holding, and controlled cooling under a protective atmosphere. The high-temperature section typically exceeds 1000 degrees Celsius, and multiple heating elements in the furnace work together to maintain thermal uniformity. Compared with ceramic sintering, metal sintering involves higher operating temperatures and more drastic dynamic changes in the thermal field. Furthermore, metal powders have a narrower tolerance window for temperature deviations. Once temperature runaway occurs, it can easily cause overheating, deformation, or performance defects in the entire batch of workpieces. Therefore, more stringent requirements are placed on the early identification and rapid response to abnormal trends.

[0003] Currently, temperature monitoring in metal sintering furnaces generally relies on alarm mechanisms based on upper and lower thresholds. Operators set permissible temperature deviation ranges for each temperature zone on the monitoring terminal, and an audible and visual alarm is triggered when the actual temperature exceeds this range. This method suffers from a lag in engineering practice: by the time an alarm signal is issued, the furnace temperature has already deviated substantially, and the thermal imbalance has often persisted for several minutes. At this point, the heating element's adjustment response still requires a certain thermal inertia delay, making remedial measures difficult to take effect before the workpiece is damaged. Furthermore, threshold settings depend on the operator's personal experience; excessively wide thresholds increase the false alarm rate and safety hazards, while excessively narrow thresholds lead to frequent false alarms and operator fatigue in responding to alarm signals, thus exacerbating the uncertainty of the production process to some extent.

[0004] Some studies have attempted to introduce deep learning techniques to predict furnace temperature. For example, prediction methods based on convolutional neural networks can extract local morphological features of temperature curves, but they are insufficient in capturing the long-term thermal inertia effects unique to high-temperature sintering processes. While traditional recurrent neural networks possess temporal modeling capabilities, the vanishing gradient problem leads to low sensitivity to slowly evolving early signs of thermal imbalance when faced with long input sequences of hundreds of sampling periods. Most of the above solutions only utilize temperature as a single parameter, failing to incorporate auxiliary information such as current signals reflecting the state of heating elements and pressure signals characterizing atmosphere flow properties into the modeling scope, thus weakening the model's ability to represent abnormal operating conditions involving multiple coupled factors. Summary of the Invention

[0005] In view of this, the present invention aims to provide an early warning method for temperature anomalies in high-temperature sintering furnaces based on LSTM, so as to solve the problems of strong lag and difficulty in identifying early thermal field anomalies in traditional threshold alarm methods, while overcoming the shortcomings of existing deep learning methods in long-term multi-parameter fusion modeling.

[0006] A method for early warning of temperature anomalies in high-temperature sintering furnaces based on LSTM, comprising: A1: Collect multi-zone temperature sequence, pressure sequence, and current sequence and preprocess them respectively to obtain the cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence. A2: Time axis alignment, windowing, and arrangement of the cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence are performed to obtain the pre-processed multi-zone temperature sequence, pre-processed pressure sequence, and pre-processed current sequence, and a spatial topology matrix is ​​constructed based on the spatial layout of the thermocouples in the furnace. A3: Based on the preprocessed multi-zone temperature sequence and spatial topology matrix, extract the enhanced multi-zone temperature window hidden state sequence, spatially enhanced temperature temporal hidden state, and temperature local trend deviation features in sequence. A4: Based on the local temperature trend deviation characteristics, the preprocessed pressure sequence, and the preprocessed current sequence, extract the temperature-pressure cross-mode features and the temperature-current cross-mode features, respectively. A5: Based on the temperature-pressure cross-modal characteristics and the temperature-current cross-modal characteristics, interactive enhancement is performed to obtain multivariate collaborative anomaly characteristics; A6: Based on the local temperature trend deviation characteristics and multivariate collaborative anomaly characteristics, extract the time-series anomaly discrimination features and the final time-series anomaly discrimination features in sequence; A7: Calculate the anomaly probability score and early warning trigger flag based on the final time-series anomaly discrimination characteristics.

[0007] Furthermore, step A1 also includes: A11: Multi-zone temperature sequences are synchronously acquired by thermocouple arrays arranged in each temperature zone of the high-temperature sintering furnace. The data type is multi-dimensional time-series numerical data, including temperature sampling values ​​of each temperature zone. Outliers are removed from the multi-zone temperature sequences using the three-standard-deviation method, and then missing values ​​generated after removal are filled in using the linear interpolation method to obtain the cleaned multi-zone temperature sequences. A12: The pressure sequence is acquired through the furnace pressure transmitter, and the current sequence is acquired through the heating circuit current transformer. Both data types are one-dimensional time-series numerical data. The pressure sequence includes furnace pressure sampling values, and the current sequence includes heating circuit current sampling values. Outliers are removed from the pressure sequence and the current sequence using the three-standard-deviation method, and then the missing values ​​generated after removal are filled in using the linear interpolation method to obtain the cleaned pressure sequence and the cleaned current sequence.

[0008] Furthermore, step A2 also includes: A21: The cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence are aligned along the time axis. Then, according to a fixed window duration and sliding step size, segments are extracted sequentially along the time axis. The extracted equal-length time series subsequences are then arranged in the original sampling time order to obtain the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence. Each of the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence contains multiple single-window time series subsequences. A22: Based on the spatial layout of the thermocouples in the furnace, a spatial topology matrix is ​​constructed. The construction method includes: using each thermocouple as a matrix node, constructing the spatial topology matrix according to the arrangement order of the nodes; wherein, for two nodes with a direct spatial adjacency relationship, the corresponding matrix elements are set to connected values; for two nodes without a direct spatial adjacency relationship, the corresponding matrix elements are set to disconnected values, and the diagonal elements of the matrix corresponding to the node itself are set to connected values.

[0009] Furthermore, step A3 also includes: A31: Based on the preprocessed multi-zone temperature sequence, it is encoded using a time-series feature extraction path to obtain the enhanced multi-zone temperature window hidden state sequence. The calculation method is as follows: ; ; in, The window is initially hidden. This is for bidirectional long short-term memory network operations. For layer normalization operation, For multilayer sensor operation, This is a single-window time series subsequence from the preprocessed multi-zone temperature sequence. For the enhanced multi-zone temperature window hidden state sequence, This is a random deactivation operation. Let Gaussian error be the activation function of the linear unit. It is a fully connected layer; A32: Based on the spatial topology matrix and the enhanced multi-region temperature window hidden state sequence, feature co-modeling is performed using a graph attention mechanism to obtain the spatially enhanced temperature temporal hidden state. The calculation method is as follows: ; ; in, Here is the spatial attention weight matrix. This is for normalizing the exponential function operation. To correct the leakage of the linear unit activation function, These are the linear projection operations for the query vector and the key vector, respectively. For transpose operation, For Hadama accumulation, It is a spatial topological matrix. This refers to spatially enhanced temperature-temporal hidden states. It is the activation function of the exponential linear unit. These are linear projection operations on the value vector and the residual vector, respectively; A33: Based on the spatially enhanced temporal hidden state of temperature, multi-scale trend differencing is used to compare and decode the current time with the historical baseline, obtaining the local temperature trend deviation characteristics. The calculation method is as follows: ; ; in, For multi-scale gating weights, for Type activation function, The core size is respectively One-dimensional convolution operation, This is a global average pooling operation. This indicates a local deviation in temperature trends. For the core size is One-dimensional causal convolution operation.

[0010] It should be further explained that in heating scenarios such as injection molding and heat treatment, the temperature field is determined by two physical mechanisms: temporal dynamics and spatial coupling. The temperature rise in a certain temperature zone will be conducted to adjacent temperature zones through the mold or furnace, forming normal lag temperature fluctuations. The normal heat conduction between adjacent temperature zones and the local abnormal temperature rise in a single zone are very close in temperature value. It is impossible to distinguish whether the local reading of a single temperature zone is the lag effect of cross-zone conduction or the local offset caused by the aging of the heating rod. If the global average temperature is used for monitoring, the local overheating signal will be submerged by most normal temperature zones, resulting in missed alarms. If the threshold alarm is set independently for each temperature zone, normal cross-zone temperature fluctuations will be misjudged as faults, leading to frequent unplanned shutdowns of the production line. The A31 sub-step first encodes the preprocessed multi-zone temperature sequence: a multilayer perceptron is used to realize cross-channel interaction between temperature zones and complete the initial feature upscaling; layer normalization is used to balance the feature distribution within each window, enabling the subsequent bidirectional LSTM to stably scan the temporal window of each temperature zone; bidirectional scanning simultaneously captures the rapid response features of the heating phase and the slow decay features of the cooling phase, encoding the asymmetric dynamic law of heating and cooling into a unified initial window hidden state; then, GELU activation, nonlinear mapping of fully connected layers, random deactivation are applied to this hidden state in sequence, and then the normalized residual connections are superimposed to obtain the enhanced multi-zone temperature window hidden state sequence; this step enables subsequent spatial modeling to be based on a unified temporal context—not recording the temperature value at a certain moment in isolation, but compressing the complete heating, holding, and cooling trajectory of each temperature zone within the window into a feature vector, providing a temporal dimension comparison benchmark for subsequent judgment of whether the current fluctuation of a certain temperature zone is a local anomaly or a normal lag of cross-zone conduction; random deactivation helps to suppress overfitting and improve the stability of the model in extracting temperature patterns with limited samples; The A32 sub-step introduces a spatial topology matrix to explicitly encode the physical adjacency relationships and heat conduction paths of each temperature zone, thus limiting the graph attention mechanism's attention allocation to physically adjacent temperature zones. When the temperature of zone 3 rises, the mechanism actively evaluates the temperature change trends of its neighboring zones 1 and 2 within the same window. If zone 1 also has a temperature rise record, the attention weights will automatically suppress the abnormal response of zone 3's current temperature. Specifically, similarity scores are calculated between pairwise temperature zone features through the linear projection of the query vector and the key vector. The Hadamard product of the spatial topology matrix forces the attention weights of irrelevant temperature zones to zero. Finally, after Softmax normalization, the output... A spatial attention weight matrix is ​​derived. This weight matrix is ​​used to weight and aggregate the linear projection results of the value vector. Combined with the linear projection of the residual vector, cross-regional neighborhood information is supplemented and fused. After ELU activation and layer normalization, the spatially enhanced temperature temporal hidden state is obtained. In this way, each temperature region can actively refer to the synchronous temperature behavior of its physical neighbors when judging its own state. When the temperature change of a certain temperature region can be reasonably explained by the heat conduction of its neighbors, the abnormal response in its hidden state is naturally attenuated. Conversely, when the temperature change cannot be explained by the state of its neighbors, the local offset component of that temperature region remains significant in the hidden space, thereby eliminating false alarms caused by normal heat conduction in the spatial dimension. Substep A33 performs a bias analysis based on historical trends on the spatially enhanced temperature temporal hidden states. Two different kernel sizes of one-dimensional convolutions are used to capture short-term fluctuations and medium-to-long-term trend patterns, respectively. The output, after GELU activation and one-dimensional convolution processing, is added to the global statistical features obtained by global average pooling, and then multi-scale gating weights are generated using Sigmoid. Simultaneously, one-dimensional causal convolution is used to fit the historical trends of the hidden states—causal convolution only uses historical information before the current time step to construct the normal evolutionary benchmark that each temperature region should follow under spatial constraints. The difference term is obtained by subtracting the current hidden state from this causal historical benchmark, and adaptively filtered out the key scales using the Hadamard product of multi-scale gating weights. After the deviation signal is superimposed with a linear projection to form a complete residual connection, the local temperature trend deviation feature is output after layer normalization. This feature does not simply compare the absolute temperature value, but judges whether the current temperature change deviates from the historical evolution trajectory of each temperature zone under the constraint of spatial neighborhood. If the heating rate of a certain temperature zone is consistent with its historical heating pattern, and the neighboring states also conform to the conduction logic, the deviation feature is suppressed. If the heating rate significantly exceeds the historical normal range, the deviation feature can still respond even if the absolute temperature value is still within the normal range. This step realizes early warning of local overheating from the time-series dimension - before the temperature reaches the absolute threshold, it identifies abnormal signs in advance by the behavioral characteristics of deviation from the historical trend. Existing technologies typically employ two approaches: independent thresholds for each temperature zone or global average temperature monitoring. The former fails to consider the thermal conduction coupling between adjacent temperature zones, misjudging normal cross-zone lag temperature fluctuations as single-zone anomalies, resulting in a high false alarm rate. While the latter provides indication of global overheating, the averaging operation smooths out local anomalies, causing slow temperature drifts caused by heating rod aging in a single zone to go undetected. This invention models bidirectional temporal dependencies as hidden window states through step A31, allowing the asymmetric laws of heating and cooling to be characterized separately. Step A32 introduces a graph attention mechanism with spatial topological constraints, enabling each temperature zone to actively reference the states of its physical neighbors, distinguishing between normal thermal conduction and local anomalies in a single zone. Step A33 uses historical causal trends for comparative decoding, allowing local overheating to be captured by deviation features before the absolute temperature value exceeds the limit. These three steps work together, under the dual constraints of spatial neighborhood constraints and temporal evolution benchmarks, to accurately identify true single-zone anomalies—avoiding both false shutdowns due to normal thermal conduction and missed detections due to averaging smoothing. This ensures that the anomaly monitoring of multi-zone heating systems remains sensitive to actual faults with higher process tolerance.

[0011] Furthermore, step A4 also includes: A41: Based on the local temperature trend deviation features and the preprocessed pressure sequence, a gated time-delay convolutional alignment mechanism is constructed to extract features from the preprocessed pressure sequence. These features are then asynchronously correlated and encoded with the local temperature trend deviation features to obtain the temperature-pressure cross-modal features. The calculation method is as follows: ; ; ; in, This is a hidden state encoded for pressure timing. Let Gaussian error be the activation function of the linear unit. The expansion rates are respectively The one-dimensional dilated convolution operation, This is a single-window time series subsequence in the preprocessed pressure sequence. This represents the global temperature deviation characteristic after pooling. For temperature Pressure cross-modal characteristics; A42: Based on the preprocessed current sequence and local temperature trend deviation characteristics, a channel-normalized time-domain convolution backbone is used to capture abrupt change features of the current sequence, and a current-temperature dual-modal response mechanism is constructed to obtain temperature-current cross-modal features. The calculation method is as follows: ; ; in, This is a hidden state for current timing encoding. For layer normalization operation, For one-dimensional convolution operations with channel normalization, This is a single-window time-series subsequence in the preprocessed current sequence. Let Gaussian error be the activation function of the linear unit. For temporal convolutional network operations, For temperature Current cross-mode characteristics, It is the hyperbolic tangent activation function.

[0012] It should be further explained that in online monitoring scenarios of industrial manufacturing, heating elements degrade after long-term operation. At this time, the controller will automatically increase power compensation to maintain the set temperature, resulting in temperature readings that appear normal but the actual current curve shows abnormal fluctuations. When a small leak occurs in the mold sealing structure, the temperature can still be maintained within the process window, but the physical coupling relationship between pressure and temperature has changed—pressure changes are no longer synchronized with temperature. Single temperature modes have inherent limitations and cannot distinguish whether the temperature rise is caused by normal physical coupling or by coupling anomalies, which directly confuses subsequent anomaly judgment. At the same time, there are asynchronous and nonlinear associated failure modes between the signals collected by pressure sensors and current sensors. Conventional methods lack effective cross-modal coupling modeling capabilities and cannot convert the degree of temperature anomaly in the context of other physical quantities into calculable and quantifiable features. Taking the construction of temperature-pressure cross-modal features as an example, A41 combines the pressure sequence processed by dilated convolution with the pressure sequence processed by a bidirectional long short-term memory network using residuals to obtain the pressure temporal encoding hidden state. Dilated convolution efficiently captures long-term dependency patterns across different time spans in the pressure sequence through different dilation rates, while the bidirectional long short-term memory network models the temporal change trend of pressure in both forward and backward directions. The complementarity of these two features allows the pressure encoding to possess both a global temporal receptive field and the ability to perceive the direction of local fluctuations. Based on this, the local temperature trend deviation features are compressed into global temperature deviation features through global average pooling. This compression operation preserves the overall strength of the temperature deviation while eliminating the influence of the deviation's location. This global temperature deviation feature and the pressure encoding hidden state are then fed into a linear transformation and processed by S... The igmoid function generates gating weights, which are then multiplied element-wise with the pressure-encoded hidden state to achieve selective feature modulation. The gating weights are dynamically determined by the degree of temperature deviation: when the temperature deviation is significant, the gating value approaches 1, amplifying the response of the pressure feature during that period; when the temperature is stable, the gating value approaches 0, suppressing irrelevant fluctuation information in the pressure feature. This allows the final temperature-pressure cross-modal features to adaptively interpret the degree of anomaly in the pressure time series within the temperature deviation context, transforming the cross-modal asynchronous coupling relationship that originally required human experience into a quantitative feature that can be computed by the machine. For example, in the early stage of mold leakage, the coupling slope between pressure and temperature changes, and the gating mechanism amplifies this change and reflects it in the feature value, so that the anomaly can be perceived by the downstream classifier. In current processing, the A42 employs one-dimensional convolution with channel normalization to capture current spikes, while using a temporal convolutional network to model the long-range temporal dependence of the current waveform. The channel normalization mechanism adapts to variations in the amplitude range of the current signal under different operating conditions, ensuring that the response sensitivity of the convolution operation to spikes is unaffected by fluctuations in the signal baseline. Furthermore, the current-encoded hidden state and global temperature deviation features are linearly transformed and then processed by a hyperbolic tangent function. The positive and negative symmetry of the hyperbolic tangent function's output range (-1 to 1) perfectly matches the current signal... The polarity reversal characteristic of the signal—when the temperature deviation is in the same direction as the current anomaly, the fused response is positive, and when the direction is opposite, it is negative; after multiplying the fused response element-by-element with the current-encoded hidden state and then normalizing it, the temperature-current cross-mode feature is obtained; this feature can identify the controller's implicit power compensation behavior—when the heating element degrades, the controller gradually increases the power, the current base value slowly rises but the temperature remains unchanged. In this scenario, the fused response in the temperature-current cross-mode feature will change from positive to negative, thus transforming this implicit degradation, which cannot be identified by conventional single temperature threshold detection, into a feature difference that can be discerned by the system; In existing technologies, anomaly detection for multi-sensor signals typically involves feeding the signals from various physical quantity sensors into the same time-series network for unified modeling. This ignores the fact that different physical quantities have different time-series response speeds and physical coupling mechanisms. When cross-modal coupling failures such as heating element degradation and mold leakage occur, temperature, pressure, and current do not change synchronously. Unified modeling can confuse coupling anomalies with irrelevant noise. This invention employs a gated modulation approach: using temperature deviation characteristics to guide the interpretation of pressure and current characteristics—actively amplifying the response when the temperature deviation is significant and actively suppressing interference when the temperature is stable. This achieves temperature-sensed pressure interpretation and temperature-sensed current interpretation, enabling temperature-pressure cross-modal characteristics and temperature-current cross-modal characteristics to encode the degree of pressure anomaly and the degree of current anomaly under temperature deviation conditions, respectively. This transforms the cross-modal coupling anomaly judgment, which originally relied on the experience of process experts, into machine-comparable quantitative characteristics.

[0013] Furthermore, step A5 also includes: First, the temperature-pressure cross-mode features and the temperature-current cross-mode features are concatenated. Then, the features are processed sequentially through a linear layer, a GeLU activation function, a linear layer, and a Sigmoid function to obtain the cross-modal competitive gating coefficients. The temperature-pressure cross-modal features are processed through a linear layer and then multiplied by the Hadamard product with the cross-modal competitive gating coefficients. The temperature-current cross-modal features are processed through a linear layer and then multiplied by the Hadamard product with the result minus the cross-modal competitive gating coefficients. The temperature-pressure cross-modal features and the temperature-current cross-modal features are concatenated and processed through a linear layer and the GeLU activation function. The results of the above three steps are added together and then normalized by layers to obtain the multivariate collaborative anomaly features.

[0014] It should be further explained that A5 first concatenates the temperature-pressure cross-modal features with the temperature-current cross-modal features. After processing by a multilayer perceptron and Gaussian error linear unit activation function, the cross-modal competition gating coefficient C is output through an sigmoid function. The coefficient takes values ​​between zero and one. After linearly projecting the coefficient C onto the temperature-pressure branch, a Hadamard product is applied for weighting. After subtracting the coefficient C from the linear projection onto the temperature-current branch, a Hadamard product is applied for weighting. The weights of the two branches increase and decrease inversely, forming a zero-sum competition, forcing the model to make a clear route selection at each time step. In addition, the soft weighting design allows the gradient to be smoothly backpropagated along the two branches. The weight updates are robust during training, and the convergence speed is improved compared to hard routing. The overall anomaly trend is extracted from the concatenated features by combining the benchmark information. When a branch fails due to sensor noise or sudden changes in operating conditions, the system can still capture weak anomalies from the overall coupling state.

[0015] Furthermore, step A6 also includes: First, the local temperature trend deviation features are processed by a linear layer. At the same time, the local temperature trend deviation features are used as query vectors, and the multivariate collaborative anomaly features are used as key vectors and value vectors for multi-head attention operation. The results of the above two items are added together and then processed by layer normalization and GeLU function in sequence to obtain the hidden state of the temperature branch with cross-enhancement. The multivariate collaborative anomaly features are then processed through a linear layer. Simultaneously, a multi-head attention operation is performed on the multivariate collaborative anomaly features as query vectors and the temperature local trend deviation features as key and value vectors. The results of the above two items are added together and then processed by layer normalization and GeLU function in sequence to obtain the cross-enhanced collaborative branch hidden state. The hidden states of the temperature branch enhanced by cross-enhancement are concatenated with the hidden states of the cooperative branch enhanced by cross-enhancement. The concatenated result is processed by a linear layer to obtain the first linear intermediate result, and by a linear layer, GeLU function, and Dropout operation to obtain the second linear intermediate result. The first linear intermediate result and the second linear intermediate result are then added together and processed by layer normalization to obtain the time-series anomaly discrimination features.

[0016] Furthermore, step A6 also includes: First, the time-series anomaly discrimination features and the local temperature trend deviation features are concatenated, and then processed sequentially through a linear layer, a GeLU function, a linear layer, and a Sigmoid function to obtain the temperature branch gating weights. The time-series anomaly discrimination features and multivariate collaborative anomaly features are then concatenated and processed sequentially through a linear layer, a GeLU function, another linear layer, and a Sigmoid function to obtain the collaborative branch gating weights. Then, the time-series anomaly discrimination features are processed through three linear layers to obtain the first linear result, the second linear result, and the third linear result. The temperature branch gating weights are multiplied by the first linear result, and the cooperative branch gating weights are multiplied by the second linear result. The two Hadamard product results are then added to the third linear result. Finally, the results are normalized to obtain the final time-series anomaly discrimination features.

[0017] It should be further explained that in industrial process monitoring, the local trend deviation signal and the multivariate collaborative anomaly signal collected by the temperature sensor reflect the system's operating status from different physical dimensions. The temperature signal is easily affected by environmental noise and operating condition switching, exhibiting periodic fluctuations. The multivariate collaborative signal relies on the coupling relationship between multiple sensor readings, and its anomaly criteria change dynamically with the shift in operating conditions. The two signals have different perspectives on the time scale: the temperature branch retains finer instantaneous fluctuation details but lacks a global collaborative view; the collaborative branch compresses local time information through pooling operations but is difficult to detect millisecond-level initial anomalies. The blind spots of the two types of information make it difficult for a single feature channel to independently complete accurate anomaly discrimination. The temperature branch and the collaborative branch carry feature information at different scales; the local temperature trend deviation feature retains fine-grained temporal fluctuation details, and can perceive minute but difficult-to-qualify temperature drifts; the multivariate collaborative anomaly feature of the collaborative branch aggregates the global coupling relationship across sensor channels, and can determine whether there is a structural break in the overall collaboration, but loses the instantaneous signal details at the start time; in step A61, this invention establishes a bidirectional information guidance path with a multi-head attention mechanism: using the local trend feature of the temperature branch as the query vector to retrieve the global features of the collaborative branch, and simultaneously using the global features of the collaborative branch as the query vector to retrieve the local detail features of the temperature branch; this cross-query allows the temperature branch to obtain the injection of the global collaborative vision while maintaining its own fine-grained features—when the pressure and current coupling change causes a normal temperature rise, the corresponding coupling response mode can be found in the collaborative branch, assisting the temperature branch. This identifies such fluctuations as normal rather than abnormal drifts. When an isolated temperature drift occurs due to an anomaly in the sensor or device itself, a corresponding response pattern is usually not found in the collaborative features, thus assisting the temperature branch in identifying such drifts as true anomalies. Conversely, the collaborative branch also injects the local temporal details of the temperature branch into its own global features: multi-layer pooling causes the collaborative features to lose the initial anomalous signal during coarsening, and the injection of the temperature branch allows the collaborative branch to recapture the instantaneous anomalous details that appeared at the beginning of sampling, improving the sensitivity to early minor anomalies. After cross-modal bidirectional information guidance, the latent states of the two branches each incorporate the complementary information of the other channel. The two are added bit by bit and fused, and after linear transformation and random discarding, the temporal anomaly discrimination feature is obtained. This feature has both fine-grained local evidence and global collaborative evidence, and is complete information of cross-modal coupling rather than a simple feature splicing. Although the latent state after cross-modal information injection possesses dual-view capability, the contribution weights of temperature evidence and collaborative evidence to anomaly detection dynamically change under different operating conditions. Significant temperature rises caused by sudden pressure surges require greater reliance on collaborative branches to eliminate false positives; slow temperature deviations caused by sensor drift require greater reliance on local details of the temperature branch for confirmation. In step A62, this invention concatenates local temperature trend features and multivariate collaborative anomaly features with time-series anomaly detection features, obtaining temperature branch gating weights and collaborative branch gating weights through two layers of nonlinear mapping. These two gating weights are then used to weightedly fuse the time-series anomaly detection features. A direct connection path without gating is retained; when the temperature branch contributes significantly to the current judgment, the gating weight actively amplifies the back-transmitted signal of temperature evidence; when the cooperative branch is more reliable, the cooperative evidence is adopted first; the direct connection path ensures that basic feature information can still pass through without being truncated under conditions of ambiguous evidence—such as the chaotic period caused by the simultaneous and drastic changes of various parameters during equipment startup; this dynamic gating fusion enables the final time-series anomaly discrimination features to maintain a reasonable ratio of evidence sources under different operating conditions, reducing the risk of misjudgment caused by pressure mutations and the risk of missed detection due to slow change drift, and improving the robustness of anomaly detection in multi-condition switching scenarios in industrial processes; In existing technologies, most anomaly detection methods based on univariate temperature time series only focus on the magnitude of trend deviation and cannot distinguish between normal pressure-coupled temperature rise and true anomalous drift. When a pressure change is caused by a shift in operating conditions, the temperature curve naturally produces a peak. Univariate methods are prone to misjudging such coupled fluctuations as anomalies, resulting in a high false positive rate. The present invention achieves bidirectional feature guidance between the temperature branch and the cooperative branch through a cross-modal multi-head attention mechanism. The temperature branch uses the global information of the cooperative branch to filter the coupling noise in its own signal, while the cooperative branch obtains lost local time series details from the temperature branch. Then, the two information paths are weighted and fused by a dynamic gating mechanism that is adaptive to operating conditions. The direct connection path provides a fallback response in the case of ambiguous evidence. At the same time, it solves the problem of information blind spots in single-channel features and the dynamic trade-off problem of the primary and secondary relationships of multiple channels changing with operating conditions. The source ratio for the final anomaly judgment is automatically adjusted according to the current data features, which has a more robust anomaly identification capability in multi-operating-condition scenarios in industrial processes than existing methods.

[0018] Furthermore, step A7 also includes: First, the temporal anomaly discrimination features at time t are processed sequentially by layer normalization, GeLU function activation, and multilayer perceptron to obtain the gated enhancement features at time t. Then, the gated enhancement features are multiplied by the score projection weight matrix, the score projection bias vector is added, and then processed by the Sigmoid function to obtain the anomaly probability score at time t. Based on the anomaly probability score sequence at all times, a quantile calculation function is used to determine the dynamic anomaly threshold with a preset anomaly sensitivity percentile parameter. Then, for each time moment, an indicator function is used to determine whether the anomaly probability score at that time is greater than the dynamic anomaly threshold. If it is, the warning trigger flag at that time is set to true; otherwise, it is set to false.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In view of the lag in traditional threshold alarm methods in high-temperature scenarios of metal sintering furnaces and the insufficient identification of early thermal anomaly signs by existing deep learning methods, this invention integrates multi-source signals such as temperature, pressure and current in multiple zones and introduces spatial topology matrix to structurally model the distribution relationship of thermocouples, so that the model can capture the slowly evolving early thermal imbalance features. On this basis, the cross-modal feature extraction and interactive enhancement of local temperature trend deviation features and pressure and current signals are further performed to identify the abnormal mode caused by multi-factor coupling. It can give an abnormal probability score and warning sign before the temperature exceeds the threshold range, so that the warning signal can be triggered before the temperature deviates substantially, effectively making up for the response window caused by thermal inertia delay and reducing the risk of overheating and deformation of the entire furnace workpiece.

[0020] (2) In order to distinguish between normal heat conduction and local anomalies in a multi-zone heating scenario, this invention first models the bidirectional time-series dependency as a window hidden state, so that the asymmetric laws of heating and cooling are represented separately, and establishes a time-series measurement benchmark for subsequent judgment of whether the current fluctuation of a certain temperature zone is a local anomaly or a cross-zone conduction lag; then, a graph attention mechanism with spatial topology matrix constraints is introduced, so that each temperature zone actively refers to the synchronous temperature behavior of its physical neighbors - when the temperature change of a certain temperature zone can be reasonably explained by the heat conduction of its neighbors, its abnormal response is naturally attenuated; otherwise, the local offset component remains significant, eliminating false alarms caused by normal heat conduction from the spatial dimension; finally, deviation analysis is performed based on the causal historical trend, and the current change is judged under the spatial neighborhood constraint to determine whether it deviates from its own historical evolution trajectory, so that the system maintains a sensitive response to actual faults in the anomaly monitoring of the multi-zone heating system.

[0021] (3) In order to solve the problems of temperature branch being susceptible to noise interference during operating condition switching and loss of initial details in multi-layer pooling of cooperative branch, this invention achieves bidirectional information guidance through cross-modal multi-head attention mechanism. The local trend features of temperature branch are used as queries to retrieve global features of cooperative branch, so that temperature branch can obtain global cooperative vision while maintaining fine-grained details. When pressure current coupling changes cause normal temperature rise, the corresponding response mode can be found in cooperative features to help temperature branch identify such fluctuations as normal rather than abnormal drift. When isolated temperature drift occurs, the cooperative features do not respond, locking it as a real anomaly. At the same time, the local details of temperature branch are retrieved by using global features of cooperative branch as queries, so that cooperative branch can recapture the instantaneous abnormal signal at the sampling start stage and improve the sensitivity of early small anomalies. After bidirectional guidance, the hidden states of the two branches are added bit by bit and fused to obtain a time-series anomaly discrimination feature with dual vision. (4) The present invention further introduces an adaptive gating mechanism based on working conditions: dynamically adjusts the contribution weights of temperature evidence and collaborative evidence according to the current data characteristics, and retains the direct connection path to provide a fallback response under fuzzy working conditions. At the same time, it solves the problems of blind spots in single-channel information and changes in the primary and secondary relationships of multiple channels with working conditions, so that the ratio of anomaly discrimination information sources is automatically adjusted, and the robustness under multiple working conditions is improved. Attached Figure Description

[0022] Figure 1 A flowchart illustrating an LSTM-based early warning method for temperature anomalies in a high-temperature sintering furnace, provided by this invention. Figure 2 The diagram shows the convergence process of the binary cross-entropy loss function provided by this invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0024] Example 1: A method for early warning of temperature anomalies in high-temperature sintering furnaces based on LSTM, such as... Figure 1 As shown, it includes the following steps: A1: Collect multi-zone temperature, pressure, and current sequences and preprocess them separately to obtain the cleaned multi-zone temperature, pressure, and current sequences, including: A11: Multi-zone temperature sequences are synchronously acquired by thermocouple arrays arranged in each temperature zone of the high-temperature sintering furnace. The data type is multi-dimensional time-series numerical data, including temperature sampling values ​​of each temperature zone. Outliers are removed from the multi-zone temperature sequences using the three-standard-deviation method, and then missing values ​​generated after removal are filled in using the linear interpolation method to obtain the cleaned multi-zone temperature sequences. A12: The pressure sequence is acquired through the furnace pressure transmitter, and the current sequence is acquired through the heating circuit current transformer. Both data types are one-dimensional time-series numerical data. The pressure sequence includes furnace pressure sampling values, and the current sequence includes heating circuit current sampling values. Outliers are removed from the pressure sequence and the current sequence using the three-standard-deviation method, and then the missing values ​​generated after removal are filled in using the linear interpolation method to obtain the cleaned pressure sequence and the cleaned current sequence.

[0025] A2: The time axis of the cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence are aligned, windowed, and arranged to obtain the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence. A spatial topology matrix is ​​then constructed based on the spatial layout of the thermocouples within the furnace, including: A21: The cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence are aligned along the time axis. Then, according to a fixed window duration and sliding step size, segments are extracted sequentially along the time axis. The extracted equal-length time series subsequences are then arranged in the original sampling time order to obtain the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence. Each of the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence contains multiple single-window time series subsequences. A22: Based on the spatial layout of the thermocouples in the furnace, a spatial topology matrix is ​​constructed. The construction method includes: using each thermocouple as a matrix node, constructing the spatial topology matrix according to the arrangement order of the nodes; wherein, for two nodes with a direct spatial adjacency relationship, the corresponding matrix elements are set to connected values; for two nodes without a direct spatial adjacency relationship, the corresponding matrix elements are set to disconnected values, and the diagonal elements of the matrix corresponding to the node itself are set to connected values.

[0026] A3: Based on the preprocessed multi-zone temperature sequence and spatial topology matrix, extract the enhanced multi-zone temperature window hidden state sequence, spatially enhanced temperature temporal hidden state, and local temperature trend deviation features sequentially, including: A31: Based on the preprocessed multi-zone temperature sequence, it is encoded using a time-series feature extraction path to obtain the enhanced multi-zone temperature window hidden state sequence. The calculation method is as follows: ; ; in, The window is initially hidden. This is for bidirectional long short-term memory network operations. For layer normalization operation, For multilayer sensor operation, This is a single-window time series subsequence from the preprocessed multi-zone temperature sequence. For the enhanced multi-zone temperature window hidden state sequence, This is a random deactivation operation. Let Gaussian error be the activation function of the linear unit. It is a fully connected layer; A32: Based on the spatial topology matrix and the enhanced multi-region temperature window hidden state sequence, feature co-modeling is performed using a graph attention mechanism to obtain the spatially enhanced temperature temporal hidden state. The calculation method is as follows: ; ; in, Here is the spatial attention weight matrix. This is for normalizing the exponential function operation. To correct the leakage of the linear unit activation function, These are the linear projection operations for the query vector and the key vector, respectively. For transpose operation, For Hadama accumulation, It is a spatial topological matrix. This refers to spatially enhanced temperature-temporal hidden states. It is the activation function of the exponential linear unit. These are linear projection operations on the value vector and the residual vector, respectively; A33: Based on the spatially enhanced temporal hidden state of temperature, multi-scale trend differencing is used to compare and decode the current time with the historical baseline, obtaining the local temperature trend deviation characteristics. The calculation method is as follows: ; ; in, For multi-scale gating weights, for Type activation function, The core size is respectively One-dimensional convolution operation, This is a global average pooling operation. This indicates a local deviation in temperature trends. For the core size is One-dimensional causal convolution operation.

[0027] Specifically, for scenarios requiring refined identification of thermal coupling imbalance anomalies within a specific temperature range, this invention also provides a spatiotemporal joint feature encoding calculation method based on two-dimensional convolution to replace step A31. The calculation method is as follows: ; ; in, For tensor dimension reshaping operations, This is a two-dimensional convolution operation. This is a flattening operation for feature tensors.

[0028] A4: Based on the local temperature trend deviation characteristics, the preprocessed pressure sequence, and the preprocessed current sequence, extract the temperature-pressure cross-mode features and temperature-current cross-mode features, including: A41: Based on the local temperature trend deviation features and the preprocessed pressure sequence, a gated time-delay convolutional alignment mechanism is constructed to extract features from the preprocessed pressure sequence. These features are then asynchronously correlated and encoded with the local temperature trend deviation features to obtain the temperature-pressure cross-modal features. The calculation method is as follows: ; ; ; in, This is a hidden state encoded for pressure timing. Let Gaussian error be the activation function of the linear unit. The expansion rates are respectively The one-dimensional dilated convolution operation, This is a single-window time series subsequence in the preprocessed pressure sequence. This represents the global temperature deviation characteristic after pooling. For temperature Pressure cross-modal characteristics; A42: Based on the preprocessed current sequence and local temperature trend deviation characteristics, a channel-normalized time-domain convolution backbone is used to capture abrupt change features of the current sequence, and a current-temperature dual-modal response mechanism is constructed to obtain temperature-current cross-modal features. The calculation method is as follows: ; ; in, This is a hidden state for current timing encoding. For layer normalization operation, For one-dimensional convolution operations with channel normalization, This is a single-window time-series subsequence in the preprocessed current sequence. Let Gaussian error be the activation function of the linear unit. For temporal convolutional network operations, For temperature Current cross-mode characteristics, It is the hyperbolic tangent activation function.

[0029] A5: Based on the temperature-pressure cross-modal characteristics and the temperature-current cross-modal characteristics, interactive enhancement is performed to obtain multivariate collaborative anomaly characteristics. The calculation method is as follows: ; ; in, For cross-modal competitive gating coefficients, For feature splicing operations, This is a multivariate collaborative anomaly characteristic.

[0030] A6: Based on the local temperature trend deviation characteristics and multivariate collaborative anomaly characteristics, extract the time-series anomaly discrimination features and the final time-series anomaly discrimination features in sequence, including: A61: Based on the local temperature trend deviation characteristics and multivariate collaborative anomaly characteristics, the two branches are mutually guided by feature mapping to obtain the hidden states of the cross-enhanced temperature branch and the cross-enhanced collaborative branch. Then, the time-series anomaly discrimination features are calculated. The calculation method is as follows: ; ; ; in, For cross-enhanced temperature branch hidden state, For the hidden state of the cooperative branch with cross-enhancement, For multi-head attention operation, Features for identifying time-series anomalies; A62: Based on the time-series anomaly detection features, the information from the two branches is adaptively filtered and integrated to obtain the final time-series anomaly detection features. The calculation method is as follows: ; ; ; in, For temperature branch gating weights, To coordinate the gating weights of branch paths, This is the final feature for identifying time-series anomalies.

[0031] A7: Based on the final time-series anomaly discrimination features, calculate the anomaly probability score and early warning trigger flag, including: A71: Based on the final time-series anomaly discrimination features, perform deep anomaly mapping and probabilization to obtain anomaly probability scores. The calculation method is as follows: ; ; in, The gating enhancement feature at time t, It is a multilayer perceptron. The time series anomaly discrimination features at time t are... Let be the anomaly probability score at time t. The weight matrix is ​​projected onto the score. The projection bias vector for the score; A72: Based on the anomaly probability score, a dynamic quantile threshold is calculated and anomaly determination is performed to obtain the warning trigger flag. The calculation method is as follows: ; ; in, For dynamic anomaly thresholds, The quantile calculation function, This is a sequence of anomaly probability scores for the entire time period. This represents the total number of time points. The preset percentile parameter for abnormal sensitivity. The warning trigger flag at time t, This is an indicator function.

[0032] In this embodiment of the invention, the parameters of the neural network module are configured as follows: In step A31: The single-window temporal subsequence of the preprocessed multi-zone temperature sequence is input into a multilayer perceptron with an input feature dimension of 4 and a hidden layer output feature dimension of 128. The activation function is a Gaussian error linear unit. The output of the multilayer perceptron is normalized and then input into a bidirectional long short-term memory network with 64 hidden units (in each direction) and an output feature dimension of 128. Subsequently, this feature is mapped to 128 dimensions through a fully connected layer, activated by Gaussian error linear units with a random deactivation ratio of 0.3, and then added to the input residual and normalized to obtain the enhanced multi-zone temperature window hidden state sequence with a dimension of 128. In step A32: the enhanced multi-zone temperature window hidden state sequence (dimension 128) is linearly projected by query, key, and value respectively, with the projection output feature dimension being 64 (per attention head), the number of attention heads being 2, and the total output dimension being 128; the residual linear projection input dimension is 128, and the output dimension is 128; the attention weights are masked by the spatial topology matrix, normalized by the normalized exponential function, and activated by the leaky linear rectifier unit (negative slope 0.2); after the attention output is added to the residual branch, it is activated by the exponential linear unit and layer normalized to obtain the spatially enhanced temperature temporal hidden state, with a dimension of 128; In step A33: The spatially enhanced temporal hidden state of temperature (128 dimensions) is calculated as a gate weight through two parallel branches: The first branch uses two one-dimensional convolutions with kernel sizes of 3 and 5, and 128 input and output channels, respectively, and is activated by Gaussian error linear units in between; The second branch performs global average pooling on the input (output 128 dimensions) and then maps it to 128 dimensions through a fully connected layer; The two branches are added together and then passed through a sigmoid activation function to obtain a 128-dimensional gate vector; In the main branch, the input is passed through a fully connected layer (128→128) and a one-dimensional causal convolution (kernel size 7, input and output channels 128), the two are subtracted and then multiplied element-wise with the gate vector, and then added to the output of another fully connected layer (128→128), and finally normalized to obtain the local temperature trend deviation feature, with a dimension of 128; In step A41: The preprocessed pressure sequence (single channel, window length 256) is sequentially passed through two one-dimensional dilated convolutions. The first convolution has a dilation rate of 2, a kernel size of 3, 1 input channel, and 8 output channels; the second convolution has a dilation rate of 4, a kernel size of 3, 8 input channels, and 64 output channels. It is then passed through a bidirectional long short-term memory network with 32 hidden units (per direction), an output dimension of 64, and a random deactivation rate of 0.3. This output is then passed through two more one-dimensional dilated convolutions (the first with a dilation rate of 2, a kernel size of 3, 1 input channel, and 8 output channels). The outputs of the first layer (8 channels; second layer with dilation rate 2, kernel size 3, input channels 8, output channels 64) are summed, normalized, and then activated by Gaussian error linear units to obtain the pressure temporal encoding hidden state with dimension 64; the temperature local trend deviation feature is obtained by global average pooling to obtain a 128-dimensional vector; this vector is added to the fully connected mapping of the pressure hidden state (64→64) through a fully connected layer (128→64), and after sigmoid activation, it is multiplied element-wise with the pressure hidden state. The residuals are summed and then normalized to obtain the temperature-pressure cross-modal feature with dimension 64; In step A42: The preprocessed current sequence (single channel, window length 256) is sequentially passed through two one-dimensional convolutions. The first convolution has a kernel size of 3, 1 input channel, and 8 output channels. The second convolution is normalized with channel normalization, has a kernel size of 3, 8 input channels, and 64 output channels. Simultaneously, it is passed through a temporal convolutional network, which consists of 3 dilated convolutional blocks, each with a kernel size of 3 and dilation rates of 1, 2, and 4, with 64 input and output channels. The two outputs are added together and then activated by exponential linear units and normalized to obtain the current temporal encoded hidden state, with a dimension of 64. The temperature local trend deviation feature is globally averaged and pooled (128 dimensions) and then passed through a fully connected layer (128→64). It is added to the fully connected mapping of the current hidden state (64→64), activated by hyperbolic tangent, and then fully mapped to 64 dimensions again. It is then multiplied element-wise with the current hidden state, and the residuals are added and normalized to obtain the temperature-current cross-modal feature, with a dimension of 64. In step A5: the temperature-pressure cross-modal features (64-dimensional) and the temperature-current cross-modal features (64-dimensional) are concatenated to obtain 128-dimensional features. After passing through a fully connected layer (128→128), Gaussian error linear unit, another fully connected layer (128→128), and S-type activation, 128-dimensional cross-modal competitive gating coefficients are obtained. The two features are then passed through fully connected layers (64→128) respectively, and then fused according to the gating coefficients. This result is then added to the result of another concatenated fully connected layer (128→128) and Gaussian error linear unit. Finally, the layers are normalized to obtain multivariate collaborative anomaly features with 128 dimensions. In step A61: the local temperature trend deviation feature (128 dimensions) and the multivariate collaborative anomaly feature (128 dimensions) are used as the query, key, and value inputs, respectively, in a multi-head attention mechanism. There are 8 attention heads, each with an output dimension of 16, for a total output of 128 dimensions. The attention output is added to the output of each of the fully connected layers (128→128), then normalized, and activated by Gaussian error linear units to obtain two cross-enhanced hidden states (each 128 dimensions). The two are concatenated to 256 dimensions, then passed through a fully connected layer (256→128), random deactivation (ratio 0.3), Gaussian error linear units, and a fully connected layer (128→128), and finally normalized to obtain the time-series anomaly discrimination feature with a dimension of 128. In step A62: the time-series anomaly discrimination feature (128 dimensions) is concatenated with the temperature local trend deviation feature (128 dimensions) and the multivariate collaborative anomaly feature (128 dimensions) respectively (each 256 dimensions). After passing through a fully connected layer (256→128), a Gaussian error linear unit, and a fully connected layer (128→128), it undergoes S-shaped activation to obtain two 128-dimensional gated weights. After passing through a fully connected layer (128→128), the time-series anomaly discrimination feature is multiplied element-wise with the two gated weights and added. Then, it is added to the output of another fully connected layer (128→128). The layers are normalized to obtain the final time-series anomaly discrimination feature with 128 dimensions. In terms of training, this invention adopts an end-to-end training mode, and the training process is as follows: Figure 2 As shown, the training configuration is as follows: Optimizer: AdamW; Initial learning rate: 0.001; Learning rate decay strategy: cosine annealing, minimum learning rate set to 0.00001; Batch size: 32; Training rounds: 100; Loss function: Binary cross-entropy loss.

[0033] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method 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, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0034] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0035] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for early warning of temperature anomalies in high-temperature sintering furnaces based on LSTM, characterized in that, Includes the following steps: A1: Collect multi-zone temperature sequence, pressure sequence, and current sequence and preprocess them respectively to obtain the cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence. A2: Time axis alignment, windowing, and arrangement of the cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence are performed to obtain the pre-processed multi-zone temperature sequence, pre-processed pressure sequence, and pre-processed current sequence, and a spatial topology matrix is ​​constructed based on the spatial layout of the thermocouples in the furnace. A3: Based on the preprocessed multi-zone temperature sequence and spatial topology matrix, extract the enhanced multi-zone temperature window hidden state sequence, spatially enhanced temperature temporal hidden state, and temperature local trend deviation features in sequence. A4: Based on the local temperature trend deviation characteristics, the preprocessed pressure sequence, and the preprocessed current sequence, extract the temperature-pressure cross-mode features and the temperature-current cross-mode features, respectively. A5: Based on the temperature-pressure cross-modal characteristics and the temperature-current cross-modal characteristics, interactive enhancement is performed to obtain multivariate collaborative anomaly characteristics; A6: Based on the local temperature trend deviation characteristics and multivariate collaborative anomaly characteristics, extract the time-series anomaly discrimination features and the final time-series anomaly discrimination features in sequence; A7: Calculate the anomaly probability score and early warning trigger flag based on the final time-series anomaly discrimination characteristics.

2. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 1, characterized in that, Step A1 includes: A11: Multi-zone temperature sequences are synchronously acquired by thermocouple arrays arranged in each temperature zone of the high-temperature sintering furnace. The data type is multi-dimensional time-series numerical data, including temperature sampling values ​​of each temperature zone. Outliers are removed from the multi-zone temperature sequences using the three-standard-deviation method, and then missing values ​​generated after removal are filled in using the linear interpolation method to obtain the cleaned multi-zone temperature sequences. A12: The pressure sequence is acquired through the furnace pressure transmitter, and the current sequence is acquired through the heating circuit current transformer. Both data types are one-dimensional time-series numerical data. The pressure sequence includes furnace pressure sampling values, and the current sequence includes heating circuit current sampling values. Outliers are removed from the pressure sequence and the current sequence using the three-standard-deviation method, and then the missing values ​​generated after removal are filled in using the linear interpolation method to obtain the cleaned pressure sequence and the cleaned current sequence.

3. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 2, characterized in that, Step A2 includes: A21: The cleaned multi-zone temperature sequence, cleaned pressure sequence, and cleaned current sequence are aligned along the time axis. Then, according to a fixed window duration and sliding step size, segments are extracted sequentially along the time axis. The extracted equal-length time series subsequences are then arranged in the original sampling time order to obtain the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence. Each of the preprocessed multi-zone temperature sequence, preprocessed pressure sequence, and preprocessed current sequence contains multiple single-window time series subsequences. A22: Based on the spatial layout of the thermocouples in the furnace, a spatial topology matrix is ​​constructed. The construction method includes: using each thermocouple as a matrix node, constructing the spatial topology matrix according to the arrangement order of the nodes; wherein, for two nodes with a direct spatial adjacency relationship, the corresponding matrix elements are set to connected values; for two nodes without a direct spatial adjacency relationship, the corresponding matrix elements are set to disconnected values, and the diagonal elements of the matrix corresponding to the node itself are set to connected values.

4. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 3, characterized in that, Step A3 includes: A31: Based on the preprocessed multi-zone temperature sequence, it is encoded using a time-series feature extraction path to obtain the enhanced multi-zone temperature window hidden state sequence. The calculation method is as follows: ； ； in, The window is initially hidden. This is for bidirectional long short-term memory network operations. For layer normalization operation, For multilayer sensor operation, This is a single-window time series subsequence from the preprocessed multi-zone temperature sequence. For the enhanced multi-zone temperature window hidden state sequence, This is a random deactivation operation. Let Gaussian error be the activation function of the linear unit. It is a fully connected layer; A32: Based on the spatial topology matrix and the enhanced multi-region temperature window hidden state sequence, feature co-modeling is performed using a graph attention mechanism to obtain the spatially enhanced temperature temporal hidden state. The calculation method is as follows: ； ； in, Here is the spatial attention weight matrix. This is for normalizing the exponential function operation. To correct the leakage of the linear unit activation function, These are the linear projection operations for the query vector and the key vector, respectively. For transpose operation, For Hadama accumulation, It is a spatial topological matrix. This refers to spatially enhanced temperature-temporal hidden states. It is the activation function of the exponential linear unit. These are linear projection operations on the value vector and the residual vector, respectively; A33: Based on the spatially enhanced temporal hidden state of temperature, multi-scale trend differencing is used to compare and decode the current time with the historical baseline, obtaining the local temperature trend deviation characteristics. The calculation method is as follows: ； ； in, For multi-scale gating weights, for Type activation function, The core size is respectively One-dimensional convolution operation, This is a global average pooling operation. This indicates a local deviation in temperature trends. For the core size is One-dimensional causal convolution operation.

5. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 4, characterized in that, The A4 step includes: A41: Based on the local temperature trend deviation features and the preprocessed pressure sequence, a gated time-delay convolutional alignment mechanism is constructed to extract features from the preprocessed pressure sequence. These features are then asynchronously correlated and encoded with the local temperature trend deviation features to obtain the temperature-pressure cross-modal features. The calculation method is as follows: ； ； ； in, This is a hidden state encoded for pressure timing. Let Gaussian error be the activation function of the linear unit. The expansion rates are respectively The one-dimensional dilated convolution operation, This is a single-window time series subsequence in the preprocessed pressure sequence. This represents the global temperature deviation characteristic after pooling. For temperature Pressure cross-modal characteristics; A42: Based on the preprocessed current sequence and local temperature trend deviation characteristics, a channel-normalized time-domain convolution backbone is used to capture abrupt change features of the current sequence, and a current-temperature dual-modal response mechanism is constructed to obtain temperature-current cross-modal features. The calculation method is as follows: ； ； in, This is a hidden state for current timing encoding. For layer normalization operation, For one-dimensional convolution operations with channel normalization, This is a single-window time-series subsequence in the preprocessed current sequence. Let Gaussian error be the activation function of the linear unit. For temporal convolutional network operations, For temperature Current cross-mode characteristics, It is the hyperbolic tangent activation function.

6. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 5, characterized in that, Step A5 includes: First, the temperature-pressure cross-mode features and the temperature-current cross-mode features are concatenated. Then, the features are processed sequentially through a linear layer, a GeLU activation function, a linear layer, and a Sigmoid function to obtain the cross-modal competitive gating coefficients. The temperature-pressure cross-modal features are processed through a linear layer and then multiplied by the Hadamard product with the cross-modal competitive gating coefficients. The temperature-current cross-modal features are processed through a linear layer and then multiplied by the Hadamard product with the result of 1 minus the cross-modal competitive gating coefficients. The temperature-pressure cross-modal features and the temperature-current cross-modal features are concatenated and processed by a linear layer and the GeLU activation function. The results of the above three steps are added together and then processed by layer normalization to obtain the multivariate collaborative anomaly features.

7. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 6, characterized in that, Step A6 includes: First, the local temperature trend deviation features are processed by a linear layer. At the same time, the local temperature trend deviation features are used as query vectors, and the multivariate collaborative anomaly features are used as key vectors and value vectors for multi-head attention operation. The results of the above two items are added together and then processed by layer normalization and GeLU function in sequence to obtain the hidden state of the temperature branch with cross-enhancement. The multivariate collaborative anomaly features are then processed through a linear layer. Simultaneously, a multi-head attention operation is performed on the multivariate collaborative anomaly features as query vectors and the temperature local trend deviation features as key and value vectors. The results of the above two items are added together and then processed by layer normalization and GeLU function in sequence to obtain the cross-enhanced collaborative branch hidden state. The hidden states of the temperature branch enhanced by cross-enhancement are concatenated with the hidden states of the cooperative branch enhanced by cross-enhancement. The concatenated result is processed by a linear layer to obtain the first linear intermediate result, and by a linear layer, GeLU function, and Dropout operation to obtain the second linear intermediate result. The first linear intermediate result and the second linear intermediate result are then added together and processed by layer normalization to obtain the time-series anomaly discrimination features.

8. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 6, characterized in that, Step A6 includes: First, the time-series anomaly discrimination features and the local temperature trend deviation features are concatenated, and then processed sequentially through a linear layer, a GeLU function, a linear layer, and a Sigmoid function to obtain the temperature branch gating weights. The time-series anomaly discrimination features and multivariate collaborative anomaly features are then concatenated and processed sequentially through a linear layer, a GeLU function, another linear layer, and a Sigmoid function to obtain the collaborative branch gating weights. Then, the time-series anomaly discrimination features are processed through three linear layers to obtain the first linear result, the second linear result, and the third linear result. The temperature branch gating weights are multiplied by the first linear result, and the cooperative branch gating weights are multiplied by the second linear result. The two Hadamard product results are then added to the third linear result. Finally, the results are normalized to obtain the final time-series anomaly discrimination features.

9. The LSTM-based high-temperature sintering furnace temperature anomaly early warning method according to claim 7, characterized in that, Step A7 includes: First, the temporal anomaly discrimination features at time t are processed sequentially by layer normalization, GeLU function activation, and multilayer perceptron to obtain the gated enhancement features at time t. Then, the gated enhancement features are multiplied by the score projection weight matrix, the score projection bias vector is added, and then processed by the Sigmoid function to obtain the anomaly probability score at time t. Based on the anomaly probability score sequence at all times, a quantile calculation function is used to determine the dynamic anomaly threshold with a preset anomaly sensitivity percentile parameter. Then, for each time moment, an indicator function is used to determine whether the anomaly probability score at that time is greater than the dynamic anomaly threshold. If it is, the warning trigger flag at that time is set to true; otherwise, it is set to false.

Images