Bell jar furnace fault alarm data processing method and system based on improved transformer

Abstract

This invention relates to the field of data processing technology, specifically to a method and system for processing fault alarm data of a bell-shaped furnace based on an improved Transformer. The method includes: acquiring the original alarm data, interlock action data, and process stage data of the bell-shaped furnace; slicing and reconstructing the original alarm data according to the interlock intervention time before and after each alarm occurrence to generate interlock intervention alarm segments corresponding to each alarm process; and inputting the interlock intervention alarm segments into an improved Transformer model to learn and analyze the alarm change relationship before and after interlock intervention within the segment. This invention, through the interlock intervention alarm segment + interlock position bias Transformer modeling framework, reconstructs alarms, interlocks, and process stages into event-level data on the same time axis. Furthermore, it introduces interlock position bias into self-attention to enhance the perception of semantic boundaries before and after interlocks, achieving unified modeling for recovering original fault semantics from mixed alarms and alarm purification.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for processing fault alarm data of bell-shaped furnaces based on an improved Transformer. Background Technology

[0002] Bell-shaped furnaces are widely used in fields such as magnetic material sintering, metal heat treatment, and powder processing. Their operation involves high temperature, multiple atmosphere switching, and long-cycle control. The system typically includes a temperature control unit, an atmosphere control unit, a cooling unit, and a multi-level interlocking protection mechanism. In actual operation, a distributed control system is relied upon to monitor abnormal states, and fault indication and safety protection are achieved through alarm systems and interlocking systems.

[0003] In existing technologies, alarm processing is mostly based on preset thresholds or rule logic. When abnormal temperature, atmospheric fluctuations, or equipment failure occur, an alarm is triggered and linked to perform interlocking actions such as power outage, gas switching, or valve closure. In practical applications, alarm data often exhibits chain propagation characteristics, that is, the initial fault will trigger multiple subsequent alarms, and these alarms are similar in form to the original fault, making them difficult to distinguish.

[0004] Interlocking actions alter the system's operating state, leading to a large number of derived alarms triggered by interlocks. This can cover up or overwhelm the original fault alarms, creating alarm masking and increasing the difficulty of fault location. The meaning of alarms differs across different process stages, but current technologies lack effective correlation between process stages and alarm data, easily leading to misjudgments. Existing alarm systems lack the ability to model the temporal relationships between alarms, making it difficult to distinguish between leading and derived alarms. In scenarios with multiple devices or multiple interlocks, alarm redundancy and unreasonable responses are likely to occur. Existing machine learning-based methods are mostly limited to single-point identification and struggle to handle the semantic changes in alarms caused by interlocking interventions.

[0005] Therefore, how to model multi-source alarm data under interlocking intervention conditions, accurately distinguish between real faults and derived alarms, and achieve the purification and structured expression of alarm information has become an urgent technical problem to be solved. Summary of the Invention

[0006] This invention provides a method and system for processing fault alarm data of bell-shaped furnaces based on an improved Transformer.

[0007] A method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer includes the following steps: The original alarm data, interlock action data, and process stage data of the bell furnace are obtained. The original alarm data are sliced ​​and recombined according to the interlock intervention time before and after each alarm to generate interlock intervention alarm segments corresponding to each alarm process. The interlocking intervention alarm segment is input into the improved Transformer model to learn and analyze the alarm change relationship before and after interlocking intervention within the segment. The leading alarm features triggered by real faults and the derived alarm features induced by interlocking actions are separated, and the corresponding original fault characterization results and alarm purification results are output. Based on the original fault characterization results and the alarm purification results, the target alarm results and interlocking handling results for the bell-shaped furnace are generated.

[0008] As a further technical solution of the present invention, original alarm data, interlock action data and process stage data are collected in real time or obtained offline from the distributed control system or data history database of the bell furnace. The original alarm data includes the alarm occurrence time, alarm tag number, alarm type and alarm status. The interlock action data includes the interlock trigger time, interlock source tag number and interlock target tag number. The process stage data includes the different process stages of the bell furnace operation and their start and end times. Based on the interlocking trigger time in the interlocking action data, determine the time point of each interlocking intervention, and take the time point of interlocking intervention as the center, trace back to the time of the first alarm occurrence associated with it, and extend backward to the time when the interlocking action is completed and the alarm status tends to stabilize, and determine the slice window of each alarm process. For any interlocking record , based on its interlocking trigger time As the central time point, its corresponding slice window is defined as: ;in: ; ; in, This indicates the start time of the slice window, corresponding to the time when the first alarm associated with this interlock action occurs before the interlock is triggered. This indicates the end time of the slice window, corresponding to the moment when the interlocking action is completed and the alarm status tends to stabilize. This indicates the intensity of the alarm state change near time t, used to characterize whether the alarm is still in a state of drastic change. This represents the stability threshold.

[0009] As a further technical solution of the present invention, the original alarm data is extracted based on the slice window, and all alarm records located in the same slice window are recombined into an independent interlocking intervention alarm segment. The interlocking intervention alarm segment is associated with the corresponding process stage data and interlocking action data to form a complete data unit for characterizing the entire process of an alarm event. For any slice window The corresponding subset of alarm data is defined as follows: Furthermore, this subset of alarm data is integrated with the corresponding interlocking action data and process stage data to construct an interlocking intervention alarm segment. , obtain: ;in: ;in, This represents the subset of alarm data within the j-th slice window. This indicates the corresponding interlocking action record. This represents the set of process stages that intersect with the time range of this slice window. This indicates the start time of the k-th process stage. This indicates the end time of the k-th process stage.

[0010] As a further technical solution of the present invention, the interlocking intervention alarm segment is time-aligned and feature-encoded to construct an input tensor containing an alarm sequence, an interlocking intervention marker, and a process stage label, wherein the alarm sequence is arranged in chronological order and the interlocking intervention marker is used to identify the exact location where the interlocking action occurs; An improved Transformer model is constructed, which includes an embedding layer, a multi-head self-attention layer, a feedforward neural network, and a feature separation output layer. The multi-head self-attention layer introduces position bias encoding at the time of interlocking intervention to enhance the ability to focus on the relationship between alarm changes before and after interlocking actions. In the multi-head self-attention layer, a position bias encoding at the moment of interlock intervention is introduced to enhance the model's sensitivity to alarm changes before and after interlock actions. Specifically, an interlock position bias term is superimposed on the standard position encoding: ;in, Indicates standard position code, Indicates interlock intervention marker, This indicates the interlock position offset weight.

[0011] As a further technical solution of the present invention, the input tensor is input to the embedding layer for feature mapping to obtain a vectorized representation of the alarm event, which is then input to the multi-head self-attention layer. The self-attention mechanism is used to model the dependency relationship and timing pattern between alarms at different times in the alarm sequence, and the influence of the interlocking intervention mark on the alarm sequence is learned. The output of the multi-head self-attention layer is then sent to the feedforward neural network for nonlinear transformation to extract deep alarm timing features.

[0012] As a further technical solution of the present invention, the output of the feedforward neural network is fed into the feature separation output layer, and two output branches are optimized simultaneously through a multi-task learning mechanism: the first branch outputs the original fault representation result. The first branch is used to identify the characteristics of a leading alarm triggered by a real fault; the second branch outputs the alarm mitigation results. Used to identify derived alarm features triggered by interlocking actions; in, This represents the original fault characterization results. Indicates the alarm and purification results. Indicates the output layer weights. The output layer bias is indicated by the original fault characterization results, which are used to identify the leading alarm features triggered by real faults; the alarm purification results are used to identify the derived alarm features induced by interlocking actions.

[0013] As a further technical solution of the present invention, the original fault characterization results and the alarm purification results are analyzed to establish the correspondence between each pilot alarm feature in the original fault characterization results and the original alarm record in the interlocking intervention alarm segment. At the same time, the derived alarm features marked in the alarm purification results are associated and mapped with the interlocking actions that trigger them to obtain the real fault alarm set and the derived alarm set.

[0014] As a further technical solution of the present invention, the target alarm result is generated based on the real fault alarm set, specifically including: extracting the leading alarm features with confidence levels higher than a preset threshold from the original fault characterization results, outputting their corresponding alarm tag numbers, alarm times and alarm types as real fault alarms, and aggregating and merging multiple alarms triggered by the same fault source to form a simplified target alarm list.

[0015] As a further technical solution of the present invention, the interlocking handling result is generated based on the derived alarm set, specifically including: identifying repeated alarms, reverse alarms, and masking alarms induced by interlocking actions in the derived alarm set, and generating corresponding suppression strategies for different types of derived alarms. Specifically, for repeated alarms, an alarm suppression instruction is generated to mask redundant alarms within the same interlocking action cycle; for reverse alarms, an alarm reversal instruction is generated to correct false alarms that have the opposite effect to the expected effect of the interlocking action; and for masking alarms, an alarm priority enhancement instruction is generated to re-expose the real fault alarms that were masked by the interlocking action.

[0016] A fault alarm data processing system for bell-shaped furnaces based on an improved Transformer includes the following modules: The data acquisition and fragment construction module is used to acquire the original alarm data, interlock action data and process stage data of the bell furnace, and to slice and reconstruct the original alarm data according to the interlock intervention time before and after each alarm to generate interlock intervention alarm fragments corresponding to each alarm process. The feature parsing module is used to input the interlocking intervention alarm segment into the improved Transformer model, learn and analyze the alarm change relationship before and after the interlocking intervention within the segment, separate the leading alarm features triggered by the real fault and the derived alarm features induced by the interlocking action, and output the corresponding original fault characterization results and alarm purification results. The result generation module is used to generate target alarm results and interlocking handling results for the bell-shaped furnace based on the original fault characterization results and the alarm purification results. The target alarm results are used to output the actual faults, and the interlocking handling results are used to suppress repeated alarms, reverse alarms, and shielding alarms induced by interlocking actions.

[0017] The beneficial effects of this invention are: This invention reconstructs alarms, interlocks, and process stages into event-level data on the same timeline using an interlocking intervention alarm segment + interlocking position bias Transformer modeling framework. It also introduces interlocking position bias into self-attention to enhance the perception of semantic boundaries before and after interlocking, thereby achieving unified modeling for recovering original fault semantics from mixed alarms and alarm cleanup.

[0018] This invention constructs a multi-task feature separation mechanism that combines native fault identification and alarm purification. This mechanism decouples the same alarm sequence into two semantic categories: real faults and interlocking derivations. By combining fault source mapping and energy-saving merging strategies, it achieves alarm redundancy removal, reverse correction, and masking recovery, significantly improving alarm accuracy and availability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the system modules in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram illustrating the alarm accuracy and false alarm rate of Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the derived alarm suppression rate and false alarm rate in Embodiment 3 of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] Example 1 like Figure 1 As shown, a method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer includes the following steps: S1. Obtain the original alarm data, interlock action data, and process stage data of the bell furnace. According to the interlock intervention time before and after each alarm, slice and reassemble the original alarm data to generate interlock intervention alarm segments corresponding to each alarm process.

[0023] S11, real-time or offline acquisition of raw alarm data, interlock action data, and process stage data from the distributed control system or historical data database of the bell-shaped furnace; among which, The original alarm data is represented as follows: ; Any alarm record Represented as: ; Where N represents the total number of alarm records. Indicates the first When the alarm occurs Indicates the alarm tag number, device or measuring point identifier. Indicates the alarm type, such as abnormal temperature or abnormal atmosphere. Indicates the alarm status, whether it is triggered, deactivated, or maintained; The interlocking action data is represented as follows: ; Any interlocking record Represented as: ; Where M represents the total number of interlocking action records. Indicates the first The triggering time of the interlocking action, This indicates the interlock source bit number, representing the signal source that triggered the interlock. It indicates the interlock target tag number, which equipment is affected by the interlock, such as valve closure, power outage, etc. The process stage data is represented as follows: ; Record of any process stage Represented as: ; The total number of K process stages. Indicates the first Each process stage is marked with indicators such as heating, heat preservation, cooling, evacuation, and inflation. Indicates the start time of the phase. Indicates the end time of the phase.

[0024] The structured modeling described above provides a unified timeline for representing alarm data, interlocking action data, and process stage data, thus providing data support for subsequent slice processing.

[0025] S12: Based on the interlocking trigger time in the interlocking action data, determine the time point of each interlocking intervention, and construct a slice window of the alarm process around this time point. Specifically, using the interlocking trigger time as the central time point, trace back to the earliest alarm occurrence time associated with the interlocking action to determine the starting boundary of the slice window. Simultaneously, extend backward to the moment when the interlocking action is completed and the alarm state no longer changes significantly to determine the ending boundary of the slice window. This forms a time window covering the entire process of alarm triggering—interlocking intervention—interlocking impact—state stabilization, used to characterize a complete alarm event process. For any interlocking record... , based on its interlocking trigger time As the central time point, its corresponding slice window is defined as: ; in: Before the interlock triggering time, from all the alarm records that have occurred, the alarms that are related to the interlocking action are selected, and the earliest alarm is selected as the starting point of the slice window. This starting point is not simply the alarm that is most recent before the interlocking is triggered, but is traced back to the initial alarm source that could trigger the interlocking action, so as to ensure that the determined time window can cover the initial stage of the fault and reflect the complete pre-process of the actual fault from its inception to the triggering of the interlock. Starting from the moment of interlock triggering, observe the changes in system state and find the first time point where the intensity of alarm state change is lower than the preset stability threshold. Use this time point as the end point of the slice window. The intensity of alarm state change is used to describe whether there are still frequent alarm triggering, deactivation, or switching behaviors around this time. When the intensity of change decreases to below the threshold, it indicates that the impact of the interlocking action has basically ended, that is, it has entered a relatively stable state. Therefore, the end time actually marks the node where the interlocking action ends and the alarm evolution process tends to stabilize, thus ensuring that the constructed time window not only includes the interlocking action itself, but also covers its subsequent impact stage. in, This indicates the start time of the slice window, corresponding to the time when the first alarm associated with this interlock action occurs before the interlock is triggered. This indicates the end time of the slice window, corresponding to the moment when the interlocking action is completed and the alarm status tends to stabilize. This indicates the intensity of the alarm state change near time t, used to characterize whether the alarm is still in a state of drastic change. This represents the stability threshold. The stability threshold is set to a low level so that the number of alarm triggers, cancellations, or state transitions per unit time is significantly reduced. When the alarm enters a stable state, it is generally determined that the number of alarm changes does not exceed the preset minimum number within a certain number of consecutive time intervals, such as close to zero change or extremely low frequency change, in order to avoid prematurely truncating the alarm process that is still under the influence of interlocking. When the intensity of alarm state change is lower than this threshold, the system enters the stable stage. The stability threshold is determined based on three main factors: First, based on the alarm fluctuation characteristics of the bell furnace under normal operating conditions, its natural disturbance range is statistically analyzed, and the threshold is set to be slightly higher than the background fluctuation level, so as to avoid misjudging normal fluctuations as instability. Second, based on the alarm change pattern after the interlocking action is triggered, the alarm usually changes drastically in the early stage of interlocking and then gradually decays. The threshold should be selected at the end of the decay stage to ensure that the slice window can cover the entire process of the interlocking effect. Third, considering the impact of different process stages on alarm fluctuations, the threshold should be appropriately increased during stages with large fluctuations such as temperature rise or atmosphere switching, and appropriately decreased during stable stages such as heat preservation, thereby improving the accuracy and adaptability of stability determination.

[0026] Each interlocking intervention corresponds to a complete time window, ensuring that the entire process from the first alarm trigger to the end of the interlocking action and stabilization is fully covered, thus avoiding semantic loss due to only capturing partial alarms.

[0027] S13: Based on the slice window, the original alarm data is extracted and reassembled to generate an interlocking intervention alarm segment.

[0028] For any slice window The corresponding subset of alarm data is defined as follows: From all the original alarm data, all alarm records whose occurrence time falls within the current slice window range are selected and these records are combined into a set. Specifically, all alarm information within a certain time period is completely extracted to form the alarm data subset corresponding to that time window. The alarm data subset set reflects all the abnormal behaviors of the system during the fault process and is the basic data source for subsequent analysis. Furthermore, this subset of alarm data is fused with the corresponding interlocking action data and process stage data to construct an interlocking intervention alarm segment. , obtain: The alarm data subset within the current slice window, the corresponding interlock action data, and related process stage data are uniformly combined to construct a complete data structure. This data structure describes the entire process of an alarm event, including not only the alarm itself but also the system's interlock response and the process background at the time. Essentially, it integrates multi-source data into an event-level data unit, enabling subsequent processing to be based on overall semantic analysis rather than judgment based on fragmented data. From all process stage data, select process stages whose time ranges intersect with the current slice window, and form a set of process stages with intersection. This set is used to describe the operating stage of the system when the alarm occurs, providing operating context for understanding the actual meaning of the alarm, thereby avoiding incorrect judgments on the same alarm under different process stages. in, This represents the subset of alarm data within the j-th slice window. This indicates the corresponding interlocking action record. This represents the set of process stages that intersect with the time range of this slice window. This indicates the start time of the k-th process stage. This indicates the end time of the k-th process stage; By extracting alarm data within a time window and uniformly associating it with interlocking actions and process stage information, an interlocking intervention alarm segment that can completely characterize the entire process of a fault is constructed. Through the above-mentioned reorganization method, each interlocking intervention alarm segment is made possible. All of them can fully characterize the entire process of an alarm event, from triggering and interlocking intervention to stable termination.

[0029] S2, input the interlocking intervention alarm segment into the improved Transformer model, learn and analyze the alarm change relationship before and after interlocking intervention within the segment, separate the leading alarm features triggered by real faults from the derived alarm features induced by interlocking actions, and output the corresponding original fault characterization results and alarm purification results.

[0030] S21, for each interlocking intervention alarm segment Temporal alignment and feature encoding are performed to construct a unified input tensor.

[0031] First, a subset of alarm data Arranged in ascending order of time, the alarm sequence is represented as follows: Any alarm event is represented as follows: For each alarm event in an alarm segment, its tag number, alarm type, alarm status, and corresponding process stage label are input into an encoding function. After mapping, a vector representation is generated. This vector not only contains the attribute information of the alarm itself but also incorporates its contextual operating conditions. This allows subsequent models to compare, correlate, and model different alarm events in a unified feature space. The feature vector representation of the alarm event is constructed as follows: ;in, This represents the set of alarms after the j-th alarm sequence is sorted by time. This represents the number of alarm events in the j-th segment, t-th segment. The label for the process stage corresponding to the alarm time. This indicates the time when the t-th alarm occurs in the j-th segment. This represents the bit number corresponding to the t-th alarm. Indicates the type of the t-th alarm. This indicates the status of the t-th alarm; The interlocking intervention marker sequence is represented as follows: ;in, Finally, the alarm sequence features, interlock intervention markers, and process stage labels are combined to form the input tensor representation as follows: ;in, Indicates the first The number of alarm events in each interlocking intervention alarm segment Indicates the first Vectorized representation of an alarm event Represents the feature encoding function. The interlocking flag at time t; Through the above processing, alarm events are aligned and displayed on the timeline, and the location of the interlocking event is marked.

[0032] An improved Transformer model is constructed, which includes an embedding layer, a multi-head self-attention layer, a feedforward neural network, and a feature separation output layer. The multi-head self-attention layer consists of multiple parallel attention substructures, each of which is a head. Each substructure independently performs attention calculations on the input sequence, learning temporal dependencies from different perspectives. The outputs of multiple attention heads are concatenated or fused, and then a linear transformation is performed to obtain the final output. The model can simultaneously capture the associated features of alarm sequences from multiple feature subspaces, improving its ability to express complex temporal relationships. In the multi-head self-attention layer, a position bias encoding at the time of interlocking intervention is introduced to enhance the model's sensitivity to alarm changes before and after interlocking actions. Specifically, an interlocking position bias term is superimposed on the standard position encoding. ;in, Indicates standard position code, Indicates interlock intervention marker, Indicates the interlock position offset weight; When modeling alarm sequences based on Transformer, the model should be able to perceive and focus on the critical time point of interlocking, so as to more accurately distinguish the relationship between alarm changes before and after the interlocking action. In the actual operation of the bell furnace, the interlocking action is often the turning point in the fault evolution process: before the interlocking occurs, the alarm reflects more the formation process of the actual fault, while after the interlocking occurs, the alarm may contain a large number of derivative changes caused by protection actions. If the model cannot identify the key dividing point, it is easy to misjudge the derivative alarms after the interlocking as new faults, thus affecting the accuracy of the overall judgment.

[0033] In the standard Transformer architecture, the multi-head self-attention layer establishes temporal dependencies by calculating the correlation between various moments in the sequence. However, its default premise is that the importance of each time position is balanced, that is, the model does not actively distinguish between critical moments and ordinary moments. This invention introduces position bias encoding of the interlocking intervention moment into the multi-head self-attention layer. Its function is to explicitly strengthen the location of the interlocking on the basis of the original time position information, so that the location has a higher influence in the attention calculation, thereby guiding the model to focus on the alarm changes before and after the interlocking.

[0034] Specifically, each alarm event, when input into the model, corresponds to a location code to represent its positional relationship in the time series. This invention does not change the original location code method, but introduces an interlocking intervention flag on it. When the corresponding interlocking occurs at a certain moment, the flag value at that moment is activated, thereby superimposing an additional bias term into the location code. This bias term changes the representation of that time point in the feature space, making it easier for other time points to pay attention to it in the subsequent attention weight calculation process. At the same time, that time point itself also has stronger selectivity when paying attention to other time points.

[0035] During the learning process, the Transformer model naturally develops a preference for data changes in the time of interlocking and its immediate vicinity. This allows it to more clearly learn how alarms before interlocking evolve and how alarms after interlocking are affected by system intervention. Improvements enable the Transformer model to distinguish between native fault alarms and interlocking-derived alarms, thus providing a foundation for subsequent alarm cleanup and fault identification.

[0036] By introducing interlocking intervention markers into the positional information of the input sequence, the encoded value at the corresponding position shifts when an interlock occurs at a certain moment. This increases the weight influence of that position during the attention weight calculation process, making the model pay more attention to alarm changes before and after the interlock occurs when establishing temporal dependencies.

[0037] While keeping the original position encoding unchanged, an offset component is added to the corresponding time position according to the interlocking intervention mark, so that the encoding representation of the interlocking occurrence position is enhanced or shifted relative to other positions, thereby highlighting the importance of the position in the feature representation space and realizing the enhanced expression of key moments.

[0038] Standard position coding is used to represent the temporal order information of each element in a sequence. By assigning different codes to different positions, the model can perceive the relative positional relationship of each alarm event on the time axis, thereby supporting the modeling of time-dependent structures.

[0039] This design allows the model to pay more attention to the temporal dependencies near the location where the interlocking occurs when calculating attention weights.

[0040] S22, input tensor Input the embedding layer to obtain the initial embedding representation: After the input tensor enters the embedding layer, it is converted into a set of vector representations. The original data is in the form of text labels, such as alarm type, tag number, etc., while the embedding layer transforms the original data into numerical vector form so that the model can perform calculations. Each alarm event at each time point becomes an information vector with multi-dimensional features, and these vectors are arranged in chronological order. The data is then fed into a multi-head self-attention layer to model the dependencies in the alarm sequence. This multi-head self-attention layer models the sequential relationships within the entire alarm sequence, enabling the model to understand how alarms at different time points influence each other. The information vectors of the multi-dimensional features are input into the multi-head self-attention layer. For each alarm at a given time point, the model examines all other alarms in the sequence to determine their relationships. For example, a later alarm might be closely related to a previous alarm, or an alarm might be the result of an interlocking action. The calculation process for any attention head is as follows: Where Q represents the query matrix, K represents the key matrix, and V represents the value matrix. This represents the query mapping weight matrix. This represents the key-map weight matrix. The value mapping weight matrix represents the matrix. Indicates the first lHidden feature representation in the Transformer layer; the model generates three different representations for each time point: one for actively paying attention to others, one for being referenced by others, and one for providing information content. By comparing the degree of matching between different time points, an attention weight is calculated. The larger the weight, the stronger the correlation between the two time points. The attention calculation function assigns an attention weight to each time point based on the similarity between different time points. Then, it weights and sums the information from all time points. Specifically, it first calculates the relevance score between the current time point and all other time points, then normalizes these scores to form a weight distribution. Finally, it uses these weights to weight and sum the information from all time points, resulting in a representation that incorporates global information. An interlocking position bias is superimposed in this calculation process, giving interlocking-related time points a more significant position in the weight distribution. The attention weight calculation is as follows: ;wherein the bias term: ;in, Representing feature dimension, Indicates the interlock position offset. This represents the bias coefficient. This represents the attention calculation function. This represents the interlocking flag at time s in the j-th sequence. This represents the position bias matrix, which is used to introduce interlocking intervention information in attention calculation and to weight the relationship between different time positions.

[0041] The purpose of adding interlocking-related bias terms is to artificially increase the weight of a point in time that corresponds to the location of an interlock when the model calculates the relationship between different time points. This makes the model pay more attention to alarms before and after the interlock occurs, rather than treating all time points equally. During the learning process, the model will automatically develop a preference for alarm changes before and after interlock, making it easier to distinguish which alarms are the real cause of the fault and which are the result of the interlock action.

[0042] By introducing This allows the model to focus more on the alarm relationship before and after an interlock occurs, thereby learning the impact of interlocks on the alarm sequence.

[0043] Output the multi-head self-attention layer: Then The input is processed by a feedforward neural network using a nonlinear transformation. The role of the feedforward neural network is to focus on reprocessing the features at the current point in time, rather than looking at the relationship between time points, so as to reveal the hidden complex patterns. This allows the model to extract more abstract features from the original linear relationship, such as whether a certain alarm combination has typical fault characteristics, or whether it conforms to the change pattern after interlocking intervention, thus obtaining deeper features. ;in, Represents the weight matrix. Indicates the bias term. The activation function performs a linear transformation on the input features, followed by nonlinear processing to bend and reorganize the features, thus expressing more complex relationships. Subsequently, a second linear transformation integrates the nonlinear features to form a new feature representation. The activation function uses a Gaussian error linear unit, which can adaptively adjust the output response according to the distribution of the input value. It performs flexible suppression when the input is small and maintains a high response when the input is large, making it more suitable for handling situations where weak and strong signals are mixed in alarm data.

[0044] By introducing a feedforward neural network, the model can not only understand the correlation between alarms, but also further determine whether the correlation has practical significance, thus providing a more discriminative feature basis for distinguishing between original fault features and interlocking derived features.

[0045] S23, after processing by the multi-head self-attention layer and feedforward neural network, the resulting deep features contain rich information. This information not only reflects the temporal sequence of alarms but also includes the dependencies between alarms and the impact of interlocking interventions. These features can be seen as a comprehensive expression of the entire alarm process. However, this information is still mixed together, containing both the characteristics of the actual fault and the changing characteristics caused by interlocking interventions. If used directly for output, it will lead to unclear judgment. This invention introduces a feature separation output layer, whose function is to interpret the same deep features from different perspectives and output results with different semantics. The model no longer only provides... Instead of issuing a single unified judgment, two interpretations are given simultaneously, thus decoupling the alarm semantics. The deep features output from the feedforward neural network are used as shared inputs, connected to two independent output sub-layers. Each sub-layer generates its corresponding output through its own linear mapping and normalization processing. One sub-layer outputs the original fault representation, focusing on alarm features triggered by real faults; the other sub-layer outputs the alarm cleanup, focusing on derived alarm features caused by interlocking actions. The two output branches share the front-end feature extraction structure, but use independent parameters for mapping in the output stage, thus achieving separate feature representation. The input feature separation output layer generates two output branches through a multi-task learning mechanism: First branch native fault characteristics: The purpose of the original fault characterization results is to identify alarm features directly triggered by real faults. The original fault characterization results focus more on the earliest abnormal patterns that appear in the alarm sequence and have already formed before interlocking. Therefore, the purpose of the original fault characterization results is to help the system find the root cause of the fault. Second branch alarm purification results: The purpose of alarm cleanup is to identify derived alarms generated by interlocking actions or system interventions. Alarm cleanup focuses more on alarm changes after an interlock occurs, especially alarms that are highly correlated with interlocking actions, thereby enabling the labeling or filtering of non-root cause alarms. in, This represents the original fault characterization results. Indicates the alarm and purification results. Indicates the output layer weights. The output layer bias is indicated by the original fault characterization results, which are used to identify the leading alarm features triggered by real faults; the alarm purification results are used to identify the derived alarm features induced by interlocking actions.

[0046] The multi-task learning mechanism in this invention adopts a structure that combines a shared backbone network with multi-task branches. Specifically, based on the same set of input data, it simultaneously completes multiple tasks that are intrinsically related but semantically different, thereby achieving multi-angle analysis of alarm data. The entire structure can be divided into three layers: a shared feature extraction layer, a task branch layer, and a joint optimization layer. Each layer is connected and works collaboratively. In the shared feature extraction layer, the input interlocking intervention alarm segment is converted into a vector representation by the embedding layer, and then processed by the multi-head self-attention layer and the feedforward neural network to obtain a unified deep feature representation. The deep feature integrates the time dependency of the alarm sequence, the impact of interlocking intervention, and process stage information, and is a comprehensive expression of the entire alarm process. In this stage, the model does not distinguish between specific tasks, but focuses on extracting general features that are useful for all tasks, thereby avoiding redundant modeling and improving feature utilization efficiency.

[0047] In the task branching layer, the shared features mentioned above are simultaneously input into multiple parallel task branches. Each branch corresponds to a specific task and has an independent parameter structure. Specifically, the first task branch is used to generate the original fault characterization result, which focuses on identifying the leading features directly caused by the real fault in the alarm sequence and pays special attention to the alarm evolution pattern before the interlock occurs. The second task branch is used to generate the alarm purification result, which focuses on identifying the derived alarms caused by the interlock action or system intervention and pays special attention to the state change features after the interlock occurs. Each branch usually consists of an independent feature mapping layer and an output layer. By transforming the shared features from different angles, the separation and expression of different semantic information can be achieved.

[0048] In the joint optimization layer, a multi-task joint training mechanism is constructed to enable each task branch to be optimized simultaneously during training. Each task branch is defined with a corresponding optimization objective, and multiple objectives are optimized collaboratively during the overall training process. This allows the shared backbone network to learn feature representations applicable to multiple tasks. Furthermore, since each task branch has independent parameters, it can be differentiated based on shared features, thereby maintaining inter-task collaboration while avoiding mutual interference. Through the structural design of sharing + separation + collaborative optimization, the model's ability to express complex alarm data can be effectively improved, enabling it to accurately identify real faults and effectively filter derived alarms caused by interlocking.

[0049] Unlike conventional Transformers that model sequences solely based on temporal order, this invention introduces a structured data unit—the interlocking intervention alarm segment—in the input stage. This unit encodes the alarm sequence, interlocking actions, and process stage information in a unified manner, enabling the model to simultaneously perceive the fault evolution process and system intervention behavior on the same timeline. By explicitly constructing an interlocking intervention marker and introducing this marker during the temporal modeling process, the Transformer model no longer treats all time points equally but can instead identify the key semantic boundary point of the interlocking trigger moment, thereby gaining the ability to distinguish semantic changes before and after the alarm. Furthermore, the introduction of interlocking position bias encoding into the multi-head self-attention mechanism constitutes a significant improvement over the standard Transformer. Traditional attention mechanisms rely solely on feature similarity to calculate weights, making it difficult to highlight the importance of critical moments. In contrast, this invention superimposes interlocking position bias during the attention calculation process, giving the location of the interlocking and its neighborhood higher weight in the attention distribution. This guides the model to focus on the alarm relationship before and after the interlocking, enabling the improved Transformer model to proactively learn which alarms are the cause of the interlocking and which alarms are the result of the interlocking, effectively solving the problem of alarm semantic confusion in industrial scenarios.

[0050] This invention also introduces a deep feature extraction mechanism for interlocking perception, which integrates the contextual dependencies in the alarm sequence with the impact of interlocking interventions. Through the combination of multi-head attention and feedforward networks, the model can not only capture long-term temporal dependencies, but also extract high-order features related to the fault evolution path. This makes the representation of each alarm event contain three types of information: its own attributes, historical dependencies, and interlocking impacts, thereby significantly improving the discriminative ability of feature representation.

[0051] This invention designs a feature separation structure based on multi-task learning in the Transformer output stage to decouple alarm semantics. On the basis of extracting unified features from a shared backbone network, two parallel output paths are set up: a native fault representation branch and an alarm purification branch. This enables the model to learn two tasks simultaneously: fault root cause identification and interlocking derived alarm identification. Compared with the traditional single output structure, this improvement enables the model to separate mixed alarm signals into different semantic categories, thereby fundamentally solving the problem of mixed alarm information under interlocking intervention.

[0052] In summary, this invention systematically improves the traditional Transformer for industrial fault alarm scenarios by introducing interlocking semantic tagging, position bias enhancement mechanism, and multi-task feature separation structure. This transforms the Transformer from a general sequence modeling model into a dedicated time-series parsing model with interlocking semantic understanding capabilities, resulting in higher accuracy and interpretability when processing alarm data under complex interlocking intervention conditions.

[0053] S3. Based on the original fault characterization results and the alarm purification results, generate the target alarm results and interlock handling results for the bell-shaped furnace; the target alarm results are used to output the actual fault, and the interlock handling results are used to suppress repeated alarms, reverse alarms, and shielding alarms induced by interlock actions.

[0054] S31. After the model has output the results, the output results are interpreted in a practical way. The abstract judgments output by the model, such as the confidence level, are mapped back to the specific alarm records to realize the actual classification and reconstruction of the alarm data. The model has output two results for each alarm event: one is the confidence level of the alarm belonging to the original fault, and the other is the confidence level of the alarm belonging to the interlocking derivation. However, the results themselves are only numerical values ​​and cannot be directly used for the output or control of the alarm system. Therefore, it is necessary to map the numerical results to the original alarm data one by one to realize the semantic classification of each alarm. For the Individual interlocking intervention alarm segment The alarms are classified and mapped based on the model output results.

[0055] First, establish the correspondence between the original fault characterization results and the original alarm records, assuming: ; The original fault scoring sequence is then defined as follows: The model has learned the temporal correlation between alarms in the multi-head self-attention layer, such as which alarms occur first, which alarms follow later, and the change features before and after the interlocking occurs. Furthermore, the feedforward neural network further extracts deep semantic features, so that the alarm representation at each time point not only contains its own information, but also integrates other information. Subsequently, in the feature separation output layer, for the original fault identification task, the model will generate a score value for each alarm based on the deep features, which is used to represent the confidence level of the alarm as a real fault trigger signal. Based on this, a real fault alarm set is constructed as follows: For the original fault characterization results, it is necessary to establish a correspondence between each alarm record. For each alarm record in the alarm segment, a corresponding original fault confidence score can be found, which indicates how likely the alarm is to be directly caused by a real fault. All alarms can then be filtered, and only alarms with higher confidence scores can be retained as real fault alarms. In summary, the transformation from model output to actual alarm filtering can be achieved. Simultaneously, a derived alarm set is constructed based on the alarm purification results, assuming: ;but: For alarm cleanup results, the model will give a derived alarm confidence score for each alarm, indicating whether the alarm was caused by an interlocking action. By filtering the confidence scores, a set of derived alarms can be obtained, which do not belong to the real fault. These derived alarms are associated with their corresponding interlocking actions. Since derived alarms are not randomly generated but triggered by specific interlocking actions, such as power outages, valve closures, and purging, each derived alarm needs to be attributed to a specific interlocking action to clarify its source. This mapping relationship allows for further analysis of which alarms are consequential results within a particular interlocking action cycle, providing a basis for subsequent alarm suppression, correction, or recovery. The association mapping between derived alarms and interlocking actions can be determined through time consistency and tag number association. For each alarm identified as a derived alarm... The record first determines whether the occurrence time falls within the effective time window after the corresponding interlocking action is triggered; secondly, it determines whether the alarm tag number belongs to the target tag number of the interlocking action or its associated device set; when both time association and device association conditions are met, the derived alarm can be mapped to the corresponding interlocking action, thereby establishing a one-to-one or one-to-many association relationship between derived alarms and interlocking actions. Through this mapping relationship, the source of each type of derived alarm can be clearly identified, providing a basis for subsequent classification and handling. Furthermore, the derived alarms and interlocking actions are associated and mapped, and the mapping relationship is defined as follows: ;in, This represents the actual set of fault alarms. Indicates a derived alarm set, This represents the sequence of original fault characterization results corresponding to the j-th interlocking intervention alarm segment. This represents the confidence level that the t-th alarm is a native fault. This represents the alarm purification result sequence corresponding to the j-th interlocking intervention alarm segment. θ represents the confidence level that the t-th alarm is a derived alarm, and θ represents the threshold for determining the original fault. This indicates the threshold for determining derived alarms. This indicates the mapping relationship between derived alarms and interlocking actions.

[0056] Through the above processing, the semantic classification of alarms is achieved, and the mixed alarms are separated into real fault alarms and interlocking derived alarms.

[0057] S32 After the set of real fault alarms has been screened out, these alarms are further organized to make the output results clearer, more concise and more engineering-usable. This is because in the actual operation of the bell furnace, a real fault often does not only trigger one alarm, but also triggers a series of related alarms. If all these alarms are output directly, it will not only be redundant, but also increase the judgment burden of the operators. Therefore, it is necessary to integrate alarms that belong to the same fault source.

[0058] First, key information for each alarm is extracted from the set of real fault alarms, including alarm tag number, occurrence time, and alarm type. This information serves as the basis for subsequent processing, aiming to transform the raw alarm records into a structured candidate alarm set for unified processing. Further, real fault alarms are filtered and extracted to construct a target alarm candidate set. ;in, Indicates the alarm tag number. Indicates the time when the alarm occurred. Indicate the alarm type; Subsequently, multiple alarms triggered by the same fault source are aggregated and merged. The fault source is the root cause or core device triggering this series of alarms. For example, an malfunction in a temperature control device may simultaneously cause a temperature alarm, a power anomaly alarm, and an interlock trigger alarm. Although these alarms are different in form, they all originate from the same fault source. Therefore, it is necessary to use certain mapping rules to assign each alarm to its corresponding fault source. The fault source mapping function is defined as follows: Based on the alarm tag number and the relationship between the equipment topology, the functional association, and the interlocking trigger path, a mapping rule is established to assign each alarm to its corresponding equipment or functional unit as the fault source. When the alarm tag number belongs to the same equipment, the same control loop, or is in the same interlocking link, it can be determined that it originates from the same fault source, thereby realizing the attribution mapping from alarm to fault source. After mapping is completed, alarms are grouped according to fault source, placing all alarms belonging to the same fault source into the same set. Each set then corresponds to one actual fault event, and the alarms within each set represent different manifestations of that fault. Alarms are grouped based on the fault source mapping: ; Alarms within each group are aggregated and merged to extract the most representative results from multiple related alarms. This can be done by retaining the earliest occurring alarm as the starting signal, selecting the alarm with the highest confidence level as the representative, or merging multiple similar alarms. This generates a concise alarm output. After processing, the alarm set, which might have contained multiple redundant pieces of information, is compressed into one or a few representative target alarms, significantly improving the readability and usability of the alarms. The resulting target alarm list is represented as follows: ;in, This represents the candidate set of real fault alarms. This represents a fault source mapping function, used to map alarms to corresponding fault sources. This indicates the fault source identifier corresponding to the alarm. Indicates the target alarm result. This represents an alarm merging function, used to combine multiple alarms from the same fault source. This represents the subset of alarms belonging to the k-th fault source.

[0059] S33. Although the model has separated real fault alarms and derived alarms, the derived alarms themselves are not of the same nature. Some are just redundant alarms that repeatedly occur within the same interlocking cycle, while others are alarms whose direction is opposite to the semantics of the real fault due to the change of equipment state caused by interlocking actions. Some may even cover up the truly important alarms. If not further subdivided, a large amount of invalid information will still be output, and it will still be difficult for the on-duty personnel to quickly grasp the key faults. Therefore, it is necessary to perform semantic classification on the derived alarms again, and then take differentiated handling methods such as suppression, correction or recovery.

[0060] First, analyze the relationship between each alarm and interlocking action in the derived alarm set, focusing on three main aspects: First, does the alarm occur immediately after the interlocking action? Second, whether the alarm tag number or the device to which it belongs is within the range of interlocking action; Third, whether the alarm's presentation format matches the expected state change after interlocking; The above can be used to determine whether a derived alarm is a redundant alarm that occurs repeatedly, a reverse alarm that has the opposite effect of the interlocking, or a masking alarm that suppresses the original critical alarm due to the intervention of the interlocking action.

[0061] The classification of derived alarms is based on the temporal relationship, tag number association, and state change direction between them and the interlocking action. Alarms that repeatedly occur within the same interlocking action cycle and have the same or highly similar alarm tag number and alarm type can be classified as duplicate alarms. Alarms that appear after the interlocking action is triggered, and whose alarm state corresponds to the expected effect of the interlocking action but are semantically easily misjudged as new faults, can be classified as reverse alarms. Alarms that cause a large number of derived alarms to appear in a concentrated manner after the interlocking action, resulting in a decrease in the display level of the actual fault alarm, a lower ranking, or difficulty in identification, can be classified as masking alarms. Further, derived alarms are categorized by type, and a classification function is defined: Based on the classification results, corresponding treatment strategies are generated. (1) Repetitive alarm suppression For duplicate alarm sets: ; Generate suppression instructions: ; Repeated alarms occur multiple times within the same interlocking action cycle due to repeated changes in status feedback, sampling updates, or signal jitter. These alarms do not provide new fault information and only increase the burden on operators and cause interference. Therefore, the focus of handling repeated alarms should not be on retention, but on compression and shielding. For records classified as repeated alarms, a suppression range is established according to the interlocking action cycle. Within this range, alarms with the same tag number, the same type, or high similarity are merged and shielded. Only the first representative alarm or the summary result is retained without repeatedly triggering displays, audible and visual prompts, and secondary interlocking responses, thus forming an alarm suppression command.

[0062] (2) Reverse alarm correction For the reverse alarm set: ; Generate inversion instructions: ; Reverse alarms refer to alarms that appear to be new anomalies but are actually the inevitable result of interlocking actions. Furthermore, this result is semantically opposite to the original fault direction. For example, after an interlock disconnects a certain output, an alarm indicating a parameter is too low may appear. This does not represent a new root cause fault but is merely the expected response after the interlock execution. Without correction, this reverse change might be mistaken for a new problem. Therefore, the direction of alarm interpretation needs to be reversed or corrected. For records classified as reverse alarms, the semantics of the alarm are redefined by combining the type of interlocking action that triggered the alarm and the expected state change after the interlock. The alarm is corrected from a newly added anomaly alarm to a state feedback after interlock execution or an accompanying change caused by the interlock. The display description, alarm interpretation direction, or handling label are adjusted accordingly, thus forming an alarm reversal instruction.

[0063] (3) Masking alarm recovery For masking alarm sets: ; Generate priority boosting instructions: ; Masked alarms refer to a situation where a large number of derived alarms appear simultaneously after interlocking actions are initiated, causing the most important real fault alarms to be buried, ranked lower, or not clearly displayed. Therefore, the focus of handling masked alarms should not be deletion, but rather adjusting priorities to highlight the truly critical alarms. For scenarios classified as masked alarms, the real fault alarms that are masked by derived alarms within the same interlocking cycle are identified. Based on their original fault confidence, occurrence order, and triggering effect on subsequent interlocking actions, the display priority, output order, or alarm level of the real fault alarm is increased, making it exposed again at the front of the alarm list or in a key prompt area, thereby forming a priority upgrade instruction.

[0064] The final interlocking response result is as follows: ;in, Represents the alarm classification function. Represents the set of repeated alarms. Represents the set of reverse alarms. Indicates a set of masking alarms. This represents the alarm suppression function. This indicates the alarm reversal function. This indicates a function with higher priority. Indicates the results of the interlocking response; For repeated alarms, an alarm suppression strategy is generated to prevent repeated output of the same type of alarm within the same interlocking cycle. For reverse alarms, an alarm reversal strategy is generated to reinterpret these alarms according to their actual semantics after interlocking, rather than outputting them directly according to their surface state. For masked alarms, a priority enhancement strategy is generated to elevate key alarms that are covered by a large amount of derived information to a higher display level or earlier output order. Ultimately, these three types of handling results together constitute the interlocking handling results, which are used to purify and reconstruct the alarm system output.

[0065] Example 2 like Figure 2 The system shown is a fault alarm data processing system for a bell-shaped furnace based on an improved Transformer, comprising the following modules: The data acquisition and fragment construction module is used to acquire the original alarm data, interlock action data and process stage data of the bell furnace, and to slice and reconstruct the original alarm data according to the interlock intervention time before and after each alarm to generate interlock intervention alarm fragments corresponding to each alarm process. The feature parsing module is used to input the interlocking intervention alarm segment into the improved Transformer model, learn and analyze the alarm change relationship before and after the interlocking intervention within the segment, separate the leading alarm features triggered by the real fault and the derived alarm features induced by the interlocking action, and output the corresponding original fault characterization results and alarm purification results. The result generation module is used to generate target alarm results and interlocking handling results for the bell-shaped furnace based on the original fault characterization results and the alarm purification results. The target alarm results are used to output the actual faults, and the interlocking handling results are used to suppress repeated alarms, reverse alarms, and shielding alarms induced by interlocking actions.

[0066] Example 3 like Figure 3-4 To verify the effectiveness of the method proposed in this invention, simulation verification was conducted using actual operating data from a bell-type furnace in a steel enterprise. Historical alarm data, interlock action data, and process stage data from three consecutive months were selected, comprising 12,678 original alarm records and 1,203 interlock action events. Using manually labeled real fault alarm tags as the evaluation benchmark, the following three methods were used for comparison: Existing technology 1: Traditional alarm processing methods based on fixed thresholds do not consider derived alarms induced by interlocking actions; Existing technology 2: An alarm filtering method based on time window rule matching relies on manually set alarm suppression rules; The method of this invention is a learning and feature separation method for interlocking intervention alarm segments based on an improved Transformer.

[0067] The simulation results are shown in the table below: This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0068] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer, characterized in that, include: The original alarm data, interlock action data, and process stage data of the bell furnace are obtained. The original alarm data are sliced ​​and recombined according to the interlock intervention time before and after each alarm to generate interlock intervention alarm segments corresponding to each alarm process. The interlocking intervention alarm segment is input into the improved Transformer model to learn and analyze the alarm change relationship before and after interlocking intervention within the segment. The leading alarm features triggered by real faults and the derived alarm features induced by interlocking actions are separated, and the corresponding original fault characterization results and alarm purification results are output. Based on the original fault characterization results and the alarm purification results, the target alarm results and interlocking handling results for the bell-shaped furnace are generated.

2. The method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer according to claim 1, characterized in that, The original alarm data, interlock action data, and process stage data are collected in real time or obtained offline from the distributed control system or data history database of the bell furnace. The original alarm data includes the alarm occurrence time, alarm tag number, alarm type, and alarm status. The interlock action data includes the interlock trigger time, interlock source tag number, and interlock target tag number. The process stage data includes the different process stages of the bell furnace operation and their start and end times. Based on the interlocking trigger time in the interlocking action data, determine the time point of each interlocking intervention, and take the time point of interlocking intervention as the center, trace back to the time of the first alarm occurrence associated with it, and extend backward to the time when the interlocking action is completed and the alarm status tends to stabilize, and determine the slice window of each alarm process. For any interlocking record , based on its interlocking trigger time As the central time point, its corresponding slice window is defined as: ;in: ； ； in, This indicates the start time of the slice window, corresponding to the time when the first alarm associated with this interlock action occurs before the interlock is triggered. This indicates the end time of the slice window, corresponding to the moment when the interlocking action is completed and the alarm status tends to stabilize. This indicates the intensity of the alarm state change near time t, used to characterize whether the alarm is still in a state of drastic change. This represents the stability threshold.

3. The method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer according to claim 2, characterized in that, Based on the slice window, the original alarm data is extracted, and all alarm records located in the same slice window are recombined into an independent interlocking intervention alarm segment. The interlocking intervention alarm segment is associated with the corresponding process stage data and interlocking action data to form a complete data unit for characterizing the entire process of an alarm event. For any slice window The corresponding subset of alarm data is defined as follows: Furthermore, this subset of alarm data is integrated with the corresponding interlocking action data and process stage data to construct an interlocking intervention alarm segment. , obtain: ;in: ;in, This represents the subset of alarm data within the j-th slice window. This indicates the corresponding interlocking action record. This represents the set of process stages that intersect with the time range of this slice window. This indicates the start time of the k-th process stage. This indicates the end time of the k-th process stage.

4. The method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer as described in claim 1, characterized in that, The interlocking intervention alarm segments are time-aligned and feature-encoded to construct an input tensor containing alarm sequences, interlocking intervention markers, and process stage labels; The alarm sequence is arranged in chronological order, and the interlocking intervention mark is used to identify the exact location where the interlocking action occurs; An improved Transformer model is constructed, which includes an embedding layer, a multi-head self-attention layer, a feedforward neural network, and a feature separation output layer. The multi-head self-attention layer introduces position bias encoding at the time of interlocking intervention to enhance the ability to focus on the relationship between alarm changes before and after interlocking actions. In the multi-head self-attention layer, a position bias encoding at the moment of interlock intervention is introduced to enhance the model's sensitivity to alarm changes before and after interlock actions. Specifically, an interlock position bias term is superimposed on the standard position encoding: ;in, Indicates standard position code, Indicates interlock intervention marker, This indicates the interlock position offset weight.

5. The method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer according to claim 4, characterized in that, The input tensor is input into the embedding layer and feature-mapped to obtain a vectorized representation of the alarm event; The multi-head self-attention layer is then input to model the dependency relationship and timing pattern between alarms at different times in the alarm sequence through the self-attention mechanism. Combined with the interlocking intervention mark, the influence law of interlocking action on the alarm sequence is learned. The output of the multi-head self-attention layer is sent to the feedforward neural network for nonlinear transformation to extract deep alarm timing features.

6. The method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer as described in claim 5, characterized in that, The output of the feedforward neural network is fed into the feature separation output layer, and two output branches are optimized simultaneously through a multi-task learning mechanism: the first branch outputs the original fault representation result. This is used to identify the characteristics of a leading alarm triggered by a real fault; The second branch outputs the alarm purification results: Used to identify derived alarm features triggered by interlocking actions; in, This represents the original fault characterization results. Indicates the alarm and purification results. Indicates the output layer weights. This indicates the output layer bias, and the original fault characterization results are used to identify the features of the leading alarms triggered by real faults. The alarm cleanup results are used to identify the characteristics of derived alarms triggered by interlocking actions.

7. The method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer according to claim 1, characterized in that, The original fault characterization results and the alarm purification results are analyzed to establish the correspondence between each pilot alarm feature in the original fault characterization results and the original alarm record in the interlocking intervention alarm segment; The derived alarm features marked in the alarm purification results are associated and mapped with the interlocking actions that trigger them to obtain the real fault alarm set and the derived alarm set.

8. A method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer, as described in claim 7, is characterized in that... The target alarm result is generated based on the actual fault alarm set; Specifically, this includes: extracting the leading alarm features with a confidence level higher than a preset threshold from the original fault characterization results, outputting their corresponding alarm tag number, alarm time, and alarm type as real fault alarms, and aggregating and merging multiple alarms triggered by the same fault source to form a simplified target alarm list.

9. A method for processing fault alarm data of a bell-shaped furnace based on an improved Transformer, as described in claim 8, characterized in that, The interlocking response result is generated based on the derived alarm set; Specifically, this includes: identifying repeated alarms, reverse alarms, and masking alarms induced by interlocking actions in the derived alarm set, and generating corresponding suppression strategies for different types of derived alarms; For repeated alarms, generate an alarm suppression command to mask redundant alarms within the same interlocking action cycle; For reverse alarms, generate alarm reversal commands to correct false alarms that are contrary to the expected effect of interlocking actions; For masked alarms, generate an alarm priority escalation command to re-expose the actual fault alarm that was masked by the interlocking action.

10. A fault alarm data processing system for a bell-shaped furnace based on an improved Transformer, used to implement the fault alarm data processing method for a bell-shaped furnace based on an improved Transformer as described in any one of claims 1-9, characterized in that, Includes the following modules: The data acquisition and fragment construction module is used to acquire the original alarm data, interlock action data and process stage data of the bell furnace, and to slice and reconstruct the original alarm data according to the interlock intervention time before and after each alarm to generate interlock intervention alarm fragments corresponding to each alarm process. The feature parsing module is used to input the interlocking intervention alarm segment into the improved Transformer model, learn and analyze the alarm change relationship before and after the interlocking intervention within the segment, separate the leading alarm features triggered by the real fault and the derived alarm features induced by the interlocking action, and output the corresponding original fault characterization results and alarm purification results. The result generation module is used to generate target alarm results and interlocking handling results for the bell-shaped furnace based on the original fault characterization results and the alarm purification results. The target alarm results are used to output the actual faults, and the interlocking handling results are used to suppress repeated alarms, reverse alarms, and shielding alarms induced by interlocking actions.

Images