Industrial production process abnormal condition early warning method and device and electronic equipment

Abstract

The application provides an industrial production process abnormal condition early warning method and device and electronic equipment, comprising: obtaining a process variable sequence of an industrial production process to be predicted; dividing the process variable sequence into an endogenous variable sequence and an exogenous variable sequence; constructing a first space projection corresponding to the endogenous variable sequence based on the endogenous variable sequence; constructing a second space projection corresponding to the exogenous variable sequence based on the exogenous variable sequence; updating the first space projection based on the first space projection and the second space projection to obtain an updated first space projection; obtaining a first space mapping result and a second space mapping result based on the updated first space projection; predicting the endogenous variable sequence at a future preset time based on the first space mapping result and the second space mapping result; and warning an abnormal condition of the industrial production process to be predicted based on the endogenous variable sequence at the future preset time and an alarm limit. The abnormal condition can be accurately warned.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial processes, and in particular to a method, device, and electronic equipment for early warning of abnormal operating conditions in industrial production processes. Background Technology

[0002] In the field of industrial production processes, early warning of abnormal operating conditions is of great significance. The core of such early warning systems lies in the accurate and timely triggering of alarms. Its judgment logic usually needs to comprehensively consider two types of key information: whether the absolute level of the measured parameter exceeds the preset safety threshold, and whether the dynamic behavior of the parameter over time meets the requirements.

[0003] According to relevant technologies, existing models use the same set of weights to model the two types of information simultaneously, causing them to interfere with each other during training: samples close to the normal range are dominated by horizontal features, while weak trend signals in the pre-limit warning segment are easily masked, making the model insensitive to the prediction of the alarm threshold.

[0004] Therefore, finding a method that can accurately provide early warning of abnormal operating conditions in industrial production processes has become a current research hotspot. Summary of the Invention

[0005] This invention provides a method, device, and electronic equipment for early warning of abnormal operating conditions in industrial production processes, which enables accurate early warning of abnormal operating conditions in industrial production processes.

[0006] This invention provides a method for early warning of abnormal operating conditions in industrial production processes. The method includes: acquiring a sequence of process variables for the industrial production process to be predicted; dividing the sequence of process variables into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified sequence of process variables; and the exogenous variable sequence is a sequence of process variables other than the endogenous variable sequence; constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for representing the endogenous variable sequence; and constructing a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence. A first spatial projection and a second spatial projection are used to update the first spatial projection, resulting in an updated first spatial projection. Based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in a preset dimension latent space; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the preset dimension latent space; based on the first spatial mapping result and the second spatial mapping result, the endogenous variable sequence at a preset future time is predicted; based on the endogenous variable sequence at the preset future time and the alarm limit, an early warning is issued for abnormal operating conditions of the industrial production process to be predicted.

[0007] According to the present invention, an early warning method for abnormal operating conditions in an industrial production process is provided. The step of constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence includes: dividing the endogenous variable sequence according to a preset time block length to obtain multiple tiles, wherein the tiles represent the endogenous variable subsequences obtained by dividing the endogenous variable sequence according to the preset time block length; for any tile, superimposing sine and cosine position codes onto the tile to obtain a superimposed tile, and projecting the superimposed tile onto a representation space of a preset dimension through an unbiased linear mapping to obtain a tile spatial projection corresponding to the tile; sorting the multiple tile spatial projections according to a time sequence to obtain a tile spatial projection sequence; and concatenating the global token at the end of the tile spatial projection sequence to construct the first spatial projection corresponding to the endogenous variable sequence.

[0008] According to the present invention, an abnormal operating condition early warning method for industrial production processes is provided, wherein the global token is obtained in the following manner: a pre-trained abnormal operating condition prediction model is invoked, wherein the abnormal operating condition prediction model includes a global token generation module, and the global token generation module includes a first unbiased linear mapping layer; the endogenous variable sequence is input into the abnormal operating condition prediction model, and linear mapping processing is performed through the first unbiased linear mapping layer in the global token generation module to obtain a global token for characterizing the endogenous variable sequence.

[0009] According to the present invention, an early warning method for abnormal operating conditions in an industrial production process is provided. The step of constructing a second spatial projection corresponding to the exogenous variable sequence includes: calling a pre-trained abnormal operating condition prediction model, wherein the abnormal operating condition prediction model includes an exogenous variable inversion embedding module, and the exogenous variable inversion embedding module includes a second unbiased linear mapping layer; inputting the exogenous variable sequence into the abnormal operating condition prediction model, and performing linear mapping processing through the second unbiased linear mapping layer in the exogenous variable inversion embedding module to obtain the second spatial projection corresponding to the exogenous variable sequence.

[0010] According to the present invention, an early warning method for abnormal operating conditions in an industrial production process is provided. After obtaining the process variable sequence of the industrial production process to be predicted, the method further includes: obtaining multiple sets of historical process variable sequences; determining the mean and standard deviation of each process variable based on the historical process variable sequences; for any one of the process variable sequences, normalizing the process variable sequence based on the mean and standard deviation to obtain a normalized process variable sequence; and using the normalized process variable sequence as the process variable sequence.

[0011] According to the present invention, an abnormal operating condition early warning method for industrial production processes is provided. The step of updating the first spatial projection based on a first spatial projection and a second spatial projection to obtain an updated first spatial projection includes: calling a pre-trained abnormal operating condition prediction model, wherein the abnormal operating condition prediction model includes a decoupled encoder module, wherein the decoupled encoder module is composed of multiple stacked encoders, wherein the initial updated first spatial projection obtained from the output of the current layer encoder is used as the first spatial projection input to the next layer encoder, until the last layer encoder outputs the initial updated first spatial projection, and the initial updated first spatial projection output by the last layer encoder is used as the updated first spatial projection; wherein each layer encoder performs the following process: performing multi-head self-attention processing on the first spatial projection to obtain a self-attention processed first spatial projection; and then... After concatenating the first spatial projection and the self-attention processed first spatial projection, layer normalization is performed to obtain the aggregated first spatial projection. Cross-attention processing is then performed using the global token in the aggregated first spatial projection as the query and the second spatial projection as the key and value to obtain the cross-attention processing result. The global token in the aggregated first spatial projection and the cross-attention processing result are then concatenated and layer normalization is performed to obtain the global token after cross-variable processing. The cross-variable processing global token is then backfilled into the aggregated first spatial projection to obtain the first spatial projection updated by the current layer encoder. A nonlinear transformation is performed on the first spatial projection updated by the current layer encoder to obtain the transformed first spatial projection. Finally, the first spatial projection updated by the current layer encoder and the transformed first spatial projection are concatenated and layer normalization is performed to obtain the initial updated first spatial projection.

[0012] According to the present invention, an abnormal operating condition early warning method for industrial production processes is provided. The step of obtaining a first spatial mapping result and a second spatial mapping result based on an updated first spatial projection includes: calling a pre-trained abnormal operating condition prediction model, wherein the abnormal operating condition prediction model includes a horizontal expert prediction module and a trend expert prediction module; wherein the horizontal expert prediction module includes a first multilayer perceptron; and the trend expert prediction module includes a second multilayer perceptron; inputting the updated first spatial projection into the abnormal operating condition prediction model, and performing [further analysis] on the updated first spatial projection through the horizontal expert prediction module. Spatial mapping processing is performed to obtain a first spatial mapping result; the updated first spatial projection is divided according to a preset time block length to obtain multiple projection tiles, wherein the projection tiles are used to represent the first spatial projection subsequence obtained by dividing the updated first spatial projection according to the preset time block length; a first-order difference processing is performed on adjacent projection tiles in the time dimension to obtain the inter-projection tile change tensor; the inter-projection tile change tensor is input into the working condition anomaly prediction model, and the trend expert prediction module performs spatial mapping processing on the inter-projection tile change tensor to obtain a second spatial mapping result.

[0013] According to the present invention, an abnormal operating condition early warning method for industrial production processes, before predicting the endogenous variable sequence at a future preset time based on the first spatial mapping result and the second spatial mapping result, the method further includes: calling a pre-trained abnormal operating condition prediction model, wherein the abnormal operating condition prediction model includes a gating module, the gating module including a third unbiased linear mapping layer, a fourth unbiased linear mapping layer, a GELU activation function, and a Sigmoid activation function; inputting the global token in the updated first spatial projection into the abnormal operating condition prediction model, and performing linear mapping processing through the third unbiased linear mapping layer in the gating module to obtain the linearly mapped global token; and applying the GELU activation function to... The global token after linear mapping is processed to obtain a first activated global token; the first activated global token is linearly mapped based on the fourth unbiased linear mapping layer to obtain a reprocessed global token; the reprocessed global token is processed based on the Sigmoid activation function to obtain a scalar gate; the prediction of the endogenous variable sequence at a future preset time based on the first spatial mapping result and the second spatial mapping result includes: subtracting 1 from the scalar gate to obtain a subtracted scalar gate; and performing a weighted summation on the first spatial mapping result, the scalar gate, the second spatial mapping result, and the subtracted scalar gate to predict the endogenous variable sequence at a future preset time.

[0014] According to the present invention, an abnormal operating condition early warning method for industrial production processes is provided. The abnormal operating condition prediction model is trained in the following manner: acquiring multiple sets of historical process variable sequences and corresponding real labels, wherein the real labels are used to characterize the historical endogenous variable sequences corresponding to the historical process variable sequences at a future preset time; constructing a training dataset based on the historical process variable sequences and the corresponding real labels; constructing a target training function, wherein the target training function is determined according to the relative relationship between the real labels and alarm limits in the training dataset; and training the abnormal operating condition prediction model in an end-to-end training manner based on the training dataset and the target training function to obtain a trained abnormal operating condition prediction model.

[0015] According to the present invention, an early warning method for abnormal operating conditions in an industrial production process is provided, wherein the target training function is determined in the following manner: an alarm mode item is determined, the alarm mode item includes a high alarm mode item, a low alarm mode item, and a dual-sided alarm mode item in which both high and low alarm mode items coexist; the target training function is constructed based on the alarm mode item, the true label, and the prediction result obtained by inputting the training data in the training dataset into the abnormal operating condition prediction model.

[0016] According to the present invention, an early warning method for abnormal operating conditions in an industrial production process is provided, wherein the alarm mode item is determined in the following manner: A high alarm mode item is determined based on the relative relationship between the true label and the high alarm limit, wherein the high alarm limit is used to characterize the alarm limit at which an alarm reminder is triggered when the endogenous variable sequence exceeds the alarm limit within a preset future time; a low alarm mode item is determined based on the relative relationship between the true label and the low alarm limit, wherein the low alarm limit is used to characterize the alarm limit at which an alarm reminder is triggered when the endogenous variable sequence falls below the alarm limit within a preset future time; and a dual-sided alarm mode item is determined based on the high alarm mode item and the low alarm mode item.

[0017] This invention also provides an early warning device for abnormal operating conditions in industrial production processes. The device includes: an acquisition module for acquiring a sequence of process variables of the industrial production process to be predicted; a division module for dividing the sequence of process variables into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified process variable sequence within the process variable sequence; and the exogenous variable sequence is a process variable sequence other than the endogenous variable sequence within the process variable sequence; a first construction module for constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for characterizing the endogenous variable sequence; and a second construction module for constructing a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence. A new module is used to update the first spatial projection based on the first spatial projection and the second spatial projection to obtain an updated first spatial projection; a mapping module is used to obtain a first spatial mapping result and a second spatial mapping result based on the updated first spatial projection, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in a preset dimension latent space; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the preset dimension latent space; a prediction module is used to predict the endogenous variable sequence at a preset future time based on the first spatial mapping result and the second spatial mapping result; and an early warning module is used to provide early warning for abnormal operating conditions of the industrial production process to be predicted based on the endogenous variable sequence at the preset future time and an alarm limit.

[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the abnormal operating condition early warning method for industrial production processes as described above.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the abnormal operating condition early warning method for industrial production processes as described above.

[0020] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the abnormal operating condition early warning method for industrial production processes as described above.

[0021] This invention provides a method, apparatus, and electronic device for early warning of abnormal operating conditions in industrial production processes. The method includes: acquiring a sequence of process variables for the industrial production process to be predicted; dividing the sequence of process variables into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified sequence of process variables; and the exogenous variable sequence is a sequence of process variables other than the endogenous variable sequence; constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for representing the endogenous variable sequence; and constructing a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence. Based on the first spatial projection and the second spatial projection, the first spatial projection is updated to obtain an updated first spatial projection. Based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in a preset dimension latent space; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the preset dimension latent space; based on the first spatial mapping result and the second spatial mapping result, the endogenous variable sequence at a preset future time is predicted; based on the endogenous variable sequence at the preset future time and the alarm limit, an early warning is given for abnormal operating conditions of the industrial production process to be predicted. This invention achieves accurate early warning for abnormal operating conditions of industrial production processes. Attached Figure Description

[0022] 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the abnormal operating condition early warning method for industrial production processes provided by the present invention.

[0024] Figure 2 This is a schematic diagram of the process of constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence provided by the present invention.

[0025] Figure 3 This is a schematic diagram of the process provided by the present invention for obtaining the first spatial mapping result and the second spatial mapping result based on the updated first spatial projection.

[0026] Figure 4 This is a schematic diagram of the structure of the early warning device for abnormal operating conditions in industrial production processes provided by the present invention.

[0027] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] The industrial production process abnormal condition early warning method provided by this invention can provide advanced early warning for multivariate time series data of large-scale continuous chemical plants (such as fluidized catalytic cracking FCC units). It can be used to predict key process variables (such as key pressure, temperature, liquid level, flow rate, etc.) within the next few minutes and provide forward-looking warnings for abnormal conditions that are about to exceed limits. For ease of explanation, this solution will be illustrated using a large-scale continuous chemical plant (such as the stabilizer tower of a fluidized catalytic cracking FCC unit) as an example.

[0030] Figure 1 This is a flowchart illustrating the abnormal operating condition early warning method for industrial production processes provided by the present invention.

[0031] The following will combine Figure 1 The process of the abnormal operating condition early warning method for industrial production processes provided by the present invention will be described.

[0032] In an exemplary embodiment of the present invention, combined with Figure 1 As can be seen, the early warning method for abnormal operating conditions in industrial production processes may include steps 110 to 180, and each step will be described below.

[0033] In step 110, the sequence of process variables for the industrial production process to be predicted is obtained.

[0034] In one embodiment, a distributed control system (DCS) or historical database can be used to collect time-series data from N process measurement points at T consecutive sampling times, forming a multivariate historical matrix. These process variables include, but are not limited to, the pressure at the top of the stabilizer, the temperature at the bottom of the stabilizer, the reflux flow rate, and the upstream pressure.

[0035] In step 120, the process variable sequence is divided into an endogenous variable sequence and an exogenous variable sequence. The endogenous variable sequence is the process variable sequence specified by the user in the process variable sequence; the exogenous variable sequence is the process variable sequence other than the endogenous variable sequence in the process variable sequence.

[0036] In one embodiment, the user can specify a key process variable as an endogenous variable based on the warning target, such as the pressure at the top of the stabilizer column. The remaining N-1 process variables (such as column bottom temperature, reflux flow rate, upstream pressure, etc.) are all treated as exogenous variables. Thus, the endogenous variable sequence and the exogenous variable sequence are separated from the original multivariate sequence.

[0037] In step 130, a first spatial projection corresponding to the endogenous variable sequence is constructed based on the endogenous variable sequence, wherein the first spatial projection includes a global token used to characterize the endogenous variable sequence.

[0038] In another embodiment, the endogenous variable sequence can be input into an embedding network. First, the endogenous variable sequence is reversibly instance-normalized to eliminate distribution bias during training and inference. Then, using a sliding window or non-overlapping segmentation method, the sequence of length L is divided into several patches, each patch having a length of... Each tile is projected onto a linear mapping layer (without bias). The latent space is constructed and positional encodings are superimposed to obtain a tile embedding sequence (corresponding to the superimposed tiles below). Furthermore, a learnable global token with the same dimensionality as the tile embeddings is appended to the end of the tile embedding sequence. This global token is independent of any specific time step; its initial value is randomly generated and optimized during training. This yields the first spatial projection. .

[0039] In step 140, a second spatial projection corresponding to the exogenous variable sequence is constructed based on the exogenous variable sequence.

[0040] For sequences of exogenous variables, an inverted embedding approach can be used, treating all L historical values ​​of each exogenous variable as a single "variable vector," and mapping it to a linear projection layer. The latent space of dimension 1 yields a token corresponding to each exogenous variable, which is also the projection of the second space. , where k represents the k-th exogenous variable.

[0041] In step 150, the first spatial projection is updated based on the first spatial projection and the second spatial projection to obtain the updated first spatial projection.

[0042] In one embodiment, the first space can be projected. Second space projection The data is fed into an update network consisting of multiple stacked encoders. Each encoder layer performs the following operations: Projection of the first space Perform multi-head self-attention computation to achieve information fusion between graph tiles and between graph tiles and the global token; Projection from the first space Extract the global token as a query and project it into the second space. As keys and values, multi-head cross-attention computation is performed, enabling global tokens to aggregate exogenous variable information; Update the global token to At the end of the first spatial projection, the image is then processed through a position-aware feedforward network (FFN) and residual connections to obtain the updated first spatial projection.

[0043] This update process can be repeated multiple times, eventually yielding a first spatial projection containing global tokens that integrates endogenous and exogenous information, and finally obtaining the updated first spatial projection.

[0044] In step 160, based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained respectively. The first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in the latent space of the preset dimension; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the latent space of the preset dimension.

[0045] In one embodiment, latent representations can be extracted from the updated first spatial projection, which may specifically include the first spatial mapping result and the second spatial mapping result.

[0046] To obtain the first spatial mapping result, the updated first spatial projection (containing all positions of tiles and global tokens) is flattened as a whole, or it is aggregated in the spatial dimension (e.g., global average pooling), and then mapped to the H dimension (H is the prediction step size) through a linear prediction head (Flatten + Linear) to obtain the first spatial mapping result. This result is used to characterize the absolute position information of the endogenous variable sequence in the latent space, such as whether the current pressure is high, medium, or low.

[0047] To obtain the second spatial mapping result, a first-order difference is calculated along the tile dimension of the updated first spatial projection, i.e., the representations of adjacent tiles are subtracted to obtain the change tensor ΔZ. This change tensor is then mapped to the H dimension through a linear prediction head with another independent parameter to obtain the second spatial mapping result. This result is used to characterize the trend changes of endogenous variable sequences in the latent space, such as whether pressure is rising or falling, and at what rate.

[0048] In step 170, based on the first spatial mapping result and the second spatial mapping result, the sequence of endogenous variables at a future preset time is predicted.

[0049] In one embodiment, the first spatial mapping result and the second spatial mapping result can be combined to obtain the final multi-step advance prediction value. The combination method can employ weighted summation (e.g., learning a gating factor g (corresponding to the scalar gate below) to make... ,in, This represents the sequence of endogenous variables at a predetermined future time. They can also be directly added together or concatenated before linear mapping. This yields the predicted sequence of endogenous variables for the next H time points.

[0050] In step 180, based on the sequence of endogenous variables and alarm limits at a preset future time, an early warning is issued for abnormal operating conditions in the industrial production process to be predicted.

[0051] In one embodiment, the predicted sequence of future H steps is compared with pre-activated process alarm limits. If a certain predicted time exists... , making or Indicates a high alarm limit; This indicates a low alarm limit. In this case, an alarm will be immediately sent to the DCS / alarm management system. A warning that the limit will be exceeded in minutes (among which) (For sampling interval), and simultaneously display the over-limit location, over-limit direction, predicted amplitude and early warning duration on the operator's HMI, allowing response time for manual or automatic coordinated intervention.

[0052] In this embodiment, by generating a "first spatial mapping result" (representing absolute position / level) and a "second spatial mapping result" (representing trend change) respectively, the two types of information that were originally mixed in the same model are decoupled, focusing on steady-state mean prediction and on capturing the direction and rate of change respectively. Each performs its own function, avoiding the problem that the trend signal is masked by the level signal at the alarm threshold, thereby significantly improving the early warning recall rate for impending over-limit conditions.

[0053] This invention provides a method for early warning of abnormal operating conditions in industrial production processes. The method includes: acquiring a sequence of process variables for the industrial production process to be predicted; dividing the sequence of process variables into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified sequence of process variables; and the exogenous variable sequence is a sequence of process variables other than the endogenous variable sequence; constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for representing the endogenous variable sequence; constructing a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence; and based on the... The invention describes a first spatial projection and a second spatial projection. The first spatial projection is updated to obtain an updated first spatial projection. Based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained. The first spatial mapping result characterizes the absolute position information of the endogenous variable sequence in a preset-dimensional latent space; the second spatial mapping result characterizes the trend change information of the endogenous variable sequence in the preset-dimensional latent space. Based on the first and second spatial mapping results, the endogenous variable sequence at a preset future time is predicted. Based on the endogenous variable sequence at the preset future time and an alarm limit, an early warning is issued for abnormal operating conditions in the industrial production process to be predicted. This invention achieves accurate early warning for abnormal operating conditions in industrial production processes.

[0054] Figure 2 This is a schematic diagram of the process of constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence provided by the present invention.

[0055] The following will combine Figure 2 The process of constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence provided by the present invention will be described.

[0056] In an exemplary embodiment of the present invention, combined with Figure 2 As can be seen, constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence may include steps 210 to 240, and each step will be described below.

[0057] In step 210, the endogenous variable sequence is divided according to a preset time block length to obtain multiple blocks, wherein the blocks are used to represent the endogenous variable subsequences obtained by dividing the endogenous variable sequence according to the preset time block length.

[0058] In one embodiment, the endogenous variable sequence can be... According to the preset time block length Perform non-overlapping segmentation to obtain There are 1000 tiles; among them... Represents a real number; L represents the length of the endogenous variable sequence. The preset time block length can be pre-set by the user according to the sampling frequency and process dynamic characteristics; each block represents a continuous time subsequence, i.e., an "endogenous variable subsequence", with a length of [missing information]. These tiles are arranged chronologically to capture patterns of change within a local time window.

[0059] In step 220, for any tile, a sine and cosine position code is superimposed on the tile to obtain the superimposed tile, and the superimposed tile is projected onto a representation space of a preset dimension through a biasless linear mapping to obtain the tile space projection corresponding to the tile.

[0060] For any tile Perform the following operations: Since the tile sequence has a natural temporal order, a sine-cosine position code corresponding to its position index i is generated for each tile. The code is then added to (or combined with) the tile itself to obtain the overlaid tile. .

[0061] Furthermore, the superimposed tiles are passed through an unbiased linear mapping layer. Projecting onto a representation space of a preset dimension yields a representation of dimension . The linear mapping does not include a bias term to reduce the number of parameters and avoid overfitting. This step maps each tile from the original numerical space to a high-dimensional latent space, forming a tile space projection corresponding to the i-th tile. Among them, the spatial projection of the tile corresponding to the tile This can be expressed as formula (1): (1) In step 230, multiple tile spatial projections are sorted according to the time sequence to obtain a tile spatial projection sequence.

[0062] In step 240, a global token is appended to the end of the tile spatial projection sequence to construct the first spatial projection corresponding to the endogenous variable sequence.

[0063] In yet another embodiment, the tiles can be arranged in chronological order (i.e., i from 1 to...). All tile spatial projections are arranged into a sequence that maintains a temporal order for subsequent encoder processing.

[0064] Furthermore, to enable the encoder to maintain explicit representation of the overall sequence level while performing self-attention within the tiles, a learnable global token is concatenated at the end of the aforementioned tile spatial projection sequence. The dimension of the global token is exactly the same as that of the tile spatial projection. Its initial value can be randomly initialized (e.g., sampled from a uniform or normal distribution) and updated via backpropagation during model training. After concatenation, the final first spatial projection is obtained, which can be expressed as Equation (2): (2) This global token acts as a "sequence-level summary" of the variable and is the only bridge for subsequent cross-variable attention with exogenous variables.

[0065] This completes the process of constructing the first spatial projection from the endogenous variable sequence. This first spatial projection will subsequently be used for self-attention, cross-attention, and horizontal / trend decoupling prediction.

[0066] In yet another exemplary embodiment of the present invention, continuing with the previously described embodiments, the global token can be obtained in the following manner: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a global token generation module, and the global token generation module includes a first unbiased linear mapping layer. The endogenous variable sequence is input into the working condition anomaly prediction model, and linear mapping is performed through the first unbiased linear mapping layer in the global token generation module to obtain a global token used to represent the endogenous variable sequence.

[0067] In one embodiment, the abnormal operating condition prediction model is a deep learning model that includes a global token generation module specifically designed to generate global tokens from a sequence of endogenous variables. The global token generation module includes at least a first unbiased linear mapping layer.

[0068] In application, the complete sequence of endogenous variables can be used as input and fed into the global token generation module. The first unbiased linear mapping layer in this module is defined as follows: This does not include bias terms. Mapping is performed through this linear mapping layer, for example, it could be... = ( ).

[0069] in, Consider it as an L-dimensional column vector. for The weight matrix, output For one A dimensional vector, which is the global token we are looking for.

[0070] It should be noted that this linear mapping layer does not contain a bias term (i.e., it is unbiased), which is consistent with the unbiased design used when linearly mapping tiles, helping to reduce the number of parameters and lower the risk of overfitting. The weight matrix of the first unbiased linear mapping layer... These are the trainable parameters of the model, which are learned end-to-end during model training through backpropagation and optimization algorithms (such as Adam).

[0071] In yet another exemplary embodiment of the present invention, continuing with the previously described embodiments, the construction of a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence can be achieved in the following manner: The pre-trained abnormal working condition prediction model is invoked. The abnormal working condition prediction model includes an exogenous variable inverted variable embedding module, which includes a second unbiased linear mapping layer. The exogenous variable sequence is input into the working condition anomaly prediction model, and linear mapping is performed through the second unbiased linear mapping layer in the exogenous variable inversion embedding module to obtain the second spatial projection corresponding to the exogenous variable sequence.

[0072] In one embodiment, the abnormal operating condition prediction model is a deep learning model. This model internally includes an exogenous variable inversion embedding module, which is specifically used to convert the exogenous variable sequence into a corresponding second spatial projection. The exogenous variable inversion embedding module includes at least one second unbiased linear mapping layer.

[0073] In application, the exogenous variable sequence can be... The input is fed into the exogenous variable inversion embedding module. This module performs an inversion embedding operation independently for each exogenous variable. For the k-th exogenous variable... The observations of its complete L historical moments constitute a vector. The vector is then passed through the second unbiased linear mapping layer. (Excluding bias terms) A linear mapping process is performed to obtain the second spatial projection corresponding to the exogenous variable sequence, wherein the second spatial projection can be expressed by formula (3): (3) in, Indicates the total number of exogenous variable sequences; The weight matrix is ​​learned end-to-end during model training and no bias term is set. It is called the "second unbiased linear mapping layer" to distinguish it from the "first unbiased linear mapping layer" in the global token generation module (although the two structures can be the same, their parameters are independent).

[0074] Unlike endogenous paths, exogenous branches treat all L historical values ​​of each exogenous variable as a single "variable vector," mapping them through linear projection to... A dimensional space is formed, representing each exogenous variable as a token, that is... This variable-level inverted embedding preserves the identifiability of exogenous variables in the variable dimension, preventing them from being overwhelmed by time-dimensional details at the cross-variable stage, while also reducing the computational complexity of the exogenous side from... Compress to This facilitates online deployment.

[0075] In yet another exemplary embodiment of the present invention, continuing with the previously described embodiments, after obtaining the process variable sequence of the industrial production process to be predicted, the method further includes: Obtain multiple sets of historical process variable sequences; Based on the historical process variable sequence, the mean and standard deviation of each process variable are determined; For any of the process variable sequences, the process variable sequences are normalized based on the mean and standard deviation to obtain normalized process variable sequences. The normalized process variable sequence is used as the process variable sequence.

[0076] Normalization can be performed before constructing the first and second spatial projections to ensure consistent numerical scales among different process variables and improve the training stability and generalization ability of the model.

[0077] In one embodiment, a large number of historical samples can be extracted from a historical database of the production process. These historical samples cover multivariate time series data under different operating conditions, including normal operating conditions, transitional operating conditions, and critical operating conditions approaching alarm limits. Each set of historical process variable sequences contains the same N process variables (including endogenous and exogenous variables) as the sample to be predicted, and has the same historical sequence length L. The number of historical samples should typically reach several thousand or even tens of thousands to robustly estimate the statistical distribution parameters of each variable.

[0078] For each process variable (a total of N variables), calculate its mean at all times across all historical samples. and standard deviation .

[0079] Furthermore, for the current process variable sequence to be predicted... It can be based on the mean. and standard deviation After performing variable-wise normalization, it can be expressed as formula (4): (4) in, This represents the sequence of process variables after normalization.

[0080] After normalization, the resulting normalized process variable sequence is used as the input process variable sequence for subsequent steps. All subsequent operations (including partitioning endogenous / exogenous variables, constructing the first / second spatial projection, encoder calculation, prediction, and loss calculation) are based on this normalized sequence.

[0081] It should be noted that at the final output of the model (i.e., after predicting the future sequence of endogenous variables), it is necessary to use the stored mean. and standard deviation Inverse normalization is performed to restore the predicted values ​​to the original process dimensions, so that they can be compared with the process alarm limits and an early warning can be issued.

[0082] Different process variables in chemical processes often have different physical units and numerical ranges (e.g., pressure in megapascals, temperature in hundreds of degrees Celsius, flow rate in tons per hour). Without normalization, variables with larger values ​​will dominate the loss function, causing the model to ignore variables with smaller values. This embodiment transforms each variable to the same scale using the mean and standard deviation, enabling the model to learn the contribution of each variable equally and improving the accuracy of multivariate coupled modeling.

[0083] In yet another exemplary embodiment of the present invention, the following description continues using the previously described embodiments. Based on the first spatial projection and the second spatial projection, updating the first spatial projection to obtain the updated first spatial projection can be achieved in the following manner: A pre-trained abnormal working condition prediction model is invoked. This model includes a decoupled encoder module, which is composed of multiple stacked encoders. The initial updated first spatial projection obtained from the output of the current encoder layer is used as the first spatial projection input to the next encoder layer, and so on, until the output of the last encoder layer is the initial updated first spatial projection. This final updated first spatial projection is then used as the updated first spatial projection. The encoder for each layer performs the following process: Multi-head self-attention processing is applied to the first spatial projection to obtain the self-attention processed first spatial projection. After stitching together the first spatial projection and the self-attention processed first spatial projection, and then performing layer normalization, the aggregated first spatial projection is obtained. The global token in the first spatial projection after aggregation is used as the query, and the second spatial projection is used as the key and value for cross-attention processing to obtain the cross-attention processing result. After concatenating the global token in the first spatial projection after aggregation and the cross-attention processing result, and then performing layer normalization, the global token after cross-variable processing is obtained. After cross-variable processing, the global token is backfilled into the aggregated first spatial projection to obtain the first spatial projection updated by the encoder of this layer. A nonlinear transformation is performed on the first spatial projection after the encoder of this layer is updated to obtain the transformed first spatial projection. After stitching together the first spatial projection updated by the encoder of this layer and the first spatial projection after transformation, and then performing layer normalization processing, the initial updated first spatial projection is obtained.

[0084] In one embodiment, the update process is performed by a decoupled encoder module in a pre-trained condition anomaly prediction model, the decoupled encoder module being composed of multiple stacked encoders.

[0085] During the application process, the first spatial projection has been obtained. (Including a global token) and second space projection The decoupled encoder module has a total of Layers, where the input of each layer is the output of the layer above (the input of the first layer is...). The output of the current layer is used as the input of the next layer. Now, let's take the first layer as an example. The layer encoder describes its execution process: Projection of the first space Multi-head self-attention is performed to enable bidirectional information aggregation between tiles and between tiles and the global token. Residual connections and layer normalization are used between the input and output of self-attention. Specifically, multi-head self-attention processing is applied to the first spatial projection to obtain the self-attention-processed first spatial projection. Then project the first space. and the first spatial projection after self-attention processing After stitching, and following layer normalization, the aggregated first spatial projection is obtained. .

[0086] In yet another embodiment, the aggregated first spatial projection can be taken. The last position (i.e., the global token) (as a query) Multi-head cross-attention is performed using the key / value pairs. Specifically, the global token in the aggregated first spatial projection is used as the query, and the second spatial projection is used as the key and value for cross-attention processing to obtain the cross-attention processing result. ; The global token in the first spatial projection after aggregation After concatenating the results of cross-attention processing, and then performing layer normalization, the global token after cross-variable processing is obtained. ;in, .

[0087] In yet another embodiment, the updated global token can be... Backfill to At the end of the layer, the position of the tile remains unchanged, forming a new layer. Specifically, the global token after cross-variable processing can be backfilled into the aggregated first spatial projection to obtain the updated first spatial projection of the current layer encoder. .

[0088] In another embodiment, a feedforward network consisting of two one-dimensional convolutions with kernel size of 1 and activation functions can also be used. A position-by-position nonlinear transformation is performed, and residual connectivity and layer normalization are superimposed for processing to obtain the first spatial projection after the initial update. Specifically, a nonlinear transformation is performed on the first spatial projection after the encoder of this layer is updated to obtain the transformed first spatial projection. Then update the first spatial projection after updating the encoder of this layer. and the transformed first spatial projection After stitching, and then undergoing layer normalization, the initial updated first spatial projection is obtained. .

[0089] through After stacking layer encoders, the initial updated first spatial projection output by the last layer encoder can be used as the updated first spatial projection. .

[0090] It should be noted that the parameters involved in the decoupled encoder module in the aforementioned embodiments are the trainable parameters of the model, which are learned end-to-end through backpropagation and optimization algorithms (such as Adam) during the model training process.

[0091] Figure 3 This is a schematic diagram of the process provided by the present invention for obtaining the first spatial mapping result and the second spatial mapping result based on the updated first spatial projection.

[0092] The following will combine Figure 3The process of obtaining the first spatial mapping result and the second spatial mapping result based on the updated first spatial projection provided by the present invention will be described.

[0093] In an exemplary embodiment of the present invention, combined with Figure 3 As can be seen, obtaining the first spatial mapping result and the second spatial mapping result based on the updated first spatial projection can include steps 310 to 350, and each step will be described below.

[0094] In step 310, the pre-trained working condition anomaly prediction model is invoked, wherein the working condition anomaly prediction model includes a horizontal expert prediction module and a trend expert prediction module; wherein the horizontal expert prediction module includes a first multilayer perceptron; and the trend expert prediction module includes a second multilayer perceptron.

[0095] In step 320, the updated first spatial projection is input into the abnormal operating condition prediction model, and the updated first spatial projection is spatially mapped by the horizontal expert prediction module to obtain the first spatial mapping result.

[0096] In one embodiment, a pre-trained abnormal working condition prediction model can be invoked, which includes a horizontal expert prediction module. The horizontal expert prediction module includes a first multilayer perceptron (first MLP). The structure of the first MLP can be: the input layer receives flattened features, passes them through one or more hidden layers (e.g., containing the GELU activation function), and finally outputs a vector with dimension H (prediction step size).

[0097] During application, the updated first spatial projection can be used. The data is input into the model. In the horizontal expert prediction module, the data is first flattened into a one-dimensional vector along the tile dimension (including the global token), and then this flattened vector is fed into the first MLP (also known as...). The spatial mapping process is performed to obtain an output vector of dimension H, which is denoted as the first spatial mapping result. Among them, the first spatial mapping result This can be achieved using formula (5): (5) The first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in the preset dimension latent space, such as the current level of the target variable (high pressure zone, medium pressure zone or low pressure zone) and the vicinity of the mean to be regressed.

[0098] Among them, horizontal experts characterize the target variable in It contains absolute positional information in the latent space, and therefore excels at predicting patterns related to absolute amplitude, such as "what level the target variable is currently at" and "which mean it will return to."

[0099] In step 330, the updated first spatial projection is divided according to a preset time block length to obtain multiple projection blocks, wherein the projection blocks are used to represent the first spatial projection subsequence obtained by dividing the updated first spatial projection according to the preset time block length.

[0100] In step 340, the adjacent projected tiles are subjected to first-order difference processing in the time dimension to obtain the change tensor between the projected tiles.

[0101] In yet another embodiment, the same updated first spatial projection can be used. The image is divided according to the preset time block length to obtain multiple projection blocks.

[0102] Furthermore, for the projected tile sequence obtained from the above segmentation, a first-order difference processing is performed on adjacent projected tiles in the time dimension (i.e., the projected tile index dimension). Specifically, the difference between two adjacent projected tiles is calculated, that is, the adjacent tile representations are subtracted element by element to obtain the tile change tensor ΔZ.

[0103] In step 350, the change tensor between projected tiles is input into the abnormal operating condition prediction model, and the change tensor between projected tiles is spatially mapped by the trend expert prediction module to obtain the second spatial mapping result.

[0104] In another embodiment, the abnormal operating condition prediction model can be invoked, which further includes a trend expert prediction module. The trend expert prediction module includes a second multilayer perceptron (second MLP), the structure of which can be the same as the first MLP but with independent parameters.

[0105] The tensor ΔZ representing the changes between the projected plots is input into the trend expert prediction module. First, ΔZ is flattened into a one-dimensional vector, and then this flattened vector is fed into the second MLP (also known as...). The spatial mapping process is performed to obtain an output vector of dimension H, which is denoted as the second spatial mapping result. Among them, the second space mapping result This can be achieved using formula (6): (6) The second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the latent space, such as whether the target variable is rising or falling, whether the rate of change is accelerating or slowing down, and is especially sensitive to the trend of change near the alarm critical zone.

[0106] Unlike direct prediction of Z, the first-order differencing operation explicitly erases the absolute level component at the latent space level, retaining only the relative dynamic information of how it evolves over time. Therefore, trend experts naturally focus on the direction and rate of increase / decrease of the target variable, are more sensitive to trend capture in the alarm threshold, and complement the functions of level experts.

[0107] It should be noted that the parameters involved in the horizontal expert prediction module and the trend expert prediction module in the aforementioned embodiments are the trainable parameters of the model, which are learned end-to-end through backpropagation and optimization algorithms (such as Adam) during the model training process.

[0108] In yet another exemplary embodiment of the present invention, continuing with the description of the aforementioned embodiments, before predicting the endogenous variable sequence at a future preset time based on the first spatial mapping result and the second spatial mapping result, the method further includes: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a gating module, and the gating module includes a third unbiased linear mapping layer, a fourth unbiased linear mapping layer, a GELU activation function, and a Sigmoid activation function; The updated global token in the first spatial projection is input into the working condition anomaly prediction model, and linear mapping is performed through the third unbiased linear mapping layer in the gating module to obtain the linearly mapped global token. The global token after linear mapping is processed based on the GELU activation function to obtain the first global token after activation. Based on the fourth unbiased linear mapping layer, the global token after the first activation process is linearly mapped to obtain the reprocessed global token. The reprocessed global token is processed based on the Sigmoid activation function to obtain a scalar gate; The prediction of the endogenous variable sequence at a predetermined future time, based on the first and second spatial mapping results, can be achieved in the following manner: The difference between 1 and the scalar gate is calculated to obtain the subtracted scalar gate. The first spatial mapping result, the scalar gate, the second spatial mapping result, and the scalar gate after the difference processing are weighted and summed to predict the endogenous variable sequence at a future preset time.

[0109] After obtaining the first and second spatial mapping results and before fusing them into the final predicted value, the following gating weight (i.e., scalar gate) generation process is performed. This process is executed by the gating module in the pre-trained abnormal condition prediction model, which includes: a third unbiased linear mapping layer, a fourth unbiased linear mapping layer, a GELU activation function, and a Sigmoid activation function. It should be noted that the parameters involved in the gating module in the aforementioned embodiment are the trainable parameters of the model, which are learned end-to-end through backpropagation and optimization algorithms (such as Adam) during model training.

[0110] In one embodiment, a global token located at the end can be extracted from the updated first spatial projection, denoted as... The global token is then input into the gating module. This is processed through the third unbiased linear mapping layer in the gating module. Perform a linear mapping on the global token to obtain the linearly mapped global token. .

[0111] The above linear mapping result is then input into the GELU activation function (Gaussian Error Linear Unit) for processing to obtain the global token after the first activation process. .

[0112] Furthermore, the global token processed by the first activation step is sent to the fourth unbiased linear mapping layer. Perform remapping to obtain the reprocessed global token. The fourth layer weight matrix outputs a scalar, and also does not contain a bias term.

[0113] In another embodiment, the reprocessed global token can be fed into a Sigmoid activation function to compress it to the (0,1) interval, resulting in a scalar gate. ,in, .

[0114] scalar gate The physical meaning of is the weight of the horizontal expert prediction results in the final prediction; correspondingly, the weight of the trend expert prediction results is 1- Also known as a differential scalar gate.

[0115] In another embodiment, the first spatial mapping result, the scalar gate, the second spatial mapping result, and the scalar gate after subtraction can be weighted and summed to predict the endogenous variable sequence at a predetermined future time. Specifically, the endogenous variable sequence at the predetermined future time can be expressed as formula (6): (6) in, This represents the sequence of endogenous variables over a predetermined future time period.

[0116] When variables are in a stationary phase, the gating output is biased towards level experts to maintain steady-state mean predictions; when variables enter a trend phase (accelerated rise or fall), the gating output is biased towards trend experts to amplify the response to change. Because the gating input comes from the global summary rather than a single tile, the gating is robust on a time scale, avoiding the high-frequency jitter caused by point-to-point routing.

[0117] In yet another exemplary embodiment of the present invention, the abnormal operating condition prediction model can be trained in the following manner: Obtain multiple sets of historical process variable sequences and corresponding real labels. The real labels are used to characterize the historical endogenous variable sequences corresponding to the historical process variable sequences at a future preset time. A training dataset is constructed based on the historical process variable sequence and the real labels corresponding to the historical process variable sequence; Construct a target training function, which is determined based on the relative relationship between the true labels and alarm limits in the training dataset; Based on the training dataset and the target training function, the working condition anomaly prediction model is trained in an end-to-end training manner to obtain the trained working condition anomaly prediction model.

[0118] In one embodiment, multiple sets of historical process variable sequences are acquired, along with corresponding ground truth labels for each set. These historical process variable sequences are collected from a historical database of industrial production processes or a distributed control system (DCS), covering multivariate time series data of the unit under different operating conditions (normal, transition, and abnormal approximation conditions). Each sequence contains observations of N process variables over L consecutive historical time points. Ground truth labels characterize the sequence of historical endogenous variables corresponding to the historical process variable sequences at a predetermined future time. For the m-th sample, its label is the actual observation value of the endogenous variable (a user-specified key process variable, such as tower top pressure) over H consecutive time points following that sample. H is the prediction step size.

[0119] In another embodiment, the aforementioned multiple sets of historical process variable sequences and their corresponding real labels can be divided into training, validation, and test sets according to time sequence (or random sampling). A rolling window time series partitioning method is typically used to ensure that the training set is earlier than the validation and test sets in time, thus avoiding future information leakage. The training dataset contains at least a training set portion for model parameter learning; the validation set is used for hyperparameter tuning and early stopping detection; and the test set is used for final performance evaluation.

[0120] In another embodiment, a target training function (i.e., a loss function) can be constructed. The core feature of this target training function is that it is determined based on the relative relationship between the true labels and the alarm limits in the training dataset. In other words, for different true label values, the loss function assigns different weights or uses different calculation forms, so that the model pays more attention to those "tail samples" that are close to or exceed the alarm limits during training.

[0121] Furthermore, the abnormal operating condition prediction model can be trained using an end-to-end training approach based on the aforementioned training dataset and target training function. Specifically: all model parameters (including the first unbiased linear mapping layer, the second unbiased linear mapping layer, the weights of each layer of the decoupled encoder, the first and second MLPs, the third and fourth unbiased linear mapping layers, etc.) are treated as trainable parameters. A mini-batch stochastic gradient descent optimizer is employed. For each batch, the historical process variable sequence is input into the current model, forward propagation is performed to obtain the predicted values, then the target training function value is calculated, and then backpropagation is performed to calculate the gradient and update all parameters.

[0122] During training, the loss is calculated on the validation set at regular intervals (epochs), and an early stopping strategy is adopted (e.g., training is stopped if the validation loss does not decrease for three consecutive epochs) to prevent overfitting. After training is completed, the model parameters that achieve the best early warning performance (such as predictive recall, false alarm rate, average lead time, etc.) on the validation set are saved as the trained anomaly prediction model.

[0123] This end-to-end training approach ensures that all components of the model (from input embeddings to the gated weights of the final prediction head) are jointly optimized toward the unified goal of minimizing tail-aware loss, rather than being trained independently in stages.

[0124] In an exemplary embodiment of the present invention, the target training function can be determined in the following manner, using the previously described embodiments as an example: Determine the alarm mode items, which include high alarm mode items, low alarm mode items, and dual alarm mode items where high alarm mode items and low alarm mode items coexist. Based on alarm mode items, real labels, and prediction results obtained by inputting training data from the training dataset into the abnormal working condition prediction model, a target training function is constructed.

[0125] In one embodiment, the target training function can be expressed as formula (7): (7) in, Represents the target training function; Indicates the prediction step size; This represents the true label corresponding to the i-th sample; This represents the prediction result obtained by inputting the i-th sample into the abnormal working condition prediction model; This represents the alarm mode item corresponding to the i-th sample.

[0126] In yet another exemplary embodiment of the present invention, the alarm mode item can be determined in the following manner: Based on the relative relationship between the real labels and the high alarm limit, the high alarm mode item is determined, wherein the high alarm limit is used to characterize the alarm limit at which an alarm reminder is triggered when the endogenous variable sequence exceeds the alarm limit at a preset time in the future; Based on the relative relationship between the real label and the low alarm limit, the low alarm mode item is determined, wherein the low alarm limit is used to characterize the alarm limit at which a warning reminder is triggered when the endogenous variable sequence falls below the alarm limit at a preset time in the future; The dual-sided alarm mode item is determined based on the high alarm mode item and the low alarm mode item.

[0127] In one embodiment, to achieve higher prediction accuracy for the model near the alarm limit, this invention abandons the equal-weighted MSE for all training samples and instead calculates the MSE based on each ground truth sample. The relative position between the (corresponding real label) and the high / low alarm limits of the process is dynamically assigned a sample-level weight. Specifically, an element-wise sample weight function is defined. The alarm mode options are as follows: For the high alarm mode item: For the low alarm mode item: For the dual-alarm mode item: in, The maximum gain for the weights (the upper bound of weight w is...) ), The steepness of the Sigmoid transition (the smaller the β, the steeper the transition); , The high and low alarm limits are given by the field instruments and process design. Far below the high alarm time limit (Equivalent to standard MSE); when Approaching or exceeding the high alarm limit The contribution of the corresponding samples to gradient updates is amplified. times.

[0128] It should be noted that any combination of one or more of the above embodiments also falls within the protection scope of this solution, and will not be elaborated further here.

[0129] As described above, the present invention provides an early warning method for abnormal operating conditions in industrial production processes, which explicitly decouples "level" and "trend" information at the prediction head level. The level expert predicts the absolute amplitude component through the output of the original encoder, while the trend expert predicts the relative evolution component through first-order patch differencing. Both are handled by prediction heads with independent parameters, avoiding mutual interference between the two types of information in existing single-head structures. Furthermore, a globally summarized Sigmoid gating adaptively fuses the outputs of the two experts under different operating conditions, favoring level prediction in stable segments and amplifying the change response in trend segments, structurally improving the model's sensitivity to trend precursors in the alarm threshold. At the training mechanism level, a weighted tail-perception sample centered on the alarm limit is introduced. Through a Sigmoid-shaped sample weight function, sparse samples approaching or exceeding the alarm limit receive a loss weight of up to 1+α times, completely changing the situation where traditional MSE is dominated by massive normal operating condition samples. Without changing the network structure or introducing additional inference overhead, this significantly improves the model's prediction accuracy and early warning recall rate in the alarm threshold. The loss function is orthogonal to the model backbone, exhibiting good versatility and can be independently applied to other time series prediction networks. It maintains computational efficiency consistent with TimeXer-like tile-global token backbones. The exogenous branch employs inverted variable embedding to keep the cross-attention complexity at ONexo, offering an order-of-magnitude advantage in inference speed compared to methods that expand exogenous variables into separate time sequences. This facilitates millisecond-level online inference deployment on field edge nodes or industrial control servers.

[0130] The following describes the industrial production process abnormal condition early warning device provided by the present invention. The industrial production process abnormal condition early warning device described below and the industrial production process abnormal condition early warning method described above can be referred to in correspondence.

[0131] Figure 4 This is a schematic diagram of the structure of the early warning device for abnormal operating conditions in industrial production processes provided by the present invention.

[0132] The following will combine Figure 4 The structure of the early warning device for abnormal operating conditions in industrial production processes provided by the present invention will be described.

[0133] In an exemplary embodiment of the present invention, combined with Figure 4 As can be seen, the industrial production process abnormal condition early warning device may include an acquisition module 410, a division module 420, a first construction module 430, a second construction module 440, an update module 450, a mapping module 460, a prediction module 470, and an early warning module 480. Each module will be described below.

[0134] The acquisition module 410 can be configured to acquire a sequence of process variables for the industrial production process to be predicted; The partitioning module 420 can be configured to partition the process variable sequence into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is the user-specified process variable sequence in the process variable sequence; and the exogenous variable sequence is the process variable sequence other than the endogenous variable sequence in the process variable sequence. The first construction module 430 can be configured to construct a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for characterizing the endogenous variable sequence; The second construction module 440 can be configured to construct a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence. The update module 450 can be configured to update the first spatial projection based on the first spatial projection and the second spatial projection to obtain the updated first spatial projection. The mapping module 460 can be configured to obtain a first spatial mapping result and a second spatial mapping result based on the updated first spatial projection, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in the latent space of the preset dimension; and the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the latent space of the preset dimension. Prediction module 470 can be configured to predict the sequence of endogenous variables at a future preset time based on the first spatial mapping result and the second spatial mapping result; The early warning module 480 can be configured to provide early warning of abnormal operating conditions of the industrial production process to be predicted based on the endogenous variable sequence and alarm limit at a future preset time.

[0135] In yet another exemplary embodiment of the present invention, the first construction module 430 may construct a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence in the following manner: The endogenous variable sequence is divided into multiple blocks according to a preset time block length, wherein the blocks are used to represent the endogenous variable subsequences obtained by dividing the endogenous variable sequence according to the preset time block length; For any given tile, a sine and cosine position code is superimposed on the tile to obtain a superimposed tile, and the superimposed tile is projected onto a representation space of a preset dimension through an unbiased linear mapping to obtain a tile space projection corresponding to the tile. According to the time sequence, the multiple spatial projections of the tiles are sorted to obtain the spatial projection sequence of the tiles; By concatenating the global token at the end of the tile spatial projection sequence, a first spatial projection corresponding to the endogenous variable sequence is constructed.

[0136] In yet another exemplary embodiment of the present invention, the first construction module 430 may obtain the global token in the following manner: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a global token generation module, and the global token generation module includes a first unbiased linear mapping layer. The endogenous variable sequence is input into the working condition anomaly prediction model, and linear mapping is performed through the first unbiased linear mapping layer in the global token generation module to obtain a global token used to characterize the endogenous variable sequence.

[0137] In yet another exemplary embodiment of the present invention, the second construction module 440 may construct a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence in the following manner: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes an exogenous variable inverted variable embedding module, and the exogenous variable inverted variable embedding module includes a second unbiased linear mapping layer; The exogenous variable sequence is input into the working condition anomaly prediction model, and linear mapping is performed through the second unbiased linear mapping layer in the exogenous variable inversion embedding module to obtain the second spatial projection corresponding to the exogenous variable sequence.

[0138] In yet another exemplary embodiment of the present invention, the acquisition module 410 may also be configured to: Obtain multiple sets of historical process variable sequences; Based on the historical process variable sequence, the mean and standard deviation of each process variable are determined; For any of the process variable sequences, the process variable sequences are normalized based on the mean and standard deviation to obtain normalized process variable sequences. The normalized process variable sequence is used as the process variable sequence.

[0139] In another exemplary embodiment of the present invention, the updating module 450 may update the first spatial projection based on the first spatial projection and the second spatial projection to obtain the updated first spatial projection in the following manner: A pre-trained abnormal working condition prediction model is invoked. This model includes a decoupled encoder module, which is composed of multiple stacked encoders. The initial updated first spatial projection obtained from the output of the current encoder layer is used as the first spatial projection input to the next encoder layer, and so on, until the last encoder layer outputs its initial updated first spatial projection. This initial updated first spatial projection from the last encoder layer is then used as the updated first spatial projection. The encoder for each layer performs the following process: Perform multi-head self-attention processing on the first spatial projection to obtain the self-attention processed first spatial projection; After stitching the first spatial projection and the self-attention processed first spatial projection together, and then performing layer normalization processing, the aggregated first spatial projection is obtained. Using the global token in the first spatial projection after aggregation as the query and the second spatial projection as the key and value, cross-attention processing is performed to obtain the cross-attention processing result. After concatenating the global token in the first spatial projection after aggregation and the cross-attention processing result, and then performing layer normalization processing, the global token after cross-variable processing is obtained. The global token after cross-variable processing is backfilled into the aggregated first spatial projection to obtain the first spatial projection after encoder update of this layer. A nonlinear transformation is performed on the updated first spatial projection of the encoder at this layer to obtain the transformed first spatial projection. After stitching together the first spatial projection updated by the encoder of this layer and the first spatial projection after transformation, and then performing layer normalization processing, the initial updated first spatial projection is obtained.

[0140] In yet another exemplary embodiment of the present invention, the mapping module 460 may obtain the first spatial mapping result and the second spatial mapping result based on the updated first spatial projection in the following manner: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a horizontal expert prediction module and a trend expert prediction module; wherein the horizontal expert prediction module includes a first multilayer perceptron; and the trend expert prediction module includes a second multilayer perceptron. The updated first spatial projection is input into the working condition anomaly prediction model, and the updated first spatial projection is spatially mapped by the horizontal expert prediction module to obtain the first spatial mapping result. The updated first spatial projection is divided according to a preset time block length to obtain multiple projection blocks, wherein the projection blocks are used to represent the first spatial projection subsequence obtained by dividing the updated first spatial projection according to the preset time block length. The first-order difference processing of adjacent projected tiles in the time dimension is performed to obtain the change tensor between projected tiles. The tensor of changes between the projected tiles is input into the abnormal operating condition prediction model, and the tensor of changes between the projected tiles is spatially mapped by the trend expert prediction module to obtain a second spatial mapping result.

[0141] In yet another exemplary embodiment of the present invention, the prediction module 470 may also be configured to: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a gating module, and the gating module includes a third unbiased linear mapping layer, a fourth unbiased linear mapping layer, a GELU activation function, and a Sigmoid activation function; The updated global token in the first spatial projection is input into the working condition anomaly prediction model, and linear mapping is performed through the third unbiased linear mapping layer in the gating module to obtain the linearly mapped global token. The global token after linear mapping is processed based on the GELU activation function to obtain the first global token after activation. Based on the fourth unbiased linear mapping layer, the global token after the first activation process is linearly mapped to obtain the reprocessed global token. The reprocessed global token is processed based on the Sigmoid activation function to obtain a scalar gate; The prediction module 470 can predict the sequence of endogenous variables at a future preset time based on the first spatial mapping result and the second spatial mapping result in the following manner: The difference between 1 and the scalar gate is calculated to obtain the subtracted scalar gate. The first spatial mapping result, the scalar gate, the second spatial mapping result, and the scalar gate after the difference processing are weighted and summed to predict the endogenous variable sequence at a future preset time.

[0142] In yet another exemplary embodiment of the present invention, the abnormal operating condition prediction model is trained in the following manner: Obtain multiple sets of historical process variable sequences and corresponding real labels, wherein the real labels are used to characterize the historical endogenous variable sequences corresponding to the historical process variable sequences at a future preset time. A training dataset is constructed based on the historical process variable sequence and the real labels corresponding to the historical process variable sequence; Construct a target training function, wherein the target training function is determined based on the relative relationship between the true labels and alarm limits in the training dataset; Based on the training dataset and the target training function, the working condition anomaly prediction model is trained in an end-to-end training manner to obtain a trained working condition anomaly prediction model.

[0143] In yet another exemplary embodiment of the present invention, the target training function is determined in the following manner: Determine alarm mode items, including high alarm mode items, low alarm mode items, and dual-sided alarm mode items where high alarm mode items and low alarm mode items coexist; Based on the alarm mode item, the real label, and the prediction result obtained by inputting the training data in the training dataset into the abnormal working condition prediction model, the target training function is constructed.

[0144] In yet another exemplary embodiment of the present invention, the alarm mode item is determined in the following manner: Based on the relative relationship between the real labels and the high alarm limit, the high alarm mode item is determined, wherein the high alarm limit is used to characterize the alarm limit at which an alarm reminder is triggered when the endogenous variable sequence exceeds the alarm limit at a preset time in the future; Based on the relative relationship between the real label and the low alarm limit, the low alarm mode item is determined, wherein the low alarm limit is used to characterize the alarm limit at which a warning reminder is triggered when the endogenous variable sequence falls below the alarm limit at a preset time in the future; The dual-sided alarm mode item is determined based on the high alarm mode item and the low alarm mode item.

[0145] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute an industrial production process abnormal condition early warning method. This method includes: acquiring a sequence of process variables for the industrial production process to be predicted; dividing the sequence of process variables into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified process variable sequence; and the exogenous variable sequence is a process variable sequence other than the endogenous variable sequence; constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for representing the endogenous variable sequence; and constructing a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence. Spatial projection; based on the first spatial projection and the second spatial projection, the first spatial projection is updated to obtain the updated first spatial projection; based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained respectively, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in the latent space of the preset dimension; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the latent space of the preset dimension; based on the first spatial mapping result and the second spatial mapping result, the endogenous variable sequence at a preset future time is predicted; based on the endogenous variable sequence at the preset future time and the alarm limit, an early warning is given for the abnormal operating conditions of the industrial production process to be predicted.

[0146] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0147] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the industrial production process abnormal condition early warning method provided by the above methods. The method includes: acquiring a process variable sequence of the industrial production process to be predicted; dividing the process variable sequence into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified process variable sequence in the process variable sequence; the exogenous variable sequence is a process variable sequence other than the endogenous variable sequence in the process variable sequence; and constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a full-space projection for characterizing the endogenous variable sequence. The system employs a token-based approach; based on the exogenous variable sequence, constructs a second spatial projection corresponding to the exogenous variable sequence; updates the first spatial projection based on the first and second spatial projections to obtain an updated first spatial projection; based on the updated first spatial projection, obtains a first spatial mapping result and a second spatial mapping result, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in a preset dimension latent space; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the preset dimension latent space; based on the first and second spatial mapping results, predicts the endogenous variable sequence at a preset future time; and based on the endogenous variable sequence at the preset future time and an alarm limit, provides an early warning for abnormal operating conditions of the industrial production process to be predicted.

[0148] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the industrial production process abnormal condition early warning method provided by the above methods. The method includes: acquiring a sequence of process variables for the industrial production process to be predicted; dividing the sequence of process variables into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is a user-specified process variable sequence within the process variable sequence; and the exogenous variable sequence is a process variable sequence other than the endogenous variable sequence within the process variable sequence; constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for characterizing the endogenous variable sequence; and based on the exogenous variable sequence... A second spatial projection corresponding to the exogenous variable sequence is constructed; based on the first spatial projection and the second spatial projection, the first spatial projection is updated to obtain an updated first spatial projection; based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained respectively, wherein the first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in a preset dimension latent space; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in a preset dimension latent space; based on the first spatial mapping result and the second spatial mapping result, the endogenous variable sequence at a preset future time is predicted; based on the endogenous variable sequence at the preset future time and the alarm limit, an early warning is given for abnormal operating conditions of the industrial production process to be predicted.

[0149] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0150] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of abnormal operating conditions in an industrial production process, characterized in that, The method includes: Obtain the sequence of process variables for the industrial production process to be predicted; The process variable sequence is divided into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is the process variable sequence specified by the user in the process variable sequence; and the exogenous variable sequence is the process variable sequence other than the endogenous variable sequence in the process variable sequence. Based on the endogenous variable sequence, a first spatial projection corresponding to the endogenous variable sequence is constructed, wherein the first spatial projection includes a global token for characterizing the endogenous variable sequence; Based on the exogenous variable sequence, a second spatial projection corresponding to the exogenous variable sequence is constructed; Based on the first spatial projection and the second spatial projection, the first spatial projection is updated to obtain the updated first spatial projection. Based on the updated first spatial projection, a first spatial mapping result and a second spatial mapping result are obtained respectively. The first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in the latent space of the preset dimension; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the latent space of the preset dimension. Based on the first spatial mapping result and the second spatial mapping result, the sequence of endogenous variables at a future preset time is predicted; Based on the endogenous variable sequence and alarm limit at a predetermined future time, an early warning is issued for abnormal operating conditions in the industrial production process to be predicted.

2. The method for early warning of abnormal operating conditions in industrial production processes according to claim 1, characterized in that, The step of constructing a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence includes: The endogenous variable sequence is divided into multiple blocks according to a preset time block length, wherein the blocks are used to represent the endogenous variable subsequences obtained by dividing the endogenous variable sequence according to the preset time block length; For any given tile, a sine and cosine position code is superimposed on the tile to obtain a superimposed tile, and the superimposed tile is projected onto a representation space of a preset dimension through an unbiased linear mapping to obtain a tile space projection corresponding to the tile. According to the time sequence, the multiple tile spatial projections are sorted to obtain a tile spatial projection sequence; By concatenating the global token at the end of the tile spatial projection sequence, a first spatial projection corresponding to the endogenous variable sequence is constructed.

3. The method for early warning of abnormal operating conditions in industrial production processes according to claim 2, characterized in that, The global token is obtained in the following way: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a global token generation module, and the global token generation module includes a first unbiased linear mapping layer. The endogenous variable sequence is input into the working condition anomaly prediction model, and linear mapping is performed through the first unbiased linear mapping layer in the global token generation module to obtain a global token used to characterize the endogenous variable sequence.

4. The method for early warning of abnormal operating conditions in industrial production processes according to claim 1, characterized in that, The step of constructing a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence includes: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes an exogenous variable inverted variable embedding module, and the exogenous variable inverted variable embedding module includes a second unbiased linear mapping layer; The exogenous variable sequence is input into the working condition anomaly prediction model, and linear mapping is performed through the second unbiased linear mapping layer in the exogenous variable inversion embedding module to obtain the second spatial projection corresponding to the exogenous variable sequence.

5. The method for early warning of abnormal operating conditions in industrial production processes according to any one of claims 1 to 4, characterized in that, After obtaining the sequence of process variables for the industrial production process to be predicted, the method further includes: Obtain multiple sets of historical process variable sequences; Based on the historical process variable sequence, the mean and standard deviation of each process variable are determined; For any of the process variable sequences, the process variable sequences are normalized based on the mean and standard deviation to obtain normalized process variable sequences. The normalized process variable sequence is used as the process variable sequence.

6. The method for early warning of abnormal operating conditions in industrial production processes according to claim 1, characterized in that, The step of updating the first spatial projection based on the first spatial projection and the second spatial projection to obtain the updated first spatial projection includes: A pre-trained abnormal working condition prediction model is invoked. This model includes a decoupled encoder module, which is composed of multiple stacked encoders. The initial updated first spatial projection obtained from the output of the current encoder layer is used as the first spatial projection input to the next encoder layer, and so on, until the output of the last encoder layer is the initial updated first spatial projection. This initial updated first spatial projection from the last encoder layer is then used as the updated first spatial projection. The encoder for each layer performs the following process: Perform multi-head self-attention processing on the first spatial projection to obtain the self-attention processed first spatial projection; After stitching the first spatial projection and the self-attention processed first spatial projection together, and then performing layer normalization processing, the aggregated first spatial projection is obtained. Using the global token in the first spatial projection after aggregation as the query and the second spatial projection as the key and value, cross-attention processing is performed to obtain the cross-attention processing result. After concatenating the global token in the first spatial projection after aggregation and the cross-attention processing result, and then performing layer normalization processing, the global token after cross-variable processing is obtained. The global token after cross-variable processing is backfilled into the aggregated first spatial projection to obtain the first spatial projection after encoder update of this layer. A nonlinear transformation is performed on the updated first spatial projection of the encoder at this layer to obtain the transformed first spatial projection. After stitching together the first spatial projection updated by the encoder of this layer and the first spatial projection after transformation, and then performing layer normalization processing, the initial updated first spatial projection is obtained.

7. The method for early warning of abnormal operating conditions in industrial production processes according to claim 1, characterized in that, The process of obtaining the first spatial mapping result and the second spatial mapping result based on the updated first spatial projection includes: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a horizontal expert prediction module and a trend expert prediction module; wherein the horizontal expert prediction module includes a first multilayer perceptron; and the trend expert prediction module includes a second multilayer perceptron. The updated first spatial projection is input into the working condition anomaly prediction model, and the updated first spatial projection is spatially mapped by the horizontal expert prediction module to obtain the first spatial mapping result. The updated first spatial projection is divided according to a preset time block length to obtain multiple projection blocks, wherein the projection blocks are used to represent the first spatial projection subsequence obtained by dividing the updated first spatial projection according to the preset time block length. The first-order difference processing of adjacent projected tiles in the time dimension is performed to obtain the change tensor between projected tiles. The tensor of changes between the projected tiles is input into the abnormal operating condition prediction model, and the tensor of changes between the projected tiles is spatially mapped by the trend expert prediction module to obtain a second spatial mapping result.

8. The method for early warning of abnormal operating conditions in industrial production processes according to claim 1, characterized in that, Before predicting the endogenous variable sequence at a future preset time based on the first spatial mapping result and the second spatial mapping result, the method further includes: The pre-trained abnormal working condition prediction model is invoked, wherein the abnormal working condition prediction model includes a gating module, and the gating module includes a third unbiased linear mapping layer, a fourth unbiased linear mapping layer, a GELU activation function, and a Sigmoid activation function; The updated global token in the first spatial projection is input into the working condition anomaly prediction model, and linear mapping is performed through the third unbiased linear mapping layer in the gating module to obtain the linearly mapped global token. The global token after linear mapping is processed based on the GELU activation function to obtain the first global token after activation. Based on the fourth unbiased linear mapping layer, the global token after the first activation process is linearly mapped to obtain the reprocessed global token. The reprocessed global token is processed based on the Sigmoid activation function to obtain a scalar gate; The prediction of the endogenous variable sequence at a future preset time based on the first spatial mapping result and the second spatial mapping result includes: The difference between 1 and the scalar gate is calculated to obtain the subtracted scalar gate. The first spatial mapping result, the scalar gate, the second spatial mapping result, and the scalar gate after the difference processing are weighted and summed to predict the endogenous variable sequence at a future preset time.

9. The method for early warning of abnormal operating conditions in industrial production processes according to any one of claims 3, 4, 6, 7, and 8, characterized in that, The abnormal operating condition prediction model was trained using the following method: Obtain multiple sets of historical process variable sequences and corresponding real labels, wherein the real labels are used to characterize the historical endogenous variable sequences corresponding to the historical process variable sequences at a future preset time. A training dataset is constructed based on the historical process variable sequence and the real labels corresponding to the historical process variable sequence; Construct a target training function, wherein the target training function is determined based on the relative relationship between the true labels and alarm limits in the training dataset; Based on the training dataset and the target training function, the working condition anomaly prediction model is trained in an end-to-end training manner to obtain a trained working condition anomaly prediction model.

10. The method for early warning of abnormal operating conditions in industrial production processes according to claim 9, characterized in that, The target training function is determined in the following manner: Determine alarm mode items, including high alarm mode items, low alarm mode items, and dual-sided alarm mode items where high alarm mode items and low alarm mode items coexist; Based on the alarm mode item, the real label, and the prediction result obtained by inputting the training data in the training dataset into the abnormal working condition prediction model, the target training function is constructed.

11. The method for early warning of abnormal operating conditions in industrial production processes according to claim 10, characterized in that, The alarm mode item is determined in the following way: Based on the relative relationship between the real labels and the high alarm limit, the high alarm mode item is determined, wherein the high alarm limit is used to characterize the alarm limit at which an alarm reminder is triggered when the endogenous variable sequence exceeds the alarm limit in a future preset time. Based on the relative relationship between the real label and the low alarm limit, the low alarm mode item is determined, wherein the low alarm limit is used to characterize the alarm limit at which a warning reminder is triggered when the endogenous variable sequence falls below the alarm limit at a preset time in the future. The dual-sided alarm mode item is determined based on the high alarm mode item and the low alarm mode item.

12. An early warning device for abnormal operating conditions in an industrial production process, characterized in that, The device includes: The acquisition module is used to acquire the sequence of process variables for the industrial production process to be predicted; A partitioning module is used to divide the process variable sequence into an endogenous variable sequence and an exogenous variable sequence, wherein the endogenous variable sequence is the process variable sequence specified by the user in the process variable sequence; and the exogenous variable sequence is the process variable sequence other than the endogenous variable sequence in the process variable sequence. A first construction module is configured to construct a first spatial projection corresponding to the endogenous variable sequence based on the endogenous variable sequence, wherein the first spatial projection includes a global token for characterizing the endogenous variable sequence; The second construction module is used to construct a second spatial projection corresponding to the exogenous variable sequence based on the exogenous variable sequence; The update module is used to update the first spatial projection based on the first spatial projection and the second spatial projection to obtain the updated first spatial projection; The mapping module is used to obtain a first spatial mapping result and a second spatial mapping result based on the updated first spatial projection. The first spatial mapping result is used to characterize the absolute position information of the endogenous variable sequence in the latent space of the preset dimension; the second spatial mapping result is used to characterize the trend change information of the endogenous variable sequence in the latent space of the preset dimension. The prediction module is used to predict the sequence of endogenous variables at a future preset time based on the first spatial mapping result and the second spatial mapping result; The early warning module is used to provide early warning of abnormal operating conditions in the industrial production process to be predicted based on the endogenous variable sequence and alarm limit at a preset future time.

Images