An industrial artificial intelligence-based device early warning method and electronic device

Abstract

The embodiment of the application discloses a kind of equipment early warning method and electronic equipment based on industrial artificial intelligence, it is related to artificial intelligence technical field, the method comprises: the working condition data sequence and exogenous data sequence of industrial equipment in current period are input into the time series causal prediction model trained in advance, time series causal prediction model includes fusion module, causal time delay representation module and prediction module;Working condition data sequence and exogenous data sequence are fused using fusion module, obtain working condition exogenous joint matrix;Causal weight and time delay probability coupling processing is carried out on working condition exogenous joint matrix using causal time delay representation module, obtain causal time delay mask matrix;Characteristic joint decoding regression processing is carried out on working condition exogenous joint matrix and causal time delay mask matrix using prediction module, obtain the prediction sequence of target parameter;The early warning strategy of industrial equipment is determined according to the prediction sequence of target parameter, and industrial equipment is controlled based on early warning strategy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a device early warning method and electronic device based on industrial artificial intelligence. Background Technology

[0002] With the continuous improvement of industrial intelligence and automation, various industrial equipment are widely used in manufacturing, energy and chemical industries, metallurgy and heavy industry, and other fields, forming the core foundation for ensuring continuous and stable industrial production. Industrial equipment operates under complex environments of high load, variable operating conditions, and multiple external interferences, making it prone to hidden degradation problems such as component wear, performance degradation, and deviations in operating parameters. If potential equipment hazards cannot be identified in advance, abnormal evolution trends cannot be predicted, and timely measures cannot be taken, unplanned downtime and production line interruptions can easily occur, potentially leading to safety accidents and huge economic losses. Therefore, equipment early warning technology with high precision and forward-looking predictive capabilities is crucial for ensuring stable industrial production, reducing operation and maintenance costs, and mitigating safety risks.

[0003] To overcome the limitations of traditional threshold alarms and simple statistical analysis, artificial intelligence (AI) technology has become a crucial support for improving early warning accuracy. With the deep integration of general artificial intelligence technology and industrial operating mechanisms, the field of industrial AI technology has emerged. It aims to utilize AI technology to achieve state perception, trend prediction, and intelligent decision-making for industrial equipment. Although this field is developing rapidly, most current mainstream equipment early warning methods rely solely on time-series analysis of operating data, making it difficult to fully reconstruct the internal physical operating mechanisms of the equipment. This results in often one-sided and inaccurate early warning conclusions, failing to meet the urgent need for refined and differentiated early warnings in complex industrial scenarios.

[0004] Therefore, there is an urgent need to propose a new method to solve the above problems. Summary of the Invention

[0005] This invention provides a device early warning method and electronic device based on industrial artificial intelligence, which can deeply integrate operating condition data with exogenous data, thereby effectively improving the accuracy and reliability of early warning and meeting the refined early warning needs of complex industrial scenarios.

[0006] In a first aspect, embodiments of the present invention provide a device early warning method based on industrial artificial intelligence, the method comprising: The operating condition data sequence and exogenous data sequence of industrial equipment in the current period are input into a pre-trained time-series causal prediction model, which includes a fusion module, a causal delay representation module and a prediction module. The fusion module is used to fuse the operating condition data sequence and the exogenous data sequence to obtain the operating condition exogenous joint matrix; The causal delay characterization module is used to couple the causal weights and delay probabilities of the exogenous joint matrix of the operating conditions to obtain the causal delay mask matrix. The prediction module is used to perform feature joint decoding and regression processing on the exogenous joint matrix of the operating conditions and the causal delay mask matrix to obtain the prediction sequence of the target parameters; The early warning strategy for the industrial equipment is determined based on the predicted sequence of the target parameters, and the industrial equipment is controlled based on the early warning strategy.

[0007] The technical solution of this invention first inputs the operating condition data sequence and exogenous data sequence of industrial equipment within the current period into a pre-trained time-series causal prediction model, providing a data foundation for determining the prediction sequence of the target parameters. Next, a fusion module is used to fuse the operating condition data sequence and exogenous data sequence to obtain an operating condition-exogenous joint matrix. This simplifies the downstream data processing flow and reduces the overall system complexity. Furthermore, it allows the model to synchronously perceive changes in equipment operating conditions and the external environment, thereby more completely reconstructing the equipment's operating mechanism and effectively adapting to the refined early warning needs of complex industrial scenarios, thus improving prediction accuracy. Subsequently, a causal delay representation module is used to couple causal weights and delay probabilities into the operating condition-exogenous joint matrix to obtain a causal delay mask matrix. This mask matrix can eliminate false correlations and invalid features based on the dual laws of causal strength and time delay, filtering out weak causal variables to reduce computational load and system computing power, while simultaneously improving the credibility of the time-series causal representation, thereby improving the prediction accuracy of the target parameters. Then, the prediction module performs feature joint decoding and regression processing on the exogenous joint matrix of operating conditions and the causal delay mask matrix to obtain the predicted sequence of target parameters. This generates a high-precision, high-efficiency, strong generalization target parameter prediction sequence with causal interpretability, providing accurate quantitative basis for subsequent determination of early warning strategies, thereby improving the timeliness and reliability of equipment anomaly early warning judgment. Finally, the early warning strategy for industrial equipment is determined based on the predicted sequence of target parameters. Controlling industrial equipment based on the early warning strategy can identify potential equipment anomalies in advance and achieve differentiated automatic control in conjunction with the early warning strategy, thereby ensuring the stability of continuous production while avoiding serious equipment failures. The entire process forms an integrated automated control closed loop of "prediction → early warning → control", significantly improving the operational stability and intelligent operation and maintenance management level of industrial equipment. Therefore, the technical solution of this invention can solve the problems of insufficient early warning accuracy and difficulty in adapting to the refined early warning needs of complex industrial scenarios in the prior art.

[0008] Secondly, embodiments of the present invention also provide an equipment early warning device based on industrial artificial intelligence, the device comprising: The input module is used to input the operating condition data sequence and exogenous data sequence of industrial equipment in the current period into the pre-trained time-series causal prediction model. The time-series causal prediction model includes a fusion module, a causal delay representation module and a prediction module. A multi-source fusion module is used to fuse the operating condition data sequence and the exogenous data sequence to obtain an operating condition exogenous joint matrix. The encoding module is used to perform causal weight and delay probability coupling processing on the exogenous joint matrix of the working conditions using the causal delay characterization module to obtain a causal delay mask matrix. The decoding module is used to perform feature joint decoding and regression processing on the exogenous joint matrix of the operating conditions and the causal delay mask matrix using the prediction module to obtain the prediction sequence of the target parameters; The early warning module is used to determine the early warning strategy of the industrial equipment based on the predicted sequence of the target parameters, and to control the industrial equipment based on the early warning strategy.

[0009] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the equipment early warning method based on industrial artificial intelligence as described in any embodiment of the present invention.

[0010] Fourthly, embodiments of the present invention also provide a storage medium containing computer-executable instructions, characterized in that the computer-executable instructions, when executed by a computer processor, implement the equipment early warning method based on industrial artificial intelligence as described in any embodiment of the present invention.

[0011] It should be noted that the aforementioned computer instructions may be stored, in whole or in part, on a computer-readable storage medium. This computer-readable storage medium may be packaged together with the processor of the industrial AI-based equipment early warning device, or it may be packaged separately from the processor; this application does not impose any limitations on this.

[0012] The descriptions of the second, third, and fourth aspects in this application can be referenced to the detailed description of the first aspect; and the beneficial effects described in the second, third, and fourth aspects can be referenced to the analysis of the beneficial effects of the first aspect, which will not be repeated here.

[0013] In this application, the name of the aforementioned equipment early warning device based on industrial artificial intelligence does not limit the equipment or functional module itself. In actual implementation, these devices or functional modules may appear under other names. As long as the function of each device or functional module is similar to that of this application, it falls within the scope of the claims of this application and its equivalents.

[0014] These or other aspects of this application will become more readily apparent in the following description. Attached Figure Description

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

[0016] Figure 1 A flowchart illustrating a device early warning method based on industrial artificial intelligence provided in an embodiment of the present invention; Figure 2 A flowchart illustrating another equipment early warning method based on industrial artificial intelligence provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of an equipment early warning device based on industrial artificial intelligence provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0018] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0019] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0020] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0021] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subroutine, etc. Moreover, without conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0022] It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0023] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0024] Figure 1 This is a flowchart illustrating a device early warning method based on industrial artificial intelligence provided in an embodiment of the present invention. This embodiment is applicable to various industrial scenarios where intelligent early warning of the health status of industrial equipment is carried out based on industrial AI technology. The method is executed by a device early warning device based on industrial artificial intelligence, which can be implemented in software and / or hardware. For example, the device can be integrated into an electronic device. The device early warning method based on industrial artificial intelligence in this embodiment specifically includes the following steps: Step 110: Input the operating condition data sequence and exogenous data sequence of industrial equipment in the current period into the pre-trained time-series causal prediction model.

[0025] Specifically, the current period is a specific time window used for data collection and evaluation during the operation of industrial equipment (such as the past 5 minutes, the last 10 sampling points, etc.). In this embodiment, only the data collected by the equipment within this time period is used for this early warning. Industrial equipment refers to production machinery in industrial production scenarios that has continuous operation and is equipped with multiple types of sensors, such as injection molding machines, motors, kilns, rolling mills, air compressors, and energy storage battery packs. It is both the source of operating condition data collection and the object of control command execution. The operating condition data sequence is a multi-dimensional time-series dataset arranged in chronological order, used to characterize the operating status of the industrial equipment itself, including intrinsic parameters of the equipment such as temperature, speed, voltage, current, vibration amplitude, load, oil pressure, and output. It only reflects the operating characteristics of the equipment itself and does not include various external related variables. The exogenous data sequence consists of external time-series driving variables that are related to the operation of the equipment but do not belong to the equipment's own state, such as ambient temperature and humidity, workshop load, power grid voltage fluctuations, raw material feeding rate, ambient dust concentration, outdoor temperature, and output of upstream and downstream equipment. The time-series causal prediction model is a prediction model trained based on historical operating condition data sequences (such as temperature, speed, voltage, etc.), historical exogenous data sequences (such as ambient temperature, grid voltage, etc.), and historical target parameter sequences (such as output pressure, remaining life, cavity pressure, etc.) that correspond one-to-one with the above data. The time-series causal prediction model includes a fusion module, a causal delay representation module, and a prediction module.

[0026] In practice, data acquisition gateways (such as factory data platforms, edge acquisition terminals, intelligent data acquisition instruments, and acquisition terminals based on monitoring and data acquisition systems) can be used to acquire the operating condition data sequence and exogenous data sequence of industrial equipment within the current period. These sequences are then input into a pre-trained time-series causal prediction model. Since the time-series causal prediction model is trained using historical operating condition data sequences, corresponding historical exogenous data sequences, and historical true target parameter sequences that match these data, the model can output a predicted sequence of target parameters in real time based on the current input data.

[0027] In this embodiment, the above steps provide a data foundation for determining the prediction sequence of the target parameters.

[0028] Step 120: Use the fusion module to fuse the operating condition data sequence and the exogenous data sequence to obtain the operating condition exogenous joint matrix.

[0029] Specifically, the fusion module is used to perform fusion processing on two types of heterogeneous data—operational data and exogenous data—that differ in dimensions and physical meaning, and output a unified structured matrix. The joint operational data and exogenous data matrix is ​​a two-dimensional data structure generated by the fusion module (e.g., rows represent time steps and columns represent feature variables), used to provide a unified input format for subsequent causal analysis and prediction.

[0030] In the specific implementation, normalization or standardization algorithms (such as Z-score standardization) are first used to standardize the operating condition data sequence and the exogenous data sequence respectively, so as to unify the numerical distribution range of each feature. After the standardization process is completed, a two-dimensional matrix is ​​constructed with time step as the row dimension (arranged in ascending time order) and feature variables as the column dimension. The standardized operating condition data is used as the first column and the standardized exogenous data is used as the second column, and they are horizontally concatenated to obtain a single two-dimensional structured matrix, namely the operating condition exogenous joint matrix. This matrix is ​​then output to the downstream causal delay characterization module as the input data basis for the module to perform causal weight calculation and delay probability analysis.

[0031] In this embodiment, the above steps can simplify the downstream data processing flow and reduce the overall system complexity. On the other hand, the model can synchronously perceive changes in equipment operating conditions and the external environment, thereby more completely restoring the equipment operating mechanism, effectively adapting to the refined early warning needs in complex industrial scenarios, and improving prediction accuracy.

[0032] Step 130: Use the causal delay characterization module to couple the causal weights and delay probabilities of the exogenous joint matrix of the operating conditions to obtain the causal delay mask matrix.

[0033] Specifically, the causal delay characterization module is an intermediate feature module for mining two core relationships between variables: the first is the strength of the causal influence between variables, i.e., the causal association weight; the second is the time lag required for the effect of one variable to be transmitted to another, i.e., the delay probability distribution. The coupling processing of causal weights and delay probabilities involves a joint calculation process that synchronously binds the causal strength of variables and the lag time distribution. The causal delay mask matrix is ​​a selective gating structure used to eliminate invalid association information such as weak causal strength and low delay confidence, while retaining feature connections that have a genuine causal driving effect on the prediction target.

[0034] In the specific implementation, the pairwise combinations of all feature variables in the exogenous joint matrix of working conditions are traversed, and the following operations are performed on each group of variables in sequence: (I) Causal weight quantification: First, a baseline regression model is constructed using the lagged terms of the dependent variable itself, and the baseline residual sum of squares is solved; then, the lagged terms of the independent variable are introduced to construct a complete regression model, and the complete residual sum of squares is solved. The difference between the two groups of residual sums of squares is compared and the F statistic is calculated. Then, the maximum F statistic of all variable pairs is used as the benchmark to perform Min-Max normalization, and the result is mapped to the 0-1 interval to obtain the causal weight corresponding to the variable pair. (II) Solving the time delay probability distribution: The maximum time delay step size L is preset, and effective causal correlation variable pairs with causal weights higher than the preset initial screening threshold are selected. For each pair of valid variables, the independent variable is sequentially lagged by k steps (k=1, 2, ..., L) and aligned with the dependent variable (sample points that cannot be matched at the beginning and end of the sequence due to lag are removed during alignment); the association score between the independent and dependent variables (such as Pearson correlation coefficient or mutual information) is calculated for each lag step; the association scores for all lag steps are Softmax normalized to obtain the time delay probability corresponding to each time delay step, thereby generating the time delay probability distribution vector of the variable pair. (III) Coupling operation of causal weight and time delay probability: the causal weight of the same variable pair is coupled with the time delay probability distribution vector, and the causal weight is used as a scaling factor to multiply the time delay probability distribution vector element by element to output the coupling coefficient vector. This coupling coefficient can simultaneously characterize the strength of causal influence and time delay law between variables; for variable pairs with weak causal effect and no stable time delay matching law, the coupling coefficients corresponding to each time delay step are close to 0, thereby realizing the automatic filtering of invalid variable associations. (iv) Binarization of Coupling Coefficients to Generate a Mask: A preset fixed threshold is used to perform binary discrimination processing on the coupling coefficient vector; if the coupling coefficient at a certain time delay step is greater than the threshold, the corresponding position in the causal delay mask matrix is ​​assigned a value of 1, indicating that the feature has an effective causal effect at the corresponding time delay and is allowed to participate in subsequent prediction calculations; if the coupling coefficient is less than or equal to the threshold, the corresponding position is assigned a value of 0, masking the invalid feature association. After traversing all variable pairs, a causal delay mask matrix matching the feature dimension of the exogenous joint matrix of the operating conditions is generated.

[0035] In this embodiment, through the above steps, false associations and invalid features can be eliminated based on the dual laws of causal strength and time delay, weak causal variables can be screened out to reduce the amount of computation and the system's computing power overhead, while improving the credibility of time-series causal representation, thereby improving the prediction accuracy of target parameters.

[0036] Step 140: Use the prediction module to perform feature joint decoding and regression processing on the exogenous joint matrix of operating conditions and the causal delay mask matrix to obtain the prediction sequence of the target parameters.

[0037] Specifically, the prediction module jointly processes the exogenous joint matrix of operating conditions and the causal delay mask matrix, outputting continuous prediction results of equipment target parameters for multiple future time steps. The feature joint decoding regression processing fuses and decodes the feature data in the exogenous joint matrix of operating conditions with the causal constraint information of the causal delay mask matrix, outputting the calculation process for future multi-step prediction values. Target parameters are core indicators requiring early warning or control, and are continuous parameters directly reflecting the health status of the equipment, such as: bearing remaining life, peak exhaust pressure, energy consumption, reaction conversion rate, and extreme values ​​of cavity pressure. The prediction sequence of target parameters is an array of equally time-interval values ​​output by the time-series causal prediction model, representing the predicted values ​​of key monitoring parameters for the equipment at multiple future (subsequent) time steps.

[0038] In the specific implementation, the causal delay mask matrix is ​​used as a feature selection gating to perform weighted filtering of features in the exogenous joint matrix of operating conditions. Specifically, for positions with a mask matrix value of 1, the corresponding feature path is retained, allowing it to participate normally in the attention weight calculation; for positions with a mask value of 0, the corresponding feature is directly masked, and the attention weight corresponding to that feature is forcibly set to zero. The features after mask weighting are input into a temporal feature extraction network (such as a temporal convolutional network, a bidirectional long short-term memory network, a Transformer encoder, etc.) to obtain deep causal temporal features. Then, the deep causal temporal features are input into a decoder network (such as an autoregressive decoder, a Transformer decoder, or a fully connected decoding layer, etc.) to obtain a decoded feature matrix aligned with the dimension of the prediction target. Finally, the decoded feature matrix is ​​input into a regression head composed of a fully connected network, and the predicted sequence of target parameters is calculated through regression mapping.

[0039] In this embodiment, the above steps can generate a target parameter prediction sequence that is highly accurate, efficient, generalizable, and causally interpretable, providing a precise quantitative basis for determining the early warning strategy and thus improving the timeliness and reliability of equipment anomaly early warning judgment.

[0040] Step 150: Determine the early warning strategy for industrial equipment based on the predicted sequence of target parameters, and control the industrial equipment based on the early warning strategy.

[0041] Specifically, the early warning strategy combines the numerical change trend of the target parameter prediction sequence with preset early warning rules to comprehensively determine the future operational risk level of the equipment and the corresponding equipment operation plan.

[0042] In practice, quantitative indicators for risk assessment (such as instantaneous parameter values, time-series change rate (slope), duration of parameter exceeding limits, peak (valley) value of the prediction sequence, etc.) can be extracted from the target parameter prediction sequence first; the extracted indicators are then matched with preset early warning rules to obtain the early warning strategy for industrial equipment. For example, the early warning rules can be as follows: 1. If all parameters in the predicted sequence fall within the normal range and the parameter change rate does not exceed the preset rate, the equipment risk level is determined to be 0, and the corresponding equipment operation is to maintain the original control strategy of the equipment unchanged; 2. If there are parameters in the predicted sequence that exceed the first-level early warning threshold but do not reach the fault threshold, and the continuous over-limit time reaches the preset time, the equipment risk level is determined to be 1, and the corresponding equipment operation is to slightly adjust the equipment auxiliary parameters (such as cooling flow adjustment, feed rate fine-tuning, speed correction, etc.), with the adjustment range controlled within 1%-5%, and at the same time push the reminder to the operation and maintenance personnel; 3. If there are parameters in the predicted sequence that exceed the fault threshold, or the parameter change rate is greater than the preset rate, the equipment risk level is determined to be 2, and the corresponding equipment operation is to perform protective shutdown of the equipment or automatically switch to the standby equipment, while triggering an audible and visual alarm and pushing the emergency alarm information to the staff's terminal. It should be noted that when multiple risk judgment conditions are triggered at the same time and correspond to different risk levels, the principle of "highest risk level" is followed, and the highest risk level is taken as the benchmark, and the equipment operation plan corresponding to that level is output.

[0043] Upon receiving the early warning strategy, standardized equipment control commands (such as speed setpoint update commands, start / stop control commands, etc.) are automatically generated based on the corresponding equipment operation plan. These commands are then sent to the equipment controller (such as the frequency converter) via industrial fieldbus or industrial Ethernet, enabling automatic control based on the early warning strategy. Simultaneously with issuing and executing the automatic control commands, the complete early warning strategy (such as risk level, judgment criteria, recommended operations, and predictive trend charts) is pushed to the operator's terminal, achieving synchronized human-machine alarm. Operators can view the early warning details on the terminal and manually intervene or confirm the execution status of control commands when necessary.

[0044] In this embodiment, the above steps enable early identification of potential equipment malfunctions and, in conjunction with early warning strategies, differentiated automatic control can be implemented, thereby ensuring the stability of continuous production while avoiding serious equipment failures. The entire process forms an integrated automated control loop of "prediction → early warning → control," significantly improving the operational stability and intelligent operation and maintenance management level of industrial equipment.

[0045] The equipment early warning system based on industrial artificial intelligence provided in this invention first inputs the operating condition data sequence and exogenous data sequence of industrial equipment within the current period into a pre-trained time-series causal prediction model, providing a data foundation for determining the prediction sequence of target parameters. Next, a fusion module is used to fuse the operating condition data sequence and exogenous data sequence to obtain an operating condition-exogenous joint matrix. This simplifies downstream data processing and reduces overall system complexity. Furthermore, it allows the model to synchronously perceive changes in equipment operating conditions and the external environment, thus more completely reconstructing the equipment's operating mechanism and effectively adapting to the refined early warning needs of complex industrial scenarios, improving prediction accuracy. Subsequently, a causal delay representation module couples the causal weights and delay probabilities of the operating condition-exogenous joint matrix to obtain a causal delay mask matrix. This mask matrix can eliminate false correlations and invalid features based on the dual laws of causal strength and time delay, filtering out weak causal variables to reduce computational load and system computing power, while simultaneously improving the credibility of the time-series causal representation, thereby improving the prediction accuracy of the target parameters. Then, the prediction module performs feature joint decoding and regression processing on the exogenous joint matrix of operating conditions and the causal delay mask matrix to obtain the predicted sequence of target parameters. This generates a high-precision, high-efficiency, strong generalization target parameter prediction sequence with causal interpretability, providing accurate quantitative basis for subsequent determination of early warning strategies, thereby improving the timeliness and reliability of equipment anomaly early warning judgment. Finally, the early warning strategy for industrial equipment is determined based on the predicted sequence of target parameters. Controlling industrial equipment based on the early warning strategy can identify potential equipment anomalies in advance and achieve differentiated automatic control in conjunction with the early warning strategy, thereby ensuring the stability of continuous production while avoiding serious equipment failures. The entire process forms an integrated automated control closed loop of "prediction → early warning → control", significantly improving the operational stability and intelligent operation and maintenance management level of industrial equipment. Therefore, the technical solution of this invention can solve the problems of insufficient early warning accuracy and difficulty in adapting to the refined early warning needs of complex industrial scenarios in the prior art.

[0046] Figure 2 This is a flowchart illustrating another equipment early warning method based on industrial artificial intelligence provided by an embodiment of the present invention. This embodiment is a specific embodiment based on the above embodiment. In this embodiment, the method may further include: Step 210: Input the operating condition data sequence and exogenous data sequence of industrial equipment in the current period into the pre-trained time-series causal prediction model.

[0047] Specifically, the time-series causal prediction model includes a fusion module, a causal delay representation module, and a prediction module.

[0048] Optionally, the training process of the time-series causal prediction model includes a first-stage training, a second-stage training, and a third-stage training. The loss function of the first-stage training is composed of the causal delay loss function; the loss functions of the second-stage and third-stage training are both composed of the errors of the physical loss function, the domain adaptation loss function, and the data loss function.

[0049] In its implementation, the time-series causal prediction model adopts a phased training strategy. The rules for unfreezing and freezing parameters and the design of the loss function for each phase are as follows: Phase 1 Training: To achieve independent convergence of the causal delay representation module, only this module is enabled for parameter updates in this phase, while the parameters of all other modules in the model are frozen. The loss function L1 for phase 1 training is composed of the causal delay loss function, as shown in the following formula: ;in, To reconstruct the error term, ; For the complete time series sequence of the k-th eigenvalue in the exogenous joint matrix of operating conditions, Let j be the time series sequence of the j-th characteristic variable in the exogenous joint matrix of operating conditions; This is the Koopman similarity kernel function; The element in the j-th row and k-th column of the causal relationship matrix; The square of the Frobenius norm; This is a causal delay loss function used for probabilistic modeling of the time lag step in causal propagation between device variables. This loss function incorporates both entropy regularization and maximum probability penalty constraints, ensuring that the delay distribution is concentrated at the actual physical delay point and avoiding fuzzy and scattered estimations. D is the time delay probability distribution vector corresponding to the current effective variable pair in the time delay distribution matrix; d is any element in D, representing the time delay probability value corresponding to a specific lag step. The maximum probability penalty hyperparameter is a preset fixed weight coefficient used to adjust the constraint strength of the maximum probability penalty term; The acyclic constraint function ensures that the learned causal graph is a directed acyclic graph; These are the weighting coefficients for the delay loss term; are the weighting coefficients of the acyclic constraint term, and both are pre-defined hyperparameters.

[0050] Second-stage training: To fix the causal selection logic and optimize the fitting ability of the prediction network separately, only the prediction module is allowed to participate in parameter updates in this stage, while the parameters of all other modules of the model are frozen. The loss function L2 in the second-stage training consists of the errors of the physical loss function, the domain adaptation loss function, and the data loss function, as shown in the following formula: ;in, The data loss function represents the fitting error between the predicted output sequence and the actual observation labels of the equipment, constraining the prediction results to fit the actual operating data. This refers to the physical loss weighting coefficient; The physical loss function is used, and the constraint derivation process conforms to the inherent physical and dynamic characteristics of the equipment. For the domain adaptation loss weight coefficients; This is the domain adaptation loss function, used to reduce the feature distribution shift; ; This represents the total number of samples in the source domain. This represents the total number of samples in the target domain. It is the i-th effective predicted feature sample in the source domain after causal masking; This is the j-th effective predicted feature sample in the target domain; For kernel mapping; For the regenerating nucleus Hilbert space; Let be the square norm in Hilbert space. ; This is the theoretical physical quantity at time i+1 obtained by recursion based on the mechanism equation h(); For the network to predict features effectively The predicted physical quantity at time i+1 is derived from the deduction output; It is the square of the L2 norm. The mechanistic equation is: ; denoted as , where is the core target physical quantity to be predicted for the equipment; ... This is the set of learnable physical parameters within the mechanistic dynamics function.

[0051] The third stage of training: In order to fine-tune the model globally, all module parameters are unfrozen in this stage, and the overall joint iterative optimization of all module parameters of the model is carried out. The loss function L3 used in the third stage of training reuses the loss combination form of the second stage, and the two calculation expressions are consistent, that is, L3=L2.

[0052] In this embodiment, a three-stage progressive training process is set up, with each stage matched with a differentiated loss function to achieve step-by-step optimization: the first stage independently learns the temporal causal relationship of variables using causal delay loss to construct a reliable feature selection logic; the second stage fixes the parameters of the causal module and uses a composite loss function that integrates data fitting, physical mechanism constraints, and domain adaptation to optimize the prediction baseline separately; the third stage reuses the composite loss function to perform global joint fine-tuning of the entire network. The above-mentioned hierarchical training strategy decomposes the multi-task optimization objective into multiple stages to be solved sequentially, effectively reducing training difficulty and improving convergence stability; at the same time, the loss function is precisely matched with the training objective of each stage, each performing its own function, and the modular decoupled training avoids the risk of causal logic being destroyed by the subsequent prediction network, while the unified loss reuse design also simplifies model implementation and reduces engineering deployment costs.

[0053] Step 211: Use the fusion module to fuse the operating condition data sequence and the exogenous data sequence to obtain the operating condition exogenous joint matrix.

[0054] Further, step 211 may specifically include: performing time-series decomposition on the operating condition data sequence to obtain a seasonal term matrix; performing time-domain reconstruction on the seasonal term matrix to obtain a time-domain reconstructed feature matrix; reducing the dimensionality of the time-domain reconstructed feature matrix to obtain a low-dimensional operating condition latent feature matrix; and fusing the low-dimensional operating condition latent feature matrix and the exogenous data sequence to obtain an operating condition exogenous joint matrix.

[0055] Specifically, the seasonal component matrix is ​​a two-dimensional matrix composed of the periodic components extracted after time-series decomposition of the operating condition data sequence. The time-domain reconstructed feature matrix is ​​the matrix obtained after time-domain reconstruction of the seasonal component matrix. The low-dimensional operating condition latent feature matrix is ​​the compressed feature matrix obtained after dimensionality reduction of the time-domain reconstructed feature matrix.

[0056] In practice, the seasonal-trend decomposition method based on local weighted regression, the X-13ARIMA-SEATS seasonal adjustment procedure, or the wavelet transform decomposition method are used to perform time series decomposition on the time series data of each feature variable in the working condition data sequence containing M (M≥1) feature variables. Each feature variable is decomposed into three components: trend term, seasonal term, and residual term. The seasonal components of all feature variables are extracted and concatenated to obtain the seasonal term matrix.

[0057] Then, orthogonal basis functions are used to perform time-domain expansion and fitting reconstruction of the seasonal term matrix variable by variable. The univariate fitting expression is as follows: ;in, Let be the i-th order orthogonal basis functions (such as Fourier basis or wavelet basis). is the learnable fitting coefficient of the i-th order basis function corresponding to the m-th feature variable, and p is the order of the basis function; For time-series reconstruction of residual terms; This represents the original value of the m-th characteristic variable in the seasonal term matrix at time step t.

[0058] Subsequently, the time series data reconstructed from all M feature variables using basis functions are concatenated according to their feature dimensions to generate a time-domain reconstructed feature matrix. Then, the singular value decomposition algorithm is used to reduce the dimensionality of the time-domain reconstructed feature matrix to obtain a low-dimensional latent feature matrix.

[0059] Finally, the low-dimensional latent feature matrix of the operating conditions and the exogenous data sequence are fused to obtain the joint exogenous matrix of the operating conditions. Specifically, the two types of features are first mapped to a unified dimensional space through three sets of learnable projection matrices. The mapping formula is as follows: , , ;in, , , There are three sets of trainable projection matrices; Let E be the low-dimensional latent feature matrix; let E be the matrix composed of exogenous data sequences; and let Q, K, and V be the query vector, key vector, and value vector, respectively. Then, using the query vector as a constraint, cross-type feature attention weighted fusion is performed, calculated using the following formula: ; where Z is the exogenous joint matrix of operating conditions. This is the scaling factor.

[0060] In this embodiment, the above steps can not only remove trends and noise interference to highlight the periodic time-series characteristics of the equipment, but also align the correlation between operating conditions and external environmental characteristics through the attention mechanism, reduce the differences in different units and data distributions, thereby improving the accuracy of subsequent causal weight calculation and time-series prediction, while significantly reducing the computational cost of model training and edge inference.

[0061] Optionally, the causal delay characterization module includes a causal feature extraction unit, a delay feature extraction unit, and a mask fusion unit.

[0062] Step 212: Use the causal feature extraction unit to perform sparse causal solution processing on the exogenous joint matrix of operating conditions to obtain the causal correlation matrix, and use the time delay feature extraction unit to perform multi-scale time delay modeling processing on the exogenous joint matrix of operating conditions to obtain the time delay distribution matrix.

[0063] Specifically, the causal feature extraction unit, the first subunit in the causal time delay representation module, is used to perform sparse causal structure learning on the input exogenous joint matrix of operating conditions. It solves for the causal directions between variables and quantifies the influence strength, outputting a causal correlation matrix. The time delay feature extraction unit, the second subunit in the causal time delay representation module, processes the input exogenous joint matrix of operating conditions and outputs a time delay distribution matrix. The causal correlation matrix is ​​a square matrix used to quantify the strength of causal interactions between each pair of feature variables in the exogenous joint matrix of operating conditions. The time delay distribution matrix is ​​a probabilistic structure tensor used to characterize the hysteresis response patterns of each feature variable at different time scales.

[0064] In the specific implementation, firstly, the causal feature extraction unit performs sparse causal solution processing on the input exogenous joint matrix of operating conditions to obtain the causal correlation matrix. Specifically, it iterates through all pairwise combinations of feature variables in the matrix. For each pair of variables, it constructs a benchmark regression model containing only the autoregressive term of the dependent variable and a complete regression model that incorporates lagged terms of the independent variable. The F-statistic is calculated by comparing the ratio of the sum of squared residuals of the two models to quantify the causal effect of the independent variable on the dependent variable. The F-statistics of all variable pairs are then Min-Max normalized to map the causal strength to the 0-1 interval. Simultaneously, acyclic sparse constraints are introduced to eliminate spurious and circular causal relationships, ultimately constructing a causal correlation matrix to characterize the strength of causality among the variables.

[0065] After solving the causal correlation matrix, the time delay feature extraction unit performs multi-scale time delay modeling on the exogenous joint matrix under the same operating condition to obtain the time delay distribution matrix. Specifically, based on a preset initial screening threshold, effective causal variable pairs with high causal weights are selected, while variable combinations with weak causal effects are eliminated. For the remaining effective variable pairs, within a preset maximum time delay step size, the independent variables are sequentially shifted and aligned at different time lengths, and invalid sample points due to time misalignment are eliminated. By calculating the correlation scores between the independent and dependent variables at each lag step size, the response intensity at different time delay scales is characterized. Then, the correlation scores for all time delay steps are Softmax normalized to obtain the action probability corresponding to each time delay step size. Finally, the time delay probability distributions of all effective variable pairs are aggregated to generate a time delay distribution matrix containing multi-scale time lag patterns.

[0066] In this embodiment, the above steps can combine the dual dimensions of causal strength and time delay to filter invalid features, improve the reliability of temporal causal representation, reduce the amount of computation and enhance the model's generalization ability to various industrial conditions, and provide reliable causal temporal feature support for subsequent masking and high-precision temporal prediction.

[0067] Furthermore, the causal feature extraction unit is used to perform sparse causal solution processing on the exogenous joint matrix of operating conditions to obtain the causal correlation matrix, including: performing frequency domain information enhancement processing on the exogenous joint matrix of operating conditions to obtain the time-frequency joint feature matrix; performing similarity measurement on the time-frequency joint features to generate a similarity kernel matrix; and performing sparse causal mapping processing on the similarity kernel matrix to obtain the causal correlation matrix.

[0068] Specifically, the time-frequency joint feature matrix is ​​obtained by enhancing the exogenous joint matrix of operating conditions with frequency domain information. The similarity metric is used to quantify the degree of cooperative change or pattern similarity between any two feature variables in the time-frequency joint features. The similarity kernel matrix is ​​a two-dimensional matrix generated based on the time-frequency joint features using a similarity kernel function; its elements correspond to the similarity scores of the two sets of feature variables, accurately representing the degree of dynamic correlation between variables. The sparse causal mapping process, based on the similarity kernel matrix and combined with sparsity constraints and causal filtering mechanisms, eliminates false and redundant associations, transforming feature similarity relationships into weights for true causal relationships through a mapping solution.

[0069] In practice, the exogenous joint matrix of operating conditions can first be enhanced with frequency domain information using Fourier transform (such as fast Fourier transform, inverse fast Fourier transform, etc.) to obtain the time-frequency joint feature matrix. Then, for any two sets of feature variables in the time-frequency joint feature matrix, their nonlinear similarity in the time-series evolution space is calculated. The specific calculation formula is as follows: ; in, for and The nonlinear similarity, i.e., the Koopman similarity kernel function value; Let j be the time series sequence of the j-th characteristic variable in the exogenous joint matrix of operating conditions; This is the time series sequence of the k-th characteristic variable in the exogenous joint matrix of operating conditions; for Time index; for Time index; T is Total number of time steps in the sequence; for Total number of time steps in the sequence; Forward attenuation coefficient; This is the backward attenuation coefficient; It is a Gaussian kernel.

[0070] Iterate through all variable pairs and construct a similarity kernel matrix from the nonlinear similarities of all pairs. Finally, input the similarity kernel matrix into the causal mapping network layer, and solve it through learnable weight mapping and sparse acyclic constraints to obtain the causal association matrix. The training objective of the causal mapping network layer is defined by the following formula: ; Where min is the optimization objective, W is the optimization variable, and the optimal solution that minimizes the loss function is found; W is the causal correlation matrix; The complete time series sequence of the kth characteristic variable in the exogenous joint matrix of operating conditions (as the dependent variable / explained); The element in the j-th row and k-th column of the causal relationship matrix; These are the weighting coefficients for L1 regularization, controlling the strength of sparsity. It is an L1 norm; is the weighting coefficient for entropy regularization; H(W) is the information entropy of the causal correlation matrix.

[0071] The training objective is related to the first-stage loss function of the overall temporal causal prediction model. Simultaneous optimization is performed to solve the problem under acyclic sparse constraints. At the same time, the diagonal of the matrix is ​​forced to be set to zero to eliminate self-causality, and the final output is the causal correlation matrix.

[0072] In this embodiment, frequency domain enhancement is used to fuse time-frequency features, and bidirectional time-decayed Koopman kernels are used to accurately quantify the nonlinear time-series similarity of variables. Then, sparse causal mapping with multiple constraints is used to purify the true causal relationship. Furthermore, the causal solution and the loss function of the first stage of the model are jointly optimized. While eliminating false causal relationships and improving causal interpretability, invalid features are compressed, reducing the computational load of subsequent time delay modeling and time series prediction, thereby improving the overall prediction accuracy of device parameters.

[0073] Furthermore, the time delay feature extraction unit is used to perform multi-scale time delay modeling on the exogenous joint matrix of the operating conditions to obtain the time delay distribution matrix. This includes: extracting local features from the exogenous joint matrix of the operating conditions to obtain local time delay features, and extracting global features from the exogenous joint matrix of the operating conditions to obtain global time delay features; fusing the local time delay features and global time delay features to obtain fused time delay features; and performing probability normalization mapping on the fused time delay features to obtain the time delay distribution matrix.

[0074] Specifically, global time delay features reflect the lag response patterns between variables over long time scales. Local time delay features reflect the lag response patterns between variables over short time scales. The fused time delay features are a comprehensive feature representation obtained by fusing local and global time delay features.

[0075] In practice, a local feature extraction network (such as CNN) is used to extract local features from the exogenous joint matrix of operating conditions to obtain the local time delay feature D. local Furthermore, a global feature extraction network (such as Mamba) is used to extract global features from the exogenous joint matrix of operating conditions, thereby obtaining the global time delay feature D. global The local and global time delay features are concatenated and fused along the feature dimension to obtain the fused time delay feature. Then, a Softmax probability normalization mapping is applied to the fused time delay feature, transforming the response intensity of each variable at each lag step into a probability distribution that sums to 1, resulting in the time delay distribution matrix D. r In summary, the specific formula for calculating the time delay distribution matrix is ​​as follows: .

[0076] In this embodiment, by fusing local and global delay features, the multi-scale lag patterns of short-term abrupt changes and long-term steady-state conditions of the device are fully covered. After probability normalization, a standardized delay probability distribution matrix is ​​obtained, which enriches the dimensions of delay representation, reduces the bias of single-scale modeling, and improves the completeness and accuracy of delay distribution. This provides accurate and reliable delay prior features for causal delay mask generation and time series prediction.

[0077] Step 213: Use the mask fusion unit to perform mask fusion mapping on the causal correlation matrix and the time delay distribution matrix to obtain the causal time delay mask matrix.

[0078] Specifically, the mask fusion unit is the third sub-unit in the causal delay characterization module, which is used to perform coupling operations and mapping on the causal correlation matrix and the delay distribution matrix to generate a causal delay mask matrix.

[0079] In the specific implementation, firstly, the causal weight scalars of each variable pair in the causal correlation matrix are extracted and multiplied element-wise along the time delay dimension by the multi-scale time delay probability vectors of the corresponding variable pairs in the time delay distribution matrix to obtain a coupling coefficient matrix that fuses causal strength and time delay probability. Then, a preset effective threshold is used to binarize the coupling coefficient matrix: if the coupling coefficient at a certain position is greater than the threshold, the corresponding mask is set to 1, indicating that the variable pair has an effective causal time delay correlation at that lag step, allowing the downstream prediction network to use this feature; otherwise, it is set to 0, blocking the invalid feature path. After traversing all variable pairs and all time delay scales, a binary 0-1 matrix, i.e., the causal time delay mask matrix, is output.

[0080] In this embodiment, the above steps can reduce redundant calculations in the downstream model and reduce computing power consumption, and can also rely on causal and time-delay dual priors to accurately screen effective features, thereby improving the accuracy of time series prediction and the physical interpretability of the model.

[0081] Further, step 213 may specifically include: performing Hadamard product operation on the causal correlation matrix and the time delay distribution matrix to obtain the causal time delay fusion feature matrix; and performing nonlinear activation mapping processing on the causal time delay fusion feature matrix to obtain the causal time delay mask matrix.

[0082] Specifically, the causal-delay fusion feature matrix is ​​the matrix obtained by multiplying the causal correlation matrix and the delay distribution matrix by Hadamard.

[0083] In practice, element-wise dot product (i.e., Hadamard product) can be performed on the causal correlation matrix and the time delay distribution matrix to obtain the causal-time delay fusion feature matrix; then, the Sigmoid activation function is used. The causal delay fusion feature matrix is ​​subjected to nonlinear activation mapping to map the values ​​to the 0-1 interval. Then, a binary partition is performed using a preset threshold to generate a causal delay mask matrix B containing only 0 and 1 values. The corresponding calculation formula for this process is as follows: .

[0084] In this embodiment, the above steps effectively improve the subsequent prediction accuracy while improving feature quality and reducing computational overhead.

[0085] Step 214: Multiply the exogenous joint matrix of operating conditions and the causal delay mask matrix element by element to obtain the effective prediction features.

[0086] Specifically, the effective prediction features are the features obtained by filtering the exogenous joint matrix of operating conditions through a causal delay mask matrix.

[0087] In practice, the exogenous joint matrix of operating conditions and the causal delay mask matrix B can be multiplied element-wise to obtain effective prediction features. The corresponding calculation formula is: .

[0088] In this embodiment, the above steps can effectively filter out invalid correlation features with low confidence, reduce noise interference, thereby improving prediction accuracy and ultimately enhancing early warning precision.

[0089] Step 215: Use the linear dynamic operator to perform time-series extrapolation on the effective prediction features to obtain the initial prediction sequence of the target parameters.

[0090] Specifically, the linear dynamic operator is a time-series deduction operator based on a linear state-space model or linear recursive relations, used to characterize the evolution of device states over time. Time-series deduction is the process of iteratively calculating the state at multiple future time steps from the current known state using established state transition relationships. In this embodiment, the linear dynamic operator receives valid prediction features, deduces forward along the time axis, and generates initial predicted values ​​for the target parameter at multiple future time steps. The initial prediction sequence is a continuous numerical sequence output by the linear dynamic operator after time-series deduction, representing the preliminary prediction results of the target parameter at multiple future time steps.

[0091] In practice, effective prediction features can be substituted into the linear dynamic operator to complete multi-step time series deduction, and the initial prediction sequence of the target parameters can be obtained by solving the problem. The specific calculation formula is as follows: ; Where y(t) represents the initial predicted value of the target parameter at time t, and the predicted values ​​at all times form a complete initial prediction sequence; n a a is the autoregressive order; i Here, y(ti) represents the autoregressive weighting coefficients, and y(ti) represents the historical true value or previously predicted value of the target parameter at time t-1; n b b is the lag order of the input feature. j Let x(tj) be the feature weight coefficient, and x(tj) be the effective predicted feature value at time tj.

[0092] In this embodiment, the above steps can construct a complementary prediction architecture of "linear skeleton + nonlinear details", which improves the training efficiency and stability of the model while ensuring prediction accuracy.

[0093] Step 216: Optimize the initial prediction sequence of the target parameters to obtain the prediction sequence of the target parameters.

[0094] Specifically, in this embodiment, the optimization of this step is to perform residual correction on the initial prediction sequence.

[0095] In practice, linear or nonlinear compensation methods are used to optimize the initial prediction sequence of the target parameters, resulting in a predicted sequence of the target parameters. Specifically, in the linear compensation method, a fully connected layer is used to linearly map the initial prediction sequence, and the calculation formula is as follows: Among them, W fc b is the learnable weight matrix for the fully connected layer. fc is the learnable bias term of the fully connected layer; Y is the predicted value of the target parameter after linear compensation. In the nonlinear compensation method, a nonlinear activation function is superimposed on the linear mapping to construct a multi-layer nonlinear mapping network, in order to capture the nonlinear residual bias that is difficult for linear models to fit under complex working conditions, and achieve higher accuracy residual correction.

[0096] In practical applications, any of the above compensation methods can be flexibly selected according to the linearity and accuracy requirements of the equipment, so as to balance computational efficiency and prediction accuracy.

[0097] In this embodiment, the above steps can effectively improve the prediction accuracy of the target parameter prediction sequence, thereby enhancing the reliability of the early warning.

[0098] Step 217: Determine the early warning strategy for industrial equipment based on the predicted sequence of target parameters.

[0099] Further, step 217 may specifically include: determining whether the predicted sequence of the target parameter is within the safe threshold range corresponding to the target parameter; if the predicted sequence of the target parameter is not within the safe threshold range corresponding to the target parameter, then determining the warning level of the industrial equipment based on the number of parameters exceeding the limit in the predicted sequence of the target parameter, and querying the preset warning strategy mapping table based on the warning level to obtain the warning strategy of the industrial equipment.

[0100] Specifically, the safety threshold range is a pre-set numerical boundary used to determine whether the predicted value of the target parameter is within the safe operating range of the equipment, based on actual conditions or needs. For example, the safety threshold range for bearing temperature is [0℃, 85℃], the safety threshold range for remaining life is [100 hours, ∞), and the safety threshold range for energy consumption is [0 kW, 500 kW]. The number of parameters exceeding the limit is the number of sample points in the target parameter prediction sequence whose predicted value exceeds the preset safety threshold range. The warning level is a classification of the severity of equipment anomalies based on the number of parameters exceeding the limit; the more parameters, the more serious the deviation from the safe range, and the higher the corresponding warning level. The preset warning strategy mapping table is a one-to-one correspondence table of "warning level → equipment control action / operation plan" pre-established based on actual conditions or needs, used to quickly convert abstract warning levels into specific equipment control commands.

[0101] In practice, the predicted values ​​at each moment in the prediction sequence of the target parameter are first evaluated to verify whether all predicted values ​​fall within the pre-configured safety threshold range for the target parameter. If all predicted values ​​are within the safety threshold range, the warning level is set to 0, and the corresponding warning strategy is to maintain the current control strategy unchanged. If some predicted values ​​exceed the safety threshold range, meaning the entire prediction sequence is not within the safe operating range, the number of parameters exceeding the limit at each exceeding moment in the prediction sequence is counted. Based on the number of parameters exceeding the limit, the warning level corresponding to the industrial equipment is determined in a pre-defined grading rule base (such as a table showing the correspondence between grading threshold conditions and warning levels). For example, the grading rule might be: if 1 ≤ number of parameters exceeding the limit ≤ 3, the warning level is 1; if 4 ≤ number of parameters exceeding the limit ≤ 8, the warning level is 2; if 9 ≤ number of parameters exceeding the limit, the warning level is 3. The warning level is an integer and does not have a decimal point.

[0102] Subsequently, using the defined warning levels as the retrieval index, a matching query is performed in the pre-built warning strategy mapping table to obtain the warning strategy adapted to the current industrial equipment. For example, when the warning level is 1, the warning strategy is to send a reminder notification to the operation and maintenance personnel and not to automatically adjust the equipment; when the warning level is 2, the warning strategy is to automatically perform auxiliary parameter fine-tuning (such as increasing the cooling flow by 3% and reducing the speed by 2%); when the warning level is 3, the warning strategy is to automatically perform protective load reduction or shutdown and switch to standby equipment.

[0103] In this embodiment, the above steps can realize advance prediction of equipment failure risks and hierarchical differentiated management, which can reduce false alarms, improve the objectivity of early warning judgment, and quickly output standardized handling solutions, thus balancing equipment operation safety and production efficiency.

[0104] The equipment early warning system based on industrial artificial intelligence provided in this invention first inputs the operating condition data sequence and exogenous data sequence of industrial equipment within the current period into a pre-trained time-series causal prediction model, providing a data foundation for determining the prediction sequence of target parameters. Next, a fusion module is used to fuse the operating condition data sequence and exogenous data sequence to obtain an operating condition-exogenous joint matrix. This simplifies downstream data processing and reduces the overall system complexity. Furthermore, it allows the model to simultaneously perceive changes in equipment operating conditions and the external environment, thus more completely reconstructing the equipment's operating mechanism and effectively adapting to the refined early warning needs of complex industrial scenarios, improving prediction accuracy. Subsequently, a causal feature extraction unit performs sparse causal solution processing on the operating condition-exogenous joint matrix to obtain a causal correlation matrix. Then, a time delay feature extraction unit performs multi-scale time delay modeling processing on the operating condition-exogenous joint matrix to obtain a time delay distribution matrix. This combines causal strength and time delay dimensions to filter invalid features, improving the reliability of time-series causal representation, reducing computational load, and enhancing the model's generalization ability to variable industrial operating conditions. This provides reliable causal time-series feature support for subsequent masking and high-precision time-series prediction. By employing a masking fusion unit to perform masking fusion mapping on the causal correlation matrix and the time delay distribution matrix, a causal time delay mask matrix is ​​obtained. This reduces redundant computations in the downstream model, lowers computational cost, and allows for precise selection of effective features based on both causal and time delay priors, thereby improving the accuracy of time-series predictions and the physical interpretability of the model. Element-wise multiplication of the exogenous joint matrix of operating conditions and the causal time delay mask matrix yields effective prediction features, effectively filtering out low-confidence invalid correlation features, reducing noise interference, and thus improving prediction accuracy and ultimately enhancing early warning precision. Using a linear dynamic operator to perform time-series extrapolation on the effective prediction features yields the initial prediction sequence of the target parameters. This allows for the construction of a complementary prediction architecture of "linear skeleton + nonlinear details," improving model training efficiency and stability while maintaining prediction accuracy. Optimizing the initial prediction sequence of the target parameters yields the final prediction sequence of the target parameters, effectively improving the prediction accuracy of the target parameter prediction sequence and thus enhancing the reliability of early warnings. Finally, an early warning strategy for industrial equipment is determined based on the predicted sequence of target parameters. Controlling the industrial equipment based on this early warning strategy enables the early identification of potential anomalies and facilitates differentiated automatic control, thereby ensuring the stability of continuous production while avoiding serious equipment failures. This entire process forms an integrated automated control loop of "prediction → early warning → control," significantly improving the operational stability and intelligent operation and maintenance management level of industrial equipment. Therefore, the technical solution of this invention can solve the problems of insufficient early warning accuracy and difficulty in adapting to the refined early warning needs of complex industrial scenarios in existing technologies.

[0105] Figure 3This is a schematic diagram of the structure of an equipment early warning device based on industrial artificial intelligence provided in an embodiment of the present invention. This device and the equipment early warning method based on industrial artificial intelligence in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the equipment early warning device based on industrial artificial intelligence, please refer to the embodiments of the equipment early warning method based on industrial artificial intelligence.

[0106] like Figure 3 As shown, the device includes: The input module 310 is used to input the operating condition data sequence and exogenous data sequence of industrial equipment in the current period into a pre-trained time-series causal prediction model. The time-series causal prediction model includes a fusion module, a causal delay representation module, and a prediction module. The multi-source fusion module 320 is used to fuse the operating condition data sequence and the exogenous data sequence using the fusion module to obtain the operating condition exogenous joint matrix; Encoding module 330 is used to perform causal weight and delay probability coupling processing on the exogenous joint matrix of the working conditions using the causal delay characterization module to obtain a causal delay mask matrix; The decoding module 340 is used to perform feature joint decoding and regression processing on the exogenous joint matrix of the operating conditions and the causal delay mask matrix using the prediction module to obtain the prediction sequence of the target parameters; The early warning module 350 is used to determine the early warning strategy of the industrial equipment based on the predicted sequence of the target parameters, and to control the industrial equipment based on the early warning strategy.

[0107] Based on the above embodiments, the multi-source fusion module 320 is specifically used for: The operating condition data sequence is decomposed into a time series to obtain a seasonal term matrix; the seasonal term matrix is ​​reconstructed in the time domain to obtain a time domain reconstructed feature matrix; the time domain reconstructed feature matrix is ​​reduced in dimensionality to obtain a low-dimensional operating condition latent feature matrix; the low-dimensional operating condition latent feature matrix and the exogenous data sequence are fused to obtain the operating condition exogenous joint matrix.

[0108] Based on the above embodiments, the causal delay characterization module includes a causal feature extraction unit, a delay feature extraction unit, and a mask fusion unit. The encoding module 330 is specifically used for: The causal feature extraction unit is used to perform sparse causal solution processing on the exogenous joint matrix of the operating conditions to obtain a causal correlation matrix. The time delay feature extraction unit is used to perform multi-scale time delay modeling processing on the exogenous joint matrix of the operating conditions to obtain a time delay distribution matrix. The mask fusion unit is used to perform mask fusion mapping processing on the causal correlation matrix and the time delay distribution matrix to obtain the causal time delay mask matrix.

[0109] Based on the above embodiments, the encoding module 330 uses the causal feature extraction unit to perform sparse causal solution processing on the exogenous joint matrix of the working conditions to obtain a causal correlation matrix, including: The exogenous joint matrix of the operating conditions is subjected to frequency domain information enhancement processing to obtain a time-frequency joint feature matrix; the time-frequency joint features are subjected to similarity measurement to generate a similarity kernel matrix; the similarity kernel matrix is ​​subjected to sparse causal mapping processing to obtain the causal correlation matrix.

[0110] Based on the above embodiments, the encoding module 330 uses the time delay feature extraction unit to perform multi-scale time delay modeling on the exogenous joint matrix of the operating conditions to obtain a time delay distribution matrix, including: Local features are extracted from the exogenous joint matrix of the operating conditions to obtain local time delay features, and global features are extracted from the exogenous joint matrix of the operating conditions to obtain global time delay features. The local time delay features and the global time delay features are fused to obtain fused time delay features. The fused time delay features are then subjected to probability normalization mapping to obtain the time delay distribution matrix.

[0111] Based on the above embodiments, the encoding module 330 uses the mask fusion unit to perform mask fusion mapping processing on the causal correlation matrix and the time delay distribution matrix to obtain the causal time delay mask matrix, including: The causal correlation matrix and the time delay distribution matrix are subjected to Hadamard product operation to obtain the causal time delay fusion feature matrix; the causal time delay fusion feature matrix is ​​subjected to nonlinear activation mapping processing to obtain the causal time delay mask matrix.

[0112] Based on the above embodiments, the decoding module 340 is specifically used for: The exogenous joint matrix of the operating conditions and the causal delay mask matrix are multiplied element-wise to obtain effective prediction features; the effective prediction features are then subjected to time-series deduction using a linear dynamic operator to obtain an initial prediction sequence of the target parameters; the initial prediction sequence of the target parameters is then optimized to obtain a prediction sequence of the target parameters.

[0113] Based on the above embodiments, the training process of the time-series causal prediction model includes a first-stage training, a second-stage training, and a third-stage training. The loss function of the first-stage training is composed of a causal delay loss function. The loss functions of the second-stage training and the third-stage training are both composed of the error of the physical loss function, the domain adaptation loss function, and the data loss function.

[0114] Based on the above embodiments, the early warning module 350 determines the early warning strategy for the industrial equipment according to the predicted sequence of the target parameters, including: Determine whether the predicted sequence of the target parameter is within the safe threshold range corresponding to the target parameter; if the predicted sequence of the target parameter is not within the safe threshold range corresponding to the target parameter, determine the warning level of the industrial equipment based on the number of parameters exceeding the limit in the predicted sequence of the target parameter, and query the preset warning strategy mapping table based on the warning level to obtain the warning strategy of the industrial equipment.

[0115] The equipment early warning device based on industrial artificial intelligence provided in the embodiments of the present invention can execute the equipment early warning method based on industrial artificial intelligence provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0116] It is worth noting that in the above embodiments of the equipment early warning device based on industrial artificial intelligence, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0117] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Figure 4 A block diagram of an exemplary electronic device 4 suitable for implementing embodiments of the present invention is shown. Figure 4 The electronic device 4 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0118] like Figure 4 As shown, electronic device 4 is represented in the form of a general-purpose computing electronic device. The components of electronic device 4 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0119] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0120] Electronic device 4 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by electronic device 4, including volatile and non-volatile media, removable and non-removable media.

[0121] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Electronic device 4 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (… Figure 4 Not shown; usually referred to as a "hard drive"). Although Figure 4 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0122] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0123] Electronic device 4 can also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), and with one or more devices that enable a user to interact with electronic device 4, and / or with any device that enables electronic device 4 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed through input / output (I / O) interface 22. Furthermore, electronic device 4 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. Figure 4 As shown, network adapter 20 communicates with other modules of electronic device 4 via bus 18. It should be understood that, although... Figure 4 Not shown, it can be combined with electronic device 4 to use other hardware and / or software modules, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0124] The processing unit 16 executes various functional applications and page displays by running programs stored in the system memory 28, such as implementing the equipment early warning method based on industrial artificial intelligence provided in the embodiments of the present invention.

[0125] Of course, those skilled in the art will understand that the processor can also implement the technical solution of the equipment early warning method based on industrial artificial intelligence provided in any embodiment of the present invention.

[0126] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements, for example, the equipment early warning method based on industrial artificial intelligence provided in this invention.

[0127] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0128] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0129] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0130] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0131] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computing device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0132] Furthermore, the acquisition, storage, use, and processing of data in the technical solution of this invention all comply with relevant laws and regulations.

[0133] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A device early warning method based on industrial artificial intelligence, characterized in that, The method includes: The operating condition data sequence and exogenous data sequence of industrial equipment in the current period are input into a pre-trained time-series causal prediction model, which includes a fusion module, a causal delay representation module and a prediction module. The fusion module is used to fuse the operating condition data sequence and the exogenous data sequence to obtain the operating condition exogenous joint matrix; The causal delay characterization module is used to couple the causal weights and delay probabilities of the exogenous joint matrix of the operating conditions to obtain the causal delay mask matrix. The prediction module is used to perform feature joint decoding and regression processing on the exogenous joint matrix of the operating conditions and the causal delay mask matrix to obtain the prediction sequence of the target parameters; The early warning strategy for the industrial equipment is determined based on the predicted sequence of the target parameters, and the industrial equipment is controlled based on the early warning strategy.

2. The method according to claim 1, characterized in that, The fusion module is used to fuse the operating condition data sequence and the exogenous data sequence to obtain an operating condition exogenous joint matrix, including: The time-series decomposition of the operating condition data sequence yields a seasonal term matrix; The seasonal term matrix is ​​reconstructed in the time domain to obtain the reconstructed time domain feature matrix; The time-domain reconstructed feature matrix is ​​reduced in dimensionality to obtain a low-dimensional latent feature matrix for operating conditions. The low-dimensional latent feature matrix of the operating conditions and the exogenous data sequence are fused to obtain the joint matrix of exogenous operating conditions.

3. The method according to claim 1, characterized in that, The causal delay characterization module includes a causal feature extraction unit, a delay feature extraction unit, and a mask fusion unit. The causal delay characterization module performs causal weight and delay probability coupling processing on the exogenous joint matrix of the operating conditions to obtain a causal delay mask matrix, including: The causal feature extraction unit is used to perform sparse causal solution processing on the exogenous joint matrix of the working conditions to obtain the causal correlation matrix, and the time delay feature extraction unit is used to perform multi-scale time delay modeling processing on the exogenous joint matrix of the working conditions to obtain the time delay distribution matrix. The causal correlation matrix and the time delay distribution matrix are processed by the mask fusion unit to obtain the causal time delay mask matrix.

4. The method according to claim 3, characterized in that, The causal feature extraction unit is used to perform sparse causal solution processing on the exogenous joint matrix of the working conditions to obtain a causal correlation matrix, including: The exogenous joint matrix of the operating conditions is subjected to frequency domain information enhancement processing to obtain the time-frequency joint feature matrix; The time-frequency joint features are subjected to similarity measurement to generate a similarity kernel matrix; The sparse causal mapping process is applied to the similarity kernel matrix to obtain the causal association matrix.

5. The method according to claim 3, characterized in that, The time delay feature extraction unit is used to perform multi-scale time delay modeling on the exogenous joint matrix of the operating conditions to obtain a time delay distribution matrix, including: Local feature extraction is performed on the exogenous joint matrix of the operating conditions to obtain local time delay features, and global feature extraction is performed on the exogenous joint matrix of the operating conditions to obtain global time delay features; The local delay features and the global delay features are fused to obtain the fused delay features; The fused delay features are subjected to probability normalization mapping to obtain the delay distribution matrix.

6. The method according to claim 3, characterized in that, The causal correlation matrix and the time delay distribution matrix are subjected to mask fusion mapping processing by the mask fusion unit to obtain the causal time delay mask matrix, including: Perform a Hadamard product operation on the causal correlation matrix and the time delay distribution matrix to obtain a causal time delay fusion feature matrix; The causal delay fusion feature matrix is ​​subjected to nonlinear activation mapping processing to obtain the causal delay mask matrix.

7. The method according to claim 1, characterized in that, The prediction module performs feature joint decoding and regression processing on the exogenous joint matrix of the operating conditions and the causal delay mask matrix to obtain a predicted sequence of target parameters, including: The effective prediction features are obtained by multiplying the exogenous joint matrix of the operating conditions and the causal delay mask matrix element by element. The effective prediction features are subjected to time-series deduction using linear dynamic operators to obtain the initial prediction sequence of the target parameters; The initial prediction sequence of the target parameters is optimized to obtain the prediction sequence of the target parameters.

8. The method according to claim 1, characterized in that, The training process of the time-series causal prediction model includes a first-stage training, a second-stage training, and a third-stage training. The loss function of the first-stage training is composed of a causal delay loss function. The loss functions of the second-stage training and the third-stage training are both composed of the errors of the physical loss function, the domain adaptation loss function, and the data loss function.

9. The method according to claim 1, characterized in that, Determining the early warning strategy for the industrial equipment based on the predicted sequence of the target parameters includes: Determine whether the predicted sequences of the target parameters are all within the safe threshold range corresponding to the target parameters; If the predicted sequence of the target parameter is not uniformly within the safe threshold range corresponding to the target parameter, the warning level of the industrial equipment is determined based on the number of parameters exceeding the limit in the predicted sequence of the target parameter, and the warning strategy of the industrial equipment is obtained by querying the preset warning strategy mapping table based on the warning level.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the equipment early warning method based on industrial artificial intelligence as described in any one of claims 1-9.

Images