Device comprehensive efficiency multi-dimensional prediction method based on time sequence hybrid sensing network

Abstract

The invention relates to the technical field of neural network application, and particularly discloses an equipment comprehensive efficiency multi-dimensional prediction method based on a time sequence hybrid sensing network, which comprises the following steps: firstly obtaining OEE historical sequence data, and preprocessing the OEE historical sequence data to form original OEE sequence data; performing time sequence decomposition and change point identification on the original OEE sequence data by adopting an STL decomposition method; segmenting the original OEE sequence data according to the identified change points, and dividing each segment of data into a training set and a test set in proportion; constructing and training a time sequence hybrid sensing network THPN; and inputting OEE data acquired in real time into the trained time sequence hybrid sensing network THPN, performing multi-dimensional fusion prediction, and finally performing OEE sensitivity analysis. According to the multi-dimensional prediction method for the comprehensive efficiency of the equipment based on the time sequence hybrid sensing network, experimental verification on a data set of a hot rolling production line of an iron and steel enterprise shows that the method is superior to a traditional end-to-end prediction model, and has a relatively good prediction effect and a relatively good practical application value.

Description

Technical Field

[0001] This invention relates to the field of neural network application technology, and in particular to a multidimensional prediction method for the overall efficiency of devices based on temporal hybrid sensing networks. Background Technology

[0002] Production line performance evaluation is crucial in modern manufacturing, especially since Overall Equipment Effectiveness (OEE), as a multi-dimensional evaluation metric, effectively reflects availability, performance efficiency, and quality in the production process. Traditional OEE calculations assume that equipment operates under ideal conditions; however, this assumption does not take into account non-value-adding activities and complex system interactions in production.

[0003] In recent years, with the continuous changes in the production environment, OEE prediction methods have gradually attracted attention. By introducing data-driven technologies such as deep learning and machine learning, accurate OEE prediction can be achieved, and production efficiency can be improved through predictive maintenance. Existing research mainly focuses on overall OEE prediction, failing to fully consider the changes in the production environment and the interrelationships between sub-dimensions, neglecting the role of sub-dimensions, and traditional methods cannot effectively handle variable points in the data.

[0004] In view of the non-stationarity and variable point characteristics of OEE data, there is an urgent need for a multi-dimensional fusion prediction method based on a temporal hybrid sensing network model to optimize production line decisions and improve production efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a multidimensional prediction method for equipment overall efficiency (OEE) based on a temporal hybrid sensing network. It introduces TSD technology for change point detection and combines it with a THPN neural network model, improving prediction accuracy through ensemble training. Furthermore, the novel sensitivity quantification method proposed in this invention can quantify the impact of different factors on OEE, providing a scientific basis for production decisions. Experimental verification on a dataset of a hot rolling production line in a steel enterprise demonstrates that the proposed method outperforms traditional end-to-end prediction models, exhibiting better prediction performance and practical application value.

[0006] To achieve the above objectives, this invention provides a multi-dimensional prediction method for the overall efficiency of devices based on a temporal hybrid sensing network, comprising the following steps: S1. Obtain historical OEE sequence data and preprocess it to form raw OEE sequence data; S2. The STL decomposition method is used to perform temporal decomposition and change point identification on the original OEE sequence data. S3. The original OEE sequence data is segmented according to the identified change points, and each segment is divided into a training set and a test set according to the change point position. S4. Construct and train the Temporal Hybrid Sensing Network (THPN); S5. Input the real-time collected OEE data into the trained temporal hybrid sensing network THPN for multi-dimensional fusion prediction; S6. Perform OEE sensitivity analysis.

[0007] Preferably, S1 is as follows: Extract monitoring data from the production line database to obtain historical OEE data collected by the equipment, including availability. Performance efficiency and quality rate The three sub-dimensions of data were then processed using a standardized residual method to remove outliers, resulting in the original OEE sequence data.

[0008] Preferably, in S1, the standardized residual method is as follows: The OEE historical sequence data is standardized using the following formula: ; in, Indicates a time index; This represents the standardized sequence data value obtained after standardizing the historical OEE sequence data. Indicates at time The original data values ​​of the collected OEE historical sequence; This represents the mean of historical OEE data within a predetermined statistical interval. This represents the standard deviation of historical OEE series data within the stated statistical interval; when When an outlier occurs, it is identified as an anomaly and removed. Anomaly removal methods refer to using the data mean with standard deviation Construct a standardized residual space and identify outliers by assessing the degree to which the data deviates from the central interval.

[0009] Preferably, S2 is as follows: The original OEE sequence data was decomposed into trend terms using the STL decomposition method. Seasonal items With residuals ; Then, the residual difference is constructed. And twice the standard deviation of the residuals was used as the threshold. ,when The time is determined as a change point, which is used to detect nodes where the operating status of the production line changes; in, This represents the value extracted from the original OEE sequence data at time step 1 using the STL decomposition method. The corresponding residual terms; It represents the absolute difference between two adjacent time points and is used to characterize the magnitude of change of the residual term in the time dimension. A change point refers to a point where the residual term undergoes a significant abrupt change within a local segment, corresponding to equipment downtime, production rate fluctuations, or changes in the pass / fail frequency during production line operation.

[0010] Preferably, S3 is as follows: Based on the change point locations determined by S2, the original OEE sequence data is divided into multiple operating state segments. An operating state segment refers to a segment within which no new change point appears. The data in the operating state segments are then divided into training and testing sets according to the location of the change point. The training and testing sets come from different production line operating states.

[0011] Preferably, in S4, the constructed Temporal Hybrid Awareness Network (THPN) includes a dynamic feature separation module, a multi-scale dynamic dependency module, a temporal embedding module, and a fusion decoding module. During the training phase, availability is prioritized. Performance efficiency and quality rate As a multidimensional input, a loss function is used. Optimize; in, Indicates a time index; This represents the value of the loss function used during model training. Indicates the number of samples used in model training; express The actual overall equipment efficiency target value corresponding to the given moment; This represents the predicted overall device efficiency obtained through the temporal hybrid sensing network (THPN). Key hyperparameters to set include trend window, kernel size, prediction stride, hidden dimension, number of attention heads, learning rate, and maximum number of iterations; The learning rate is used to control the step size of the gradient descent parameter update. The iteration rounds represent the number of times the model has completely traversed the training set; Convolutional kernels are used to extract local temporal patterns, and their size determines the model's ability to capture changes at different scales.

[0012] Preferably, S5 is as follows: sequence availability Performance efficiency and quality rate The predicted sequences are obtained by inputting the data into the Temporal Hybrid Sensing Network (THPN). , , Then, the final Overall Equipment Effectiveness (OEE) prediction result is obtained by using the multiplicative fusion formula, which is expressed as follows: .

[0013] Preferably, S6 is as follows: First, consider availability. Performance efficiency and quality rate The sensitivity is obtained by applying perturbations to the three indicators and calculating the output difference. , , Sensitivity refers to the magnitude of change in the overall equipment efficiency (OEE) output caused by a change in a single input variable while keeping other variables constant. It is used to reflect the differences in importance of each component indicator. Then, calculate the average relative sensitivity of the overall equipment efficiency (OEE) to each sub-dimension. Relative sensitivity refers to the average relative sensitivity of the overall equipment efficiency (OEE) to each sub-dimension. , , The sensitivity of the three indicators, after normalization at the same time, is used to measure the relative contribution of the indicators to the overall equipment efficiency (OEE) prediction. The average relative sensitivity is calculated as follows: ; in, As an indicator i At any moment t The relative sensitivity, For availability at any time t Sensitivity, For performance efficiency at all times t Sensitivity, For quality rate at time t Sensitivity, indicators i Including availability Performance efficiency and quality rate .

[0014] Therefore, the present invention employs the above-mentioned multi-dimensional prediction method for the overall efficiency of devices based on temporal hybrid sensing networks, and the beneficial effects are as follows: (1) This invention introduces a temporal hybrid sensing network structure to perform multi-scale dynamic modeling of the multi-dimensional sub-indicators of Equipment Overall Efficiency (OEE) in the time domain, and establishes a segmented prediction system that can distinguish operating states by combining time series decomposition and change point identification mechanisms. Unlike traditional single-model prediction methods, this invention simultaneously introduces low-frequency and high-frequency feature separation, temporal embedding, and multi-head attention fusion mechanisms into the model structure, thereby achieving joint perception of trend changes and short-term disturbances.

[0015] (2) In addition, the present invention performs parallel fusion of different state models based on the ensemble learning strategy, and designs a sensitivity analysis unit to quantify the relative influence of each sub-dimension on the OEE change, thereby achieving a balance between interpretability and prediction accuracy.

[0016] (3) This invention has been validated on equipment operation datasets from actual production enterprises. Compared to traditional single LSTM, GRU, and convolutional time series models, it achieves significant improvements in OEE and its sub-dimension prediction tasks. The results are measured using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination. As an evaluation metric, the multidimensional fusion prediction method proposed in this invention reduces RMSE and MAE by an average of approximately 30% to 50% in the comprehensive OEE prediction. On average, the increase is about 10% or more.

[0017] (4) The present invention demonstrates higher stability and generalization ability in multi-step prediction of availability, performance efficiency and quality, and can identify equipment performance abnormalities earlier and provide interpretable analysis results of the contribution of each dimension, providing reliable data support for production scheduling and maintenance decisions.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of an embodiment of the multi-dimensional prediction method for device overall efficiency based on a time-series hybrid sensing network according to the present invention; Figure 2 This is a detailed flowchart of the overall process of an embodiment of the multi-dimensional prediction method for device comprehensive efficiency based on a time-series hybrid sensing network of the present invention. Figure 3 This is a structural framework diagram of the temporal hybrid sensing network (THPN) according to an embodiment of the multidimensional prediction method for device overall efficiency based on temporal hybrid sensing network of the present invention. Figure 4 This is an OEE prediction result diagram of the end-to-end prediction method of the multidimensional prediction method for device overall efficiency based on temporal hybrid sensing network according to an embodiment of the present invention; Figure 5 This is an OEE prediction result diagram of the multi-dimensional fusion prediction method based on the device comprehensive efficiency multi-dimensional prediction method of the present invention, wherein (a) is a comparison of availability, (b) is a comparison of performance efficiency, (c) is a comparison of quality, and (d) is a comparison of pseudo OEE. Figure 6 This is a graph showing the OEE sensitivity analysis results of an embodiment of the device overall efficiency multidimensional prediction method based on temporal hybrid sensing network of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0022] This invention is achieved through the following technical solution: This invention first identifies change points in historical OEE (Output Effectiveness) data collected by equipment through time-series decomposition in the offline phase. The data is divided into multiple subsets, and the training set is standardized and trained using a Temporal Hybrid Sensing Network (THPN). In the online phase, real-time collected OEE data is input into the trained model, and combined with the prediction results of ensemble learning, real-time monitoring of OEE changes on the production line is conducted, providing early warnings. Furthermore, this invention utilizes sensitivity analysis methods based on derivatives and relative quantities to quantify the impact of each sub-dimension of OEE on the overall OEE, providing a scientific basis for production decisions.

[0023] like Figure 1 , Figure 2 As shown, the multidimensional prediction method for device overall efficiency based on temporal hybrid sensing networks includes the following steps: S1. Obtain historical OEE sequence data and preprocess it to form raw OEE sequence data.

[0024] Based on production monitoring data from multiple batches of a company's production line, a complete modeling and prediction process was conducted, specifically as follows: Extract monitoring data from the production line database to obtain historical OEE data collected by the equipment, including availability. Performance efficiency and quality rate The three sub-dimensions of data were then processed using a standardized residual method to remove outliers, resulting in the original OEE sequence data.

[0025] The standardized residual method is as follows: The OEE historical sequence data is standardized using the following formula: ; in, Indicates a time index; This represents the standardized sequence data value obtained after standardizing the historical OEE sequence data. Indicates at time The original data values ​​of the collected OEE historical sequence; This represents the mean of historical OEE data within a predetermined statistical interval. This represents the standard deviation of the historical OEE series data within the stated statistical interval.

[0026] when When an outlier is identified, it is removed to improve the stability of subsequent modeling and prediction results.

[0027] Anomaly removal methods refer to using the data mean with standard deviation A standardized residual space is constructed, and extreme outliers are identified by assessing the degree to which the data deviates from the central interval, so as to ensure the continuity and interpretability of the input sequence.

[0028] S2. To characterize the variation features of the OEE sequence under different operating conditions, the STL decomposition method is used to perform time-series decomposition and change point identification on the original OEE sequence data, specifically: The original OEE sequence data was decomposed into trend terms using the STL decomposition method. Seasonal items With residuals .

[0029] Then, the residual difference is constructed. And twice the standard deviation of the residuals was used as the threshold. ,when The time is determined as a change point, which is used to detect nodes where the operating status of the production line changes.

[0030] in, This represents the value extracted from the original OEE sequence data at time step 1 using the STL decomposition method. The corresponding residual terms; It represents the absolute difference between two adjacent time points of the residual term, and is used to characterize the magnitude of change of the residual term in the time dimension.

[0031] A change point is a point in the residual term where a significant abrupt change occurs within a local segment. These typically correspond to critical events in the production line's operation, such as equipment downtime, production rate fluctuations, or changes in the pass / fail rate. Utilizing change points allows the overall time series to be divided into multiple segments with consistent operating states, improving the consistency of model training.

[0032] Regarding the number of variable points and production line characteristics, this embodiment detected 5 to 8 variable points in twelve years of historical data, which is basically consistent with the production rhythm, steel grade switching frequency and equipment maintenance cycle of the production line, indicating that the residual difference method has good sensitivity and stability in this scenario.

[0033] S3. Segment the original OEE sequence data according to the identified change points, and divide each segment into a training set and a test set according to the change point positions, specifically: Based on the change point positions determined by S2, the original OEE sequence data is divided into multiple operating state segments. The operating state segment data is then divided into training and test sets according to the change point positions. The training and test sets come from different production line operating states, which can effectively improve the model's generalization ability.

[0034] The operational status segment refers to the segment within which the OEE trend structure basically shows a phased pattern and no new change points appear; it can be used as input model learning for phased time series data.

[0035] S4. Construct and train the Temporal Hybrid Sensing Network (THPN).

[0036] like Figure 3 As shown, the Temporal Hybrid Sensing Network (THPN) includes a dynamic feature separation module, a multi-scale dynamic dependency module, a temporal embedding module, and a fusion decoding module, and is a multi-branch neural network. Input multidimensional time series and First, the data passes through a dynamic feature separation module, then two parallel multi-scale convolutional branches (with kernel sizes of 3, 5, and 7, combined with ReLU activation and feature fusion), before entering stacked RNN units for temporal dependency modeling. Simultaneously, sine / cosine temporal position embeddings are introduced and fused with the backbone features through a gating mechanism. The right side includes a multi-head attention mechanism (matrix multiplication, soft max, mask scaling) and a linear decoding attention module. Finally, the prediction result is output through a fully connected layer and a fusion decoder. The entire network combines multi-scale convolution, recurrent units, positional encoding, attention, and gating fusion to achieve hybrid perception and prediction of OEE temporal data.

[0037] Training phase with availability Performance efficiency and quality rate As a multidimensional input, a loss function is used. Optimize; among them, Indicates a time index; This represents the value of the loss function used during model training. Indicates the number of samples used in model training; express The actual overall equipment efficiency target value corresponding to the given moment; This represents the predicted overall device efficiency obtained through the Temporal Hybrid Sensing Network (THPN).

[0038] Key hyperparameters include a trend window of 3, kernel size of 3, 5, or 7, prediction stride of 4, hidden dimension of 32, number of attention heads of 4, and learning rate of 1×10⁻⁶. -3 The maximum number of iterations is 100.

[0039] The learning rate controls the step size for parameter updates in gradient descent. The number of iterations represents the total number of times the model traverses the training set. The convolutional kernel is used to extract local temporal patterns, and its size determines the model's ability to capture changes at different scales.

[0040] The convergence during training indicates that the model loss decreased rapidly in the early stages of training, stabilized after about 20 epochs, and converged completely around 70-90 epochs. The smooth, oscillating training curve demonstrates that the model can effectively capture the temporal patterns of OEE's multidimensional features.

[0041] S5. Input the real-time collected OEE data into the trained temporal hybrid sensing network THPN for multi-dimensional fusion prediction, specifically: sequence availability Performance efficiency and quality rate The three subsequences are input into the temporal hybrid sensing network (THPN) to obtain the predicted sequence. , , Then, the final Overall Equipment Effectiveness (OEE) prediction result is obtained by using the multiplicative fusion formula, as shown below. Figure 4 The graph shows the end-to-end single-sequence prediction results. The graph compares the actual and predicted values ​​of OEE, with the horizontal axis representing sample points and the vertical axis representing numerical values. Region A is located in the initial stage, where the actual value rises rapidly from a low point, while the predicted value is low at the edge points. Region B is located at the turning point in the middle and later stages, where the actual value drops sharply, and the predicted value shows a downward bias at the edge points, highlighting the local prediction bias of the model at sequence boundaries and state abrupt changes.

[0042] like Figure 5 The image shows the multidimensional fusion prediction effect. The multiplicative fusion formula expression is as follows: .

[0043] This embodiment first predicts the three component sequences of OEE separately and then performs multidimensional fusion, thereby avoiding error amplification when directly modeling the overall OEE, making the final prediction output more stable and robust.

[0044] S6. To measure the sensitivity of OEE to the three indicators, a sensitivity analysis is conducted, specifically as follows: First, consider availability. Performance efficiency and quality rate By applying perturbations to the three indicators and calculating the output differences, the corresponding sensitivity can be obtained. , , ,like Figure 6 The sensitivity curve is shown, where the performance efficiency is... The sensitivity is the highest; in addition, in order to further obtain the influence ratio of the three indicators in the overall prediction process, this embodiment calculates their average relative sensitivity and presents it in Table 1.

[0045] Table 1. Average relative sensitivity of subdimensions

[0046] Sensitivity refers to the magnitude of change in the overall equipment efficiency (OEE) output caused by a change in a single input variable while keeping other variables constant. It is used to reflect the differences in importance of each component indicator.

[0047] Then, calculate the average relative sensitivity of the overall equipment efficiency (OEE) to each sub-dimension. Relative sensitivity refers to the average relative sensitivity of the overall equipment efficiency (OEE) to each sub-dimension. , , The sensitivity of the three indicators, after normalization at the same time, is used to measure the relative contribution of the indicators to the overall equipment efficiency (OEE) prediction. The average relative sensitivity is calculated as follows: ; in, As an indicator i At any moment t The relative sensitivity, For availability at any time t Sensitivity, For performance efficiency at all times t Sensitivity, For quality rate at time t Sensitivity, indicators i Including availability Performance efficiency and quality rate .

[0048] from Figure 6 The sensitivity analysis results show that, It has the greatest impact on OEE output, which is consistent with the actual characteristics of hot rolling production lines that operate at high speed and continuously and are sensitive to cycle time fluctuations. Its stable structure results in relatively low sensitivity.

[0049] THPN was compared with five mainstream models—GRU, LSTM, TCN, Autoformer, and Informer—under the same experimental settings. The comparison results are shown in Table 2. THPN significantly outperformed the other models in terms of MSE, RMSE, MAE, and MAPE.

[0050] Table 2. Comparison of different models using a fusion strategy to predict the OEE test set.

[0051] When comparing the above models, it is necessary to compare the prediction errors and fitting degrees of different models under a unified evaluation system. As shown in Table 2, the THPN model of this invention is optimal in terms of overall prediction performance, stability, and error control.

[0052] Comparative experimental analysis results show that, compared with Autoformer, the THPN model of this invention reduces the error in capturing local changes by approximately 10%–15%; compared with LSTM, the trend fitting error is reduced by more than 20%; R 2 The performance metrics are also higher than those of GRU and Informer, indicating that the method of this invention has significant advantages in multidimensional OEE time series prediction.

[0053] Therefore, the above-mentioned multidimensional prediction method for equipment comprehensive efficiency based on time-series hybrid sensing network in this invention has excellent performance in trend preservation, fluctuation capture, sensitivity interpretation and overall prediction accuracy. It can meet the online OEE prediction requirements of hot rolling production lines and has good engineering application prospects.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A device comprehensive efficiency multi-dimensional prediction method based on a timing hybrid perception network, characterized in that, The method comprises the following steps: S1, obtaining OEE historical sequence data, and forming original OEE sequence data after preprocessing; S2, performing time series decomposition and change point identification on the original OEE sequence data by using an STL decomposition method; S3, segmenting the original OEE sequence data according to the identified change points, and dividing each segment of data into a training set and a test set according to the positions of the change points; S4, constructing and training a time series hybrid perception network THPN; S5, inputting real-time collected OEE data into the trained time series hybrid perception network THPN, and performing multi-dimensional fusion prediction; S6, performing OEE sensitivity analysis.

2. The device comprehensive efficiency multi-dimensional prediction method based on a timing hybrid perception network according to claim 1, characterized in that, S1 is specifically: Extracting monitoring quantity data from the production line database to obtain OEE historical data collected by equipment includes three sub-dimension data of availability , performance efficiency and quality rate , and then the original OEE sequence data is obtained after removing abnormal values by using a standard residual method.

3. The device comprehensive efficiency multi-dimensional prediction method based on the timing hybrid perception network according to claim 2, characterized in that, In S1, the standardized residual method is specifically: The OEE historical sequence data is standardized, and the formula is as follows: ； wherein, denotes a time index; denotes a standardized sequence data value obtained after standardizing OEE historical sequence data; denotes an OEE historical sequence original data value collected at a time point denotes an OEE historical sequence original data value collected at a time point denotes a mean value of OEE historical sequence data within a predetermined statistical interval; denotes a standard deviation of OEE historical sequence data within the statistical interval; When an abnormal point is determined and eliminated; The anomaly elimination method refers to using data mean and standard deviation A standardized residual space is constructed to identify outliers by assessing the degree of data deviation from the central interval.

4. The device comprehensive efficiency multi-dimensional prediction method based on a timing hybrid perception network according to claim 3, characterized in that, S2 is specifically: The original OEE sequence data is decomposed into a trend item, a seasonal item and a residual item by using an STL decomposition method ​​​ Then construct residual difference And take residual standard deviation twice as threshold When Determine as turning point, used to detect the change node of production line running state wherein, denotes the absolute difference between the residual terms at two adjacent time instants, which is used to characterize the variation amplitude of the residual terms in the time dimension. denotes the absolute difference between the residual terms at two adjacent time instants, which is used to characterize the variation amplitude of the residual terms in the time dimension. denotes the absolute difference between the residual terms at two adjacent time instants, which is used to characterize the variation amplitude of the residual terms in the time dimension. The change point refers to a point at which a residual term produces a significant mutation in a local section, corresponding to equipment downtime, production rate fluctuation or qualified frequency change in the production line running state.

5. The device comprehensive efficiency multi-dimensional prediction method based on the timing hybrid perception network according to claim 4, characterized in that, S3 is specifically: According to the change point position determined in S2, the original OEE sequence data is divided into multiple running state sections, and the running state section refers to a section in which no new change point appears in the section; And the data of the running state section is divided into a training set and a test set according to the position of the change point, and the training set and the test set come from different production line running states.

6. The device comprehensive efficiency multi-dimensional prediction method based on a timing hybrid perception network according to claim 5, characterized in that, In S4, the constructed time-hybrid perception network THPN includes a dynamic feature separation module, a multi-scale dynamic dependence module, a time embedding module and a fusion decoding module, and the availability , performance efficiency and quality rate are taken as multi-dimensional inputs, a loss function is used for optimization; wherein, denotes a time index; denotes a loss function value adopted in a model training process; denotes a number of samples participating in the model training; denotes a real device comprehensive efficiency target value corresponding to the time point; a device comprehensive efficiency prediction value obtained through a time sequence hybrid perception network THPN. Setting key hyperparameters includes trend window, convolution kernel size, prediction step, hidden dimension, number of attention heads, learning rate and maximum iteration round; The learning rate is used to control the parameter update step of gradient descent; The iteration round represents the number of complete traversals of the model on the training set; The convolution kernel is used to extract local time series patterns, and its size determines the model's ability to capture different scale changes.

7. The device comprehensive efficiency multi-dimensional prediction method based on a timing hybrid perception network according to claim 6, characterized in that, S5 is specifically: sequence availability , performance efficiency , and quality rate are input into the timing hybrid perception network THPN respectively to obtain a predicted sequence , , , and the final equipment overall efficiency OEE prediction result is obtained by using a multiplicative fusion formula, and the multiplicative fusion formula expression is as follows: 。 8. The device synthetic efficiency multi-dimensional prediction method based on a timing hybrid perception network according to claim 7, characterized in that, S6 is specifically: First, perturb the three indexes of availability , performance efficiency and quality rate , and calculate the output difference to obtain the sensitivity , , ; wherein the sensitivity refers to the change range of the overall equipment efficiency OEE output caused by the change of a single input variable while keeping other variables unchanged, and is used to reflect the importance difference of each component index. Then the average relative sensitivity of the overall equipment effectiveness OEE to each sub-dimension is calculated, the relative sensitivity refers to the influence proportion obtained by normalizing the sensitivities of the three indexes at the same time, and is used to measure the relative contribution of the indexes in the overall equipment effectiveness OEE prediction, and the average relative sensitivity is calculated in the following manner: , , ​ ； wherein, is the index of availability i at time t is the relative sensitivity of the index of availability, is the sensitivity of the index of availability at time t is the sensitivity of the index of performance efficiency at time is the sensitivity of the index of quality rate at time t is the sensitivity of the index of quality rate at time is the sensitivity of the index of quality rate at time t is the sensitivity of the index of quality rate at time i the index comprises availability , performance efficiency and quality rate .

Images