A device state prediction method based on parallelized cyclic network

Abstract

A device state prediction method based on a parallelized recurrent network includes the following steps: S1: Improve the GRU network architecture; S2: Use the GELU feedforward network to extract deep feature patterns in the input time series; S3: Add a noise module to ensure parallel training and consistent sequential prediction; S4: Use RMS Norm to improve model efficiency and stability; S5: Use causal convolution to capture long-term causal time series information; S6: Use residual connections to deepen and stack modules; S7: Use the ADAMW optimizer to optimize by superimposing dynamic and prediction losses; S8: Input historical and future control variable information, output the difference between historical and future predictions, and calculate the mean squared error to improve performance; S9: Provide device state warnings based on the prediction results. Compared with traditional LSTM models, this method reduces prediction error and significantly shortens single-sequence inference time, providing a powerful solution for accurate and efficient prediction of industrial equipment states.

Description

Technical Field

[0001] The present invention belongs to the field of industrial equipment state prediction, and in particular relates to an equipment state prediction method based on a parallelized cyclic network. Background Art

[0002] Industrial equipment continuously generates massive amounts of complex time series data during operation. Wind turbines, for example, can generate tens of thousands of time series records daily, including parameters such as speed, temperature, and vibration. While traditional recurrent networks (such as LSTM and GRU) have some processing capabilities for time series data, their architecture is limited. Their serial computation model means that each calculation step depends on the results of the previous step, significantly increasing computation time as the amount of data increases. A typical GRU model would take several minutes to process a 10,000-character sequence of equipment operating parameters on a standard CPU. If the sequence length increases to 50,000, the processing time soars to over half an hour, far from meeting the requirements for real-time monitoring and rapid prediction.

[0003] Furthermore, traditional recurrent networks have a low degree of parallelization. In multi-core CPU or GPU environments, due to the strong interdependencies between computational steps, they struggle to fully utilize hardware parallel computing resources. For example, using a dataset for industrial equipment failure prediction, the parallel speedup ratio of a traditional LSTM model, when equipped with an NVIDIA RTX 3090 GPU, was only approximately 2.5, significantly lower than the speedup ratio of convolutional neural networks, which can exceed 10 on the same hardware. This makes traditional recurrent networks inefficient in both training and prediction for industrial applications when faced with massive amounts of equipment data. New technical solutions are urgently needed to improve the processing efficiency of massive amounts of time series data. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a device status prediction method based on a parallelized recurrent network. Through the synergistic effect of the GELU feedforward network and causal convolution, it can accurately extract deep features and long-term causal information from the device runtime data; and effectively avoid production interruptions caused by sudden equipment failures.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A device state prediction method based on a parallelized cyclic network, comprising the following steps:

[0007] S1. Improve the GRU network architecture;

[0008] S2: Use GELU feedforward network to extract deep feature patterns of input time series;

[0009] S3: Noise adding module to achieve parallel training and sequential prediction consistency;

[0010] S4: Use RMSNorm to improve model efficiency and stability;

[0011] S5: Use causal convolution to capture long-term causal temporal information;

[0012] S6: Use residual connections to deepen and stack modules; S7: Use the ADAMW optimizer to optimize by superimposing dynamics and prediction losses;

[0013] S8: Input historical and future control information, output the difference between historical and future predictions, and calculate the mean square error to improve performance;

[0014] S9: Provide equipment status warning based on the prediction results.

[0015] Preferably, the sub-steps of S1 are:

[0016] S1.1 Definitions At time step The hidden state of the recurrent unit is calculated as:

[0017] ;

[0018] Where, Is the update gate, used to control the hidden state of the previous moment and the current candidate hidden state The combination ratio, is the element-wise multiplication operator, is in the time step The hidden state of the recurrent unit is is the current time step Candidate hidden states of ;

[0019] in, ,

[0020] in Is the logistic sigmoid function, which maps the input value to arrive between, It is the input Perform linear transformation, the output dimension is , is the time step Input data; ;

[0021] S1.2: Use Hesinsen Associative Scan to perform cumulative scans in log space to achieve high-speed parallel training and inference.

[0022] Preferably, sub-step S1.2 is:

[0023] S1.2.1: Use the Heinsen Associative Scan Log module to implement recursive accumulation operations for sequence data involving logarithmic operations.

[0024] S1.2.2: Mathematical formula analysis:

[0025] calculate : logarithmic coefficient array The cumulative sum of the logarithms of all elements is calculated as follows:

[0026] ;

[0027] like is in logarithmic form, then ;

[0028] in is the array of logarithmic coefficients, is an array of logarithmic coefficients The number of elements of is the coefficient array, when When it is not in logarithmic form, ;

[0029] Calculate the log-weighted sum : Combined with the initial value and logarithmic values The calculated logarithmic weighted sum is as follows:

[0030] ;

[0031] is the initial value, a constant or initial state, is through The exponential translation sum of , is an array of logarithmic values, is a logarithmic array The number of elements in ;

[0032] S1.2.3: Confirm variable consistency, dimensionality issues, and clarify initial values ;

[0033] S1.2.4: Final output formula:

[0034] Formula 1: ;

[0035] : The final logarithmic output is the logarithmic coefficient in the first step The cumulative sum The logarithmically weighted sum obtained by exponential shift and summation in the second step is It consists of two parts;

[0036] Formula 2: ;

[0037] :The final result of the logarithmic domain Take the original value obtained by exponential reduction, keep the calculation in the numerically stable logarithmic domain, and finally output the actual value.

[0038] Preferably, the sub-steps of S2 are:

[0039] S2.1: Feedforward network structure: information transfer and nonlinear transformation are achieved through the combination of two linear transformations and activation functions;

[0040] S2.2: Mathematical formula analysis:

[0041] S2.2.1: Step 1: Perform linear transformation 1:

[0042] :enter The result after linear transformation 1 is calculated as follows: ,and ;

[0043] in is the weight matrix of linear transformation 1, with dimension , is the input data of the feedforward network, with dimension , is the bias of linear transformation 1, with dimension , Is the input data Dimensions, is the internal dimension, i.e. the linear transformation 1 output Dimensions;

[0044] S2.2.2: Step 2: Perform nonlinear transformation:

[0045] : The result after the activation function GELU transformation is: ;in is the Gaussian Error Linear Unit activation function, and its formula is , is the cumulative distribution function of the standard normal distribution;

[0046] S2.2.3: Step 3: Linear Transformation 2:

[0047] : The final output of the feedforward network, the same dimension as the input Same, for , the calculation formula is ;here is the weight matrix of linear transformation 2, with dimension , is the bias of linear transformation 2, with dimension ;

[0048] S2.3: Clarify input The dimension is , intermediate state The dimension is , output The dimension is restored to ; GELU provides nonlinear capabilities, enabling the network to learn complex patterns; weight matrix and The dimensions are and .

[0049] Preferably, the S3 sub-step is as follows: S3.1: the BlockNoise1D module performs an operation of injecting noise into the input data in units of blocks;

[0050] S3.2: Perform mathematical formula analysis:

[0051] S3.2.1: Step 1: Generate mask matrix

[0052] : Mask matrix, used to determine which blocks need to be noised, its elements The definition of

[0053] ,

[0054] in is the row index of the mask matrix, is the column index of the mask matrix;

[0055] S3.2.2: Step 2: Generate noise as input data:

[0056] :For input data The noise generated is calculated as ;

[0057] in The mean is , the standard deviation is The normal distribution of is the standard deviation of the normal distribution, controlling the noise intensity, The first elements;

[0058] S3.2.3: Step 3: Perform noise injection operation:

[0059] S3.3: Clarify the definition of noise blocks (such as time steps, feature dimensions, etc.), noise standard deviation (needs to be adjusted according to task requirements), the method of generating the mask matrix (based on random sampling or fixed pattern);

[0060] S3.4: Apply BlockNoise1D module for data augmentation, regularization, and robust training.

[0061] Preferably, the sub-steps of S4 are:

[0062] S4.1: Normalize the input vector using RMSNorm and introduce a learnable scaling factor , balance the gradient scale and improve network stability;

[0063] S4.2: Mathematical formula analysis:

[0064] S4.2.1: Step 1: Normalization:

[0065] :enter After normalization, elements, the calculation formula is ;in is the input vector No. elements, is the input vector The L2 norm of ;

[0066] S4.2.2: Step 2: Scaling and Adjustment:

[0067] : The final output of the RMSNorm module is calculated as ;in is a learnable scaling factor used to adjust the normalized output, Determined by the square root of the input dimension (either a constant or a function based on the input dimension, such as ), used to further adjust the output scale, is the input vector Dimensions;

[0068] S4.3: Explicitly normalize the last dimension, suitable for multi-dimensional input (such as sequence data or feature vectors); learnable parameters Allow the model to dynamically adjust the normalized output according to task requirements; if The formula is not specified, it is a constant or a function based on the input dimension.

[0069] Preferably, the sub-steps of S5 are:

[0070] S5: Use causal convolution to capture richer long-term causal timing information:

[0071] S5.1: Use Causal Depthwise Convolution for time series data, ensuring that the model can only access past information through depthwise separable convolution and causal padding;

[0072] S5.2: Mathematical formula analysis:

[0073] S5.2.1: Step 1: Causal Filling:

[0074] :enter The result after causal filling is calculated as follows: ;in is the input time series data of the causal convolution module, is the filling function, Indicates filling at the beginning of the sequence Zeros, is the size of the convolution kernel;

[0075] S5.2.2: Step 2: Depthwise Separable Convolution:

[0076] : The output result after depth-wise separable convolution, the dimension is the same as the input The same calculation formula is ,in It is a depth-separable convolution operation;

[0077] S5.3: Causal padding is used to ensure that the model can only access past information, which is suitable for time series data modeling (such as time series prediction, speech processing, etc.); depthwise separable convolution reduces the number of parameters and computation while maintaining the nonlinear capability of the convolution operation; padding zeros to ensure that the convolution kernel does not access future data points.

[0078] Preferably, the specific content of S8 is: inputting all historical information of a period of time: the data size is L1, D;

[0079] Enter the future control quantity information: size is L2, D - 1;

[0080] The output is a differential history prediction of length L1 and a future prediction of L2. The mean squared error loss is calculated for both predictions to improve performance.

[0081] Where: L1, D: feature dimensions of input historical information data;

[0082] L2: the length dimension of future control quantity information data;

[0083] D-1: characteristic dimension of future control quantity information data;

[0084] Preferably, sub-step S9:

[0085] S9.1: Obtain measurement points such as unit status, active power, outlet current, air cooler cooling water main inlet and outlet temperatures, and stator winding temperature;

[0086] S9.2: Preprocess the data using the specified method;

[0087] S9.3: Design a new network architecture that can learn the operating patterns of equipment and adapt to the differences in equipment dynamics caused by different measurement points. After the model is trained, it can also show good results when applied to equipment in other units.

[0088] S9.4: Test model performance: Make short-term and long-term data predictions for target measurement points and compare indicators.

[0089] A device state prediction system based on a parallelized cyclic network adopts the device state prediction method based on a parallelized cyclic network.

[0090] The present invention can achieve the following beneficial effects:

[0091] 1. The synergistic effect of the GELU feedforward network and causal convolution can accurately extract deep features and long-term causal information from equipment runtime data. In motor fault prediction scenarios, the ability to capture early and subtle fault characteristics is greatly enhanced, effectively avoiding production interruptions caused by sudden equipment failures.

[0092] 2. The improved GRU network architecture, combined with the parallel training and inference mechanism of Hesinsen Associative Scan, significantly reduces training time. When processing large-scale device operation data, training speed is 3-5 times faster than traditional recurrent networks, enabling rapid completion of model training and parameter optimization, meeting enterprise needs for rapid deployment and iteration of models. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The present invention will be further described below with reference to the accompanying drawings and examples:

[0094] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0095] The preferred solution is Figure 1 As shown in FIG, a device state prediction method based on a parallelized cyclic network includes the following steps:

[0096] S1: Improved GRU network architecture:

[0097] S1.1: Define the hidden state of the recurrent unit at time step t , and its calculation formula is .

[0098] is the update gate, through Calculated, where Is the logistic sigmoid function, which maps the input value to between 0 and 1. It is the input (Input data at time step t) is linearly transformed, and the output dimension is (set to 50). is in the time step The hidden state of the recurrent unit is is the candidate hidden state for the current time step, is the element-wise multiplication operator.

[0099] S1.2: Sub-steps:

[0100] S1.2.1: Use the Heinsen Associative Scan Log module to perform recursive accumulation operations on sequence data containing logarithmic operations.

[0101] S1.2.2: Calculation , if the logarithmic coefficient array It is not in logarithmic form, and its calculation formula is ;

[0102] like is in logarithmic form, then .

[0103] here Set to 100 (the number of elements in the logarithmic coefficient array), is the coefficient array, when When it is not in logarithmic form, .

[0104] Then calculate the log-weighted sum ,

[0105] in Set to 0.5 (initial value, constant), is through The exponential translation sum of is an array of logarithmic values, Set to 200 (the number of elements in the logarithmic array).

[0106] S1.2.3: Carefully confirm variable consistency, check dimensionality issues (such as the matching of the dimensions of the logarithmic coefficient array and the logarithmic value array), and clarify initial values. settings.

[0107] S1.2.4: Get the final output formula, formula 1 is , is the final logarithmic output, given by and It consists of two parts; Formula 2 is , The final result of the logarithmic domain is Take the original value obtained by exponential restoration to ensure that the calculation is performed in the numerically stable logarithmic domain, and finally output the actual value.

[0108] S2: Use GELU feedforward network to extract deep feature patterns of input time series:

[0109] S2.1: Construct a feedforward network structure and realize information transfer and nonlinear transformation through the combination of two linear transformations and activation functions.

[0110] S2.2: The sub-steps are:

[0111] S2.2.1: Perform linear transformation 1, input (Dimensions are , set to 20 dimensions) after linear transformation 1 , the calculation formula is ,and (Assumed to be 30 dimensions). is the weight matrix of linear transformation 1, with dimension (i.e. 30×20), is the bias of linear transformation 1, with dimension (30 dimensions).

[0112] S2.2.2: Perform nonlinear transformations, The result after the activation function GELU transformation , the formula is .in is the Gaussian Error Linear Unit activation function, and its formula is , is the cumulative distribution function of the standard normal distribution.

[0113] S2.2.3: Perform linear transformation 2, the final output of the feedforward network , dimension and input Same, for (20 dimensions), the calculation formula is .here is the weight matrix of linear transformation 2, with dimension (20×30), is the bias of linear transformation 2, with dimension (20 dimensions).

[0114] S2.3: Clarify input The dimension is (20 dimensions), intermediate state The dimension is (30 dimensions), output The dimension is restored to (20 dimensions). Recognizing the nonlinear capabilities provided by GELU, the network is able to learn complex patterns, and the weight matrix and The dimensions are 30×20 and 20×30 respectively.

[0115] S3: Add noise module to achieve consistency between parallel training and sequential prediction.

[0116] S3.1: Use the BlockNoise1D module to inject noise into the input data in blocks. Here, the data is divided into blocks of 10 time steps.

[0117] S3.2: Sub-steps:

[0118] S3.2.1: Generate mask matrix , whose elements is defined as:

[0119] ,

[0120] in is the row index of the mask matrix, are the column indices of the mask matrix.

[0121] S3.2.2: Noise generated as input data , the calculation formula is .

[0122] in The mean is 0 and the standard deviation is (set to 0.1) of the normal distribution, The first elements.

[0123] S3.2.3: Perform a noise injection operation.

[0124] S3.3: The noise block is defined as a block of 10 time steps, and the noise standard deviation It is set to 0.1 (pre-set according to data characteristics and training results), and the mask matrix is ​​generated by random sampling.

[0125] S3.4: Apply the BlockNoise1D module for data augmentation, regularization, and robust training.

[0126] S4: Use RMSNorm to improve model efficiency and stability:

[0127] S4.1: Use RMSNorm to normalize the input vector (dimension set to 20) and introduce a learnable scaling factor , balance the gradient scale and improve network stability.

[0128] S4.2: Sub-steps:

[0129] S4.2.1: Perform normalization and input After normalization, Elements , the calculation formula is ,in is the first elements, is the L2 norm of the input vector.

[0130] S4.2.2: Scaling and adjustment, the final output of the RMSNorm module , the calculation formula is .in Set to 1.2 (learnable parameter, initial value), Set as ( is the dimension of the input vector, i.e. 20, for ), which is used to further adjust the output scale.

[0131] S4.3: Explicitly normalize the last dimension and apply it to the current 20-dimensional input vector. Learnable parameters Allows the model to dynamically adjust the normalized output according to task requirements. The formula is not clear, this embodiment is set as a function based on the input dimension .

[0132] S5: Using causal convolution to capture long-term causal timing information:

[0133] S5.1: Use Causal Depthwise Convolution for time series data, ensuring that the model can only access past information through depth-wise separable convolution and causal padding.

[0134] S5.2: The sub-steps are:

[0135] S5.2.1: Perform causal filling and input The result after causal filling , the calculation formula is .in is the input time series data of the causal convolution module, is the filling function, Set to 3 (the size of the convolution kernel).

[0136] S5.2.2: Perform depthwise separable convolution, output after depthwise separable convolution , dimension and input The same calculation formula is ,in is a depth-wise separable convolution operation.

[0137] S5.3: Explicitly ensure that the model can only access past information through causal filling, which is suitable for modeling time series data of equipment operation. Depthwise separable convolution reduces the number of parameters and computation while maintaining the nonlinear capability of convolution operation. zeros to ensure that the convolution kernel does not access future data points.

[0138] S6: Use residual connections to deepen and stack modules:

[0139] The modules processed in the above steps are stacked in order, and the output of the previous layer is directly connected to the input of the next layer through residual connections to ensure the effective transmission of information in the model and deepen the model structure. For example, the output of the improved GRU network is connected to the input of the GELU feedforward network through residual connections, which enables the model to learn more complex features.

[0140] S7: Optimization using the ADAMW optimizer, superimposing dynamics and prediction losses:

[0141] The model was trained and optimized using the ADAMW optimizer, with a learning rate of 0.001 and a weight decay of 0.01. The loss function superimposed the dynamic loss and prediction loss. The dynamic loss was calculated based on the physical laws of device operation, while the prediction loss used a mean squared error loss to calculate the error between the model's predicted value and the true value. By continuously adjusting the model parameters, the superimposed loss value was minimized, thereby improving model performance.

[0142] S8: Input historical and future control information, output differential historical and future predictions, and calculate the mean square error to improve performance. The prepared historical information data of size L1 (1000), D (20) and the future control information data of size L2 (10), D-1 (19) are input into the model. The model outputs the differential historical prediction of length L1 (1000) and the future prediction of length L2 (10). The mean square error loss is calculated for both prediction values. By continuously adjusting the model parameters, the mean square error is gradually reduced to improve the model prediction performance.

[0143] S9: Provide equipment status warning based on prediction results:

[0144] S9.1: Obtain measurement point data such as unit status, active power, outlet current, air cooler cooling water main inlet and outlet temperatures, and stator winding temperature.

[0145] S9.2: Use normalization to preprocess the data and unify the data into the range of 0-1.

[0146] S9.3: Design a new network architecture that incorporates the above modules. This network architecture learns the operating patterns of the equipment and adapts to the differences in equipment dynamics at different measurement points. After training the model on the current plant equipment, apply it to other units of the same type for testing.

[0147] S9.4: Test model performance by performing short-term (one hour into the future) and long-term (one day into the future) forecasts for target measurement points and comparing metrics. Model prediction accuracy is evaluated using metrics such as root mean square error (RMSE) and mean absolute error (MAE). For example, in the temperature forecast for a key measurement point, the short-term RMSE is 0.5°C and the MAE is 0.3°C; the long-term RMSE is 1.2°C and the MAE is 0.8°C. Based on pre-set thresholds (e.g., the temperature anomaly threshold is ±2°C from the normal temperature), a timely equipment status warning signal is issued when the predicted value exceeds the threshold.

[0148] The above embodiments demonstrate the specific implementation process and effect of the device state prediction method based on the parallelized cyclic network in actual device state prediction.

[0149] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A device state prediction method based on a parallelized cyclic network, characterized in that The following steps are involved: S1. Improve the GRU network architecture; S2: Use GELU feedforward network to extract deep feature patterns of input time series; S3: Noise adding module to achieve parallel training and sequential prediction consistency; S4: Use RMSNorm to improve model efficiency and stability; S5: Use causal convolution to capture long-term causal temporal information; S6: Use residual connections to deepen and stack modules; S7: Use the ADAMW optimizer to optimize by superimposing dynamics and prediction losses; S8: Input historical and future control information, output the difference between historical and future predictions, and calculate the mean square error to improve performance; S9: Provide equipment status warning based on the prediction results; The sub-steps of S1 are: S1.1 Definitions At time step The hidden state of the recurrent unit is calculated as: ； Where, Is the update gate, used to control the hidden state of the previous moment and the current candidate hidden state The combination ratio, is the element-wise multiplication operator, is in the time step The hidden state of the recurrent unit is is the current time step Candidate hidden states of ; in, , in Is the logistic sigmoid function, which maps the input value to arrive between, It is the input Perform linear transformation, the output dimension is , is the time step Input data; ; S1.2: Use Hesinsen Associative Scan to perform cumulative scans in log space to achieve high-speed parallel training and inference; Sub-steps S1.2 are: S1.2.1: Use the Heinsen Associative Scan Log module to implement recursive accumulation operations for sequence data involving logarithmic operations. S1.2.2: Mathematical formula analysis: calculate : logarithmic coefficient array The cumulative sum of the logarithms of all elements is calculated as follows: ； like is in logarithmic form, then ; in is the array of logarithmic coefficients, is an array of logarithmic coefficients The number of elements of is the coefficient array, when When it is not in logarithmic form, ; Calculate the log-weighted sum : Combined with the initial value and logarithmic values The calculated logarithmic weighted sum is as follows: ； is the initial value, a constant or initial state, is through The exponential translation sum of , is an array of logarithmic values, is a logarithmic array The number of elements in ; S1.2.3: Confirm variable consistency, dimensionality issues, and clarify initial values ; S1.2.4: Final output formula: Formula 1: ; : The final logarithmic output is the logarithmic coefficient in the first step The cumulative sum The logarithmically weighted sum obtained by exponential shift and summation in the second step is It consists of two parts; Formula 2: ; :The final result of the logarithmic domain Take the original value obtained by exponential reduction, keep the calculation in the numerically stable logarithmic domain, and finally output the actual value; S9 sub-step: S9.1: Obtain unit status, active power, export current, air cooler cooling water main inlet and outlet temperatures, and stator winding temperature; S9.2: Preprocess the data using the specified method; S9.3: Make short-term and long-term data forecasts for target measuring points and compare indicators.

2. The device state prediction method based on a parallelized cyclic network according to claim 1, characterized in that: The sub-steps of S2 are: S2.1: Feedforward network structure: information transfer and nonlinear transformation are achieved through the combination of two linear transformations and activation functions; S2.2: Mathematical formula analysis: S2.2.1: Step 1: Perform linear transformation 1: :enter The result after linear transformation 1 is calculated as follows: ,and ; in is the weight matrix of linear transformation 1, with dimension , is the input data of the feedforward network, with dimension , is the bias of linear transformation 1, with dimension , Is the input data Dimensions, is the internal dimension, i.e. the linear transformation 1 output Dimensions; S2.2.2: Step 2: Perform nonlinear transformation: : The result after the activation function GELU transformation is: ;in is the Gaussian Error Linear Unit activation function, and its formula is , is the cumulative distribution function of the standard normal distribution; S2.2.3: Step 3: Linear Transformation 2: : The final output of the feedforward network, the same dimension as the input Same, for , the calculation formula is ;here is the weight matrix of linear transformation 2, with dimension , is the bias of linear transformation 2, with dimension ; S2.3: Clarify input The dimension is , intermediate state The dimension is , output The dimension is restored to ; GELU provides nonlinear capabilities, enabling the network to learn complex patterns; weight matrix and The dimensions are and .

3. The device state prediction method based on a parallelized cyclic network according to claim 1, characterized in that: The sub-steps of S3 are as follows: S3.1: Use the BlockNoise1D module to inject noise into the input data in units of blocks; S3.2: Perform mathematical formula analysis: S3.2.1: Step 1: Generate mask matrix : Mask matrix, used to determine which blocks need to be noised, its elements The definition of ， in is the row index of the mask matrix, is the column index of the mask matrix; S3.2.2: Step 2: Generate noise as input data: :For input data The noise generated is calculated as ; in The mean is , the standard deviation is The normal distribution of is the standard deviation of the normal distribution, controlling the noise intensity, The first elements; S3.2.3: Step 3: Perform noise injection operation: S3.3: Clarify the definition of noise block and noise standard deviation , the generation method of the mask matrix; S3.4: Apply BlockNoise1D module for data augmentation, regularization, and robust training.

4. The device state prediction method based on a parallelized cyclic network according to claim 1, characterized in that: The sub-steps of S4 are: S4.1: Normalize the input vector using RMSNorm and introduce a learnable scaling factor , balance the gradient scale and improve network stability; S4.2: Mathematical formula analysis: S4.2.1: Step 1: Normalization: :enter After normalization, elements, the calculation formula is ;in is the input vector No. elements, is the input vector The L2 norm of ; S4.2.2: Step 2: Scaling and Adjustment: : The final output of the RMSNorm module is calculated as ;in is a learnable scaling factor used to adjust the normalized output, Determined by the square root of the input dimension, is the input vector Dimensions; S4.3: Explicitly normalize the operation on the last dimension, applicable to multidimensional input; Learning parameters Allows the model to dynamically adjust the normalized output according to task requirements.

5. The device state prediction method based on a parallelized cyclic network according to claim 1, characterized in that: The sub-steps of S5 are: S5: Use causal convolution to capture richer long-term causal timing information: S5.1: Use Causal Depthwise Convolution for time series data, ensuring that the model can only access past information through depthwise separable convolution and causal padding; S5.2: Mathematical formula analysis: S5.2.1: Step 1: Causal Filling: :enter The result after causal filling is calculated as follows: ;in is the input time series data of the causal convolution module, is the filling function, Indicates filling at the beginning of the sequence Zeros, is the size of the convolution kernel; S5.2.2: Step 2: Depthwise Separable Convolution: : The output result after depth-wise separable convolution, the dimension is the same as the input The same calculation formula is ,in It is a depth-separable convolution operation; S5.3: Explicitly ensure that the model can only access past information through causal padding; maintain the nonlinearity of convolution operations; padding zeros to ensure that the convolution kernel does not access future data points.

6. The device state prediction method based on a parallelized cyclic network according to claim 1, characterized in that: The specific content of S8 is: input all historical information for a period of time: the data size is L1, D; Enter the future control quantity information: size is L2, D - 1; The output is a differential history prediction of length L1 and a future prediction of L2. The mean squared error loss is calculated for both predictions to improve performance. Where: L1, D: feature dimensions of input historical information data; L2: the length dimension of future control quantity information data; D-1: Characteristic dimension of future control quantity information data.

7. A device state prediction system based on a parallelized cyclic network, characterized by: The device state prediction method based on a parallelized cyclic network according to claim 1 is adopted.

Images