Dynamic simulation prediction model construction method based on data driving

Abstract

The invention discloses a dynamic simulation prediction model construction method based on data driving. The dynamic simulation prediction model construction method comprises the following steps: S1, collecting multi-source heterogeneous multi-modal original data and preprocessing to obtain a standardized data set; s2, extracting a long-term dependent feature sequence by adopting a Mamba depth state space model; s3, establishing a parameter diffusion space by using the self-adaptive multi-mode diffusion model, and obtaining diffusion characteristic parameters; s4, performing reverse denoising convergence processing on the diffusion characteristic parameters to generate real-time adaptive model parameters; s5, performing dynamic weighted fusion of the feature sequence based on real-time adaptive model parameters; s6, inputting the fusion sequence into a dynamic simulation prediction model framework for training, and obtaining trained model parameters; and S7, solidifying model parameters to complete model construction. According to the invention, the dynamic prediction precision and generalization performance are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of simulation prediction model construction, and in particular to a data-driven dynamic simulation prediction model construction method. Background Art

[0002] Dynamic simulation prediction technology is widely used in many fields such as industrial equipment monitoring, traffic flow forecasting, power load forecasting and financial risk assessment. It is an important technical means to achieve complex system status analysis and accurate decision-making.

[0003] Among existing data-driven dynamic simulation prediction models, typical technical approaches include those based on traditional statistical models, those based on machine learning models, and the recently emerging deep learning-based prediction methods. Traditional statistical models such as ARIMA (Autoregressive Integrated Moving Average) and Kalman filtering offer advantages such as simple modeling and low computational complexity. However, these methods typically assume stationary time series and are difficult to handle with high-dimensional, nonlinear, and nonstationary data, resulting in limited generalization performance.

[0004] In recent years, machine learning methods represented by support vector machines (SVM), random forests (RF), and XGBoost have been introduced into the field of dynamic simulation and prediction. These methods improve the model's fitting ability and prediction performance for complex data through feature engineering and complex nonlinear mapping relationships. However, these methods are highly dependent on manual feature selection and design, and face problems such as difficulty in feature extraction, poor transferability, and unstable model generalization performance. They have obvious limitations in actual complex environments.

[0005] With the rise of deep learning technology, recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and Transformer models have gradually demonstrated their advantages in the field of dynamic simulation and prediction. These methods, through automatic feature extraction and end-to-end training, significantly reduce reliance on manual feature engineering and demonstrate superior performance in sequence feature extraction and modeling. For example, RNNs and LSTMs can capture long-term data dependencies through state propagation, while Transformers leverage self-attention mechanisms to efficiently model long-range dependencies within sequences. However, existing deep learning methods still face significant bottlenecks when processing sequence data with longer time spans. For example, RNNs and LSTMs suffer from vanishing and exploding gradients, making it difficult for these models to effectively capture long-term dependencies. While Transformers overcome this gradient problem, the amount of self-attention computation required for these models increases quadratically with increasing sequence length, resulting in high computational complexity and memory requirements. These limitations limit their applicability to large-scale data and long-sequence prediction scenarios.

[0006] In addition, in practical applications, it is often necessary to fuse data from multiple modalities in order to more comprehensively describe the state characteristics of complex systems. However, multimodal data usually have characteristics such as inconsistent dimensions, large distribution differences, and complex correlations between modalities. Traditional data fusion methods mostly use simple feature splicing or weighted combination methods, lacking an effective dynamic fusion mechanism, making it difficult to fully explore the correlation information between different modal data, resulting in unsatisfactory fusion effects, which in turn reduces the accuracy and generalization ability of the overall model prediction. In addition, existing models usually have fixed parameters or can only rely on manual periodic adjustments, making it difficult to achieve real-time adaptive adjustment of model parameters according to environmental changes, making it difficult for the model to effectively adapt to the dynamic changes of data in actual working conditions, further limiting the practical application performance of existing methods.

[0007] Therefore, how to provide a data-driven dynamic simulation prediction model construction method is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0008] One object of the present invention is to propose a data-driven dynamic simulation prediction model construction method. The present invention adopts the Mamba deep state space model in combination with an adaptive multimodal diffusion learning algorithm. By collecting multi-source heterogeneous multimodal raw data and performing standardized preprocessing, the Mamba deep state space model is used to extract the long-term dependent feature sequence of each modal data. The adaptive multimodal diffusion model is further used to establish a parameter diffusion space. The feature parameters are diffused and reversely denoised in the diffusion space in real time to generate real-time adaptive model parameters. The long-term dependent feature sequence is dynamically fused based on the real-time adaptive model parameters to obtain a fused feature sequence. The fused feature sequence is then input into a dynamic simulation prediction model framework comprising a multi-scale feature extraction unit, a bidirectional gated recurrent unit layer, a dynamic feature selection self-attention layer and a multi-task collaborative prediction output layer. The model parameters are generated and solidified through iterative training of forward propagation and backpropagation, which effectively solves the problems of insufficient long-sequence feature capture capability of traditional models, poor multimodal data fusion effect, and difficulty in real-time adaptation to environmental changes due to fixed parameters. The framework has the technical advantages of strong ability to capture long-term sequence dependency features, high generalization performance, precise multimodal fusion and obvious real-time adaptive optimization effect, effectively improving the accuracy, stability and practical applicability of dynamic simulation prediction.

[0009] A data-driven dynamic simulation prediction model construction method according to an embodiment of the present invention includes the following steps:

[0010] S1. Collecting multi-source heterogeneous multimodal raw data through a data acquisition device, and preprocessing the multimodal raw data to obtain a standardized data set of each modality data;

[0011] S2. Use the Mamba deep state space model to extract sequence features from the standardized data sets of each modality data to obtain the long-term dependency feature sequence of each modality;

[0012] S3. Establishing a parameter diffusion space using an adaptive multimodal diffusion model, inputting a long-term dependent feature sequence into the parameter diffusion space, performing real-time diffusion processing on the parameters, and obtaining characteristic parameters after diffusion;

[0013] S4, performing reverse denoising and convergence processing on the diffused characteristic parameters to generate real-time adaptive model parameters;

[0014] S5. Dynamically weighted fusion of the long-term dependent feature sequences of each modality is performed based on the real-time adaptive model parameters to obtain a fused feature sequence;

[0015] S6. Inputting the fused feature sequence into a preset dynamic simulation prediction model framework, training the dynamic simulation prediction model framework in combination with real-time adaptive model parameters to generate trained model parameters;

[0016] S7. Solidify the trained model parameters into the dynamic simulation prediction model framework to complete the construction of the dynamic simulation prediction model.

[0017] Optionally, the S1 specifically includes:

[0018] S11. Using a vibration sensor array to collect vibration modal data, and performing filtering, noise reduction, and amplitude normalization on the vibration modal data to obtain a standardized vibration modal data set;

[0019] S12. Acoustic modal data is collected using a sound sensor array, and the acoustic modal data is subjected to denoising, feature filtering, and sampling rate adjustment to obtain a standardized acoustic modal data set.

[0020] S13, using a temperature sensor array to collect temperature modal data, and performing outlier removal, smoothing filtering, and scaling processing on the temperature modal data to obtain a standardized temperature modal data set;

[0021] S14, using a current sensor array to collect electrical modal data, and performing denoising filtering, feature extraction, and normalization processing on the electrical modal data to obtain a standardized electrical modal data set;

[0022] S15. Using an image sensor array to collect visual modality data, and performing image enhancement, deblurring, and image size normalization on the visual modality data to obtain a standardized visual modality dataset;

[0023] S16. Using a wireless signal acquisition device to collect communication modal data, and performing signal demodulation, filtering, and amplitude normalization processing on the communication modal data to obtain a standardized communication modal data set;

[0024] S17. When the data collection stage does not set any one or more of vibration modal data, acoustic modal data, temperature modal data, electrical modal data, visual modal data, and communication modal data to be collected, the corresponding modal data position that is not set to be collected is marked with an empty data marker for placeholder processing;

[0025] S18. When data is missing in the modalities that have been set for collection, the missing modal data are reconstructed using the historical modal data interpolation method to obtain a complete standardized data set;

[0026] S19: Perform timestamp alignment and data fusion processing on the standardized data sets of each modality obtained in steps S11 to S18 to obtain a standardized data set of each modality data that is time-synchronized.

[0027] Optionally, the S2 specifically includes:

[0028] S21. Construct a Mamba deep state-space model consisting of one input mapping layer, three selective state-space layers, two residual connection layers, and one output mapping layer connected in series.

[0029] S22. Determine the number of input neurons of the input mapping layer to be 256 according to the data dimension of the standardized data set of each modal data, and use a linear mapping method to map the standardized data set of each modal data to an initial input feature vector;

[0030] S23. Initialize the hidden state vector in the first selective state space layer to an all-zero vector with a dimension of 256, construct a linear state space equation with a state dimension of 256, use the initial input feature vector as input to perform state update, and obtain the first state output vector;

[0031] S24, using the first residual connection layer to superimpose and fuse the first state output vector and the initial input feature vector element by element to obtain a first fused state feature representation, and using the first fused state feature representation as the input of the second selective state space layer;

[0032] S25. Initialize the hidden state vector in the second selective state space layer to the first state output vector with a dimension of 128, and construct a linear state space equation with a state dimension of 128. Use the first fused state feature representation as input to update the state to obtain the second state output vector. Then, use the second residual connection layer to superimpose and fuse the second state output vector with the first fused state feature representation element by element to obtain the second fused state feature representation. Use the second fused state feature representation as the input of the third selective state space layer.

[0033] S26. Initialize the hidden state vector in the third selective state space layer to the second state output vector with a dimension of 64, construct a linear state space equation with a state dimension of 64, use the second fused state feature representation as input to perform state update, and obtain the third state output vector. Then, map the third state output vector to a feature space with a dimension of 64 through the output mapping layer to obtain the sequence feature representation of each modal data.

[0034] S27. Using the sequence feature representation of each modal data as the training target, perform forward propagation and backpropagation iterative training on the Mamba deep state-space model using the standardized data set of each modal data. The number of iterative training times is 300, and the learning rate is 0.0005. The state space parameter matrices of the first to third selective state space layers are updated by the gradient descent algorithm to obtain a converged Mamba deep state-space model.

[0035] S28. Input the standardized data sets of each modality data into the converged Mamba deep state space model respectively to extract sequence features and obtain the long-term dependency feature sequence of each modality.

[0036] Optionally, the S3 specifically includes:

[0037] S31. Construct a parameter diffusion space with a dimension of 64 based on the long-term dependency feature sequence of each modality, wherein the initial probability distribution of any feature parameter in the parameter diffusion space is a standard Gaussian distribution;

[0038] S32. Construct a three-layer feedforward neural network for predicting diffusion process noise, wherein the input layer dimension is 64 neurons, the hidden layer dimension is 128 neurons, and the output layer dimension is 64 neurons. The neural network takes the diffusion characteristic parameter and the number of diffusion steps as input and uses the predicted noise as output;

[0039] S33. Set the diffusion step length of the parameter diffusion space to 1000 steps, and define the state transfer equation of the forward diffusion process as:

[0040]

[0041] Among them, z t represents the diffusion characteristic parameter of the tth step, ∈ represents the random noise sampled from the standard Gaussian distribution, β t It represents the diffusion variance adjustment coefficient of the t-th step and satisfies the linear adjustment method. The value range is 0.0001 to 0.02;

[0042] S34, taking the long-term dependency feature sequence of each mode as the initial input, performing step-by-step iteration through the forward diffusion process, and generating a diffusion feature parameter sequence evolving from low noise to high noise in the parameter diffusion space;

[0043] S35. For the diffusion characteristic parameter sequence, the state transfer equation of the reverse diffusion process is defined as:

[0044]

[0045] in, ∈ θ (z t ,t) is the noise of the t-th step predicted by the neural network, and the neural network parameters θ are obtained through training;

[0046] S36, iteratively optimize the neural network parameters θ in the back diffusion process, the number of training times is 400 times, and the optimization goal is to make the predicted noise ∈ θ (z t ,t) and the mean square error between the real noise ∈ added in the forward diffusion process is minimized;

[0047] S36, through the back diffusion process, the real added noise ∈ and the predicted noise ∈ θ (z t ,t) between the mean square error loss function, the gradient descent algorithm is used to iteratively train the parameters of the three-layer feedforward neural network, the number of training times is 400 times, the learning rate is 0.001, and the three-layer feedforward neural network with training convergence is obtained;

[0048] S37, applying the three-layer feedforward neural network after training convergence in the back diffusion process of the parameter diffusion space in real time to generate a feature parameter sequence after diffusion in real time;

[0049] S38. Extract the feature parameters of the first diffusion step from the feature parameter sequence after diffusion generated in real time as the feature parameters after diffusion.

[0050] Optionally, the S4 specifically includes:

[0051] S41. Construct a reverse denoising convergence space with a dimension of 64 based on the characteristic parameters after diffusion, and initialize the convergence step size in the reverse denoising process to 1000 steps;

[0052] S42. Define the state transfer equation of the reverse denoising process as:

[0053]

[0054] in, Represents the characteristic parameters after the t-th step reverse denoising process. Initially is the characteristic parameter after diffusion, γ t It represents the denoising coefficient in the t-th step reverse denoising process, and its value range is 0.0001 to 0.02. represents the predicted noise of the t-th step reverse denoising;

[0055] S43. Construct a two-layer fully connected neural network for predicting the noise of the reverse denoising process, with an input layer dimension of 64 neurons and an output layer dimension of 64 neurons, taking the characteristic parameters of the reverse denoising and the number of denoising steps as input, and predicting the denoised noise;

[0056] S44. Using a standardized dataset of historically collected data from each modality, the mean square error between the actual added noise and the predicted denoised noise during the reverse denoising process was used as the loss function. The parameters of the two-layer fully connected neural network were iteratively trained using the stochastic gradient descent algorithm. The number of training times was 300 and the learning rate was 0.001. A converged two-layer fully connected neural network was obtained.

[0057] S45, using a two-layer fully connected neural network after training convergence to predict the noise of the reverse denoising process in real time, executing the reverse denoising process, and generating a denoising feature parameter sequence evolving from high noise to low noise in real time;

[0058] S46. Extracting the feature parameters of the first step of reverse denoising from the denoising feature parameter sequence generated in real time as real-time adaptive model parameters.

[0059] Optionally, the S5 specifically includes:

[0060] S51, determining a real-time fusion weight corresponding to each modal feature sequence based on real-time adaptive model parameters;

[0061] S52, using the real-time fusion weight to perform weighted processing on the long-term dependent feature sequence of each modality to obtain a weighted feature sequence of each modality;

[0062] S53, according to the timestamp information of the feature sequence, performing a fusion operation on the sequence elements of the weighted feature sequence of each modality according to the corresponding timestamp to obtain an initial fused feature sequence;

[0063] S54, calculating the stability index of the initial fusion feature sequence, and determining whether the fusion sequence meets the preset stability condition;

[0064] S55: If the fused sequence does not meet the preset stability condition, the real-time fusion weights of each modal feature sequence are updated based on the real-time adaptive model parameters, and the process returns to step S52 to perform dynamic weighted fusion of the feature sequence again.

[0065] S56. If the fused sequence meets the preset stability condition, the current fused feature sequence is output as the final fused feature sequence.

[0066] Optionally, the preset dynamic simulation prediction model framework specifically includes:

[0067] A multi-scale feature extraction unit is constructed, which includes three parallel one-dimensional convolution channels with convolution kernel sizes of 3, 5, and 7, the number of convolution kernels is 32, and the convolution step size is 1. It is used to extract local features of different scales from the input fused feature sequence and obtain a multi-scale fused feature sequence by element-by-element splicing;

[0068] Constructing a bidirectional gated recurrent unit layer, including two layers of bidirectional gated recurrent units connected in series, each layer of bidirectional gated recurrent units containing 128 hidden units, for performing forward and backward deep sequence feature modeling on the multi-scale fusion feature sequence;

[0069] Construct a dynamic feature selection self-attention layer, including a multi-head self-attention mechanism with 8 self-attention heads and 128 dimensions. The attention weight of each sequence element is dynamically calculated based on the real-time adaptive model parameters to highlight the key features in the fused feature sequence that have an important impact on the dynamic prediction results.

[0070] Construct a multi-task collaborative prediction output layer, which includes two parallel fully connected prediction sublayers. The first prediction sublayer is used for short-term prediction and contains 64 neurons. The second prediction sublayer is used for long-term prediction and contains 64 neurons. Both prediction sublayers receive features output from the attention layer in parallel.

[0071] Constructing a fusion output layer, performing linear mapping fusion on the prediction results of the two prediction sub-layers in the multi-task collaborative prediction output layer to obtain the final fusion prediction numerical output;

[0072] The fused feature sequence and real-time adaptive model parameters are input into the dynamic simulation prediction model framework. Through forward propagation and backpropagation iterative optimization, the model parameters of the multi-scale feature extraction unit, bidirectional gated recurrent unit layer, dynamic feature selection self-attention layer and multi-task collaborative prediction output layer are trained to generate trained model parameters.

[0073] Optionally, the S6 specifically includes:

[0074] S61, inputting the fused feature sequence into a multi-scale feature extraction unit, and performing a forward propagation operation of the one-dimensional convolution channel to obtain a multi-scale fused feature sequence;

[0075] S62: Input the multi-scale fusion feature sequence into a bidirectional gated recurrent unit layer, and obtain a dynamic feature sequence containing forward and backward depth sequence feature information through a forward propagation operation of the bidirectional gated recurrent unit;

[0076] S63. Initialize the dynamic feature selection self-attention layer using the real-time adaptive model parameters and determine the initial attention weight of each attention head in the multi-head self-attention mechanism;

[0077] S64: Input the dynamic feature sequence into the dynamic feature selection self-attention layer, calculate the attention weight of each sequence element through the multi-head self-attention mechanism, and obtain a dynamically weighted feature sequence;

[0078] S65, inputting the dynamically weighted feature sequence in parallel to the short-term prediction sublayer and the long-term prediction sublayer of the multi-task collaborative prediction output layer, performing linear mapping calculations on each prediction sublayer to obtain a short-term prediction value and a long-term prediction value;

[0079] S66: Input the short-term prediction value and the long-term prediction value into the fusion output layer for linear mapping fusion to obtain a final fusion prediction value;

[0080] S67. Compare the final fusion prediction value with the actual observation value, calculate the mean square error loss, and use the stochastic gradient descent algorithm to iteratively optimize the model parameters of the multi-scale feature extraction unit, the bidirectional gated recurrent unit layer, the dynamic feature selection self-attention layer, and the multi-task collaborative prediction output layer through back propagation. The number of iterative training times is 500 times, the learning rate is 0.0005, and the trained model parameters are obtained.

[0081] The beneficial effects of the present invention are:

[0082] (1) The present invention uses the Mamba deep state space model to extract long-term dependency feature sequences of each modal data, thereby accurately capturing the long-term dependency relationships in long-sequence data, effectively improving the accuracy and generalization performance of sequence feature extraction, and enhancing the prediction stability and accuracy of the model in complex environments and long-sequence prediction scenarios.

[0083] (2) The present invention can realize real-time dynamic optimization of model parameters by constructing a diffusion space of adaptive multimodal diffusion model parameters and performing real-time diffusion and reverse denoising convergence processing, significantly improving the adaptability of model parameters to dynamic environmental changes, and showing better adaptability and prediction accuracy in real-time application scenarios of dynamic simulation prediction.

[0084] (3) In terms of multimodal data fusion, the present invention performs dynamic weighted fusion of the long-term dependent feature sequences of each modality based on real-time adaptive model parameters, effectively solving the problems of single modal data fusion mode and low fusion accuracy in the existing technology, breaking through the limitations of traditional static fusion methods, and achieving more accurate feature fusion effects, thereby effectively improving the fusion quality and generalization ability of the multimodal data-driven prediction model.

[0085] (4) The present invention designs a dynamic simulation prediction model framework including a multi-scale feature extraction unit, a bidirectional gated recurrent unit layer, a dynamic feature selection self-attention layer and a multi-task collaborative prediction output layer, which can realize efficient prediction modeling of fused feature sequences, significantly improve the model's multi-scale feature capture capability and prediction performance, and show better comprehensive prediction effects in complex scenarios such as industrial equipment monitoring, traffic flow prediction, and power load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0087] Figure 1 This is a schematic diagram of the overall process of a data-driven dynamic simulation prediction model construction method proposed by the present invention. DETAILED DESCRIPTION

[0088] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0089] refer to Figure 1 , a data-driven dynamic simulation prediction model construction method, comprising the following steps:

[0090] S1. Collecting multi-source heterogeneous multimodal raw data through a data acquisition device, and preprocessing the multimodal raw data to obtain a standardized data set of each modality data;

[0091] S2. Use the Mamba deep state space model to extract sequence features from the standardized data sets of each modality data to obtain the long-term dependency feature sequence of each modality;

[0092] S3. Establishing a parameter diffusion space using an adaptive multimodal diffusion model, inputting a long-term dependent feature sequence into the parameter diffusion space, performing real-time diffusion processing on the parameters, and obtaining characteristic parameters after diffusion;

[0093] S4, performing reverse denoising and convergence processing on the diffused characteristic parameters to generate real-time adaptive model parameters;

[0094] S5. Dynamically weighted fusion of the long-term dependent feature sequences of each modality is performed based on the real-time adaptive model parameters to obtain a fused feature sequence;

[0095] S6. Inputting the fused feature sequence into a preset dynamic simulation prediction model framework, training the dynamic simulation prediction model framework in combination with real-time adaptive model parameters to generate trained model parameters;

[0096] S7. Solidify the trained model parameters into the dynamic simulation prediction model framework to complete the construction of the dynamic simulation prediction model.

[0097] When S7 of the present invention is implemented, after the dynamic simulation prediction model training is completed, all model parameters of the multi-scale feature extraction unit, the bidirectional gated recurrent unit layer, the dynamic feature selection self-attention layer and the multi-task collaborative prediction output layer in the model framework are obtained, and the binary serialization method is clearly used to perform parameter solidification processing. The parameter solidification processing method is specifically completed using the pickle serialization function provided by the Python language. The model parameters include convolution kernel weights, convolution biases, gated recurrent unit weights and hidden state parameters, the attention weight matrix of the self-attention mechanism and the fully connected layer weights and bias parameters of the multi-task prediction output layer; the serialized parameter file is clearly saved in the specified computer system directory in the format of a binary file, and the file is saved in the specified computer system directory in the format of a binary file. The file suffix is ​​clearly ".pkl", and the saving path is specifically specified in the preset storage directory of the computer system; then the model parameter loading function is clearly constructed, and the loading function is specifically implemented using the pickle deserialization method to read the binary parameter file saved in the preset directory, and the loaded parameters are sequentially assigned to the corresponding modules in the multi-scale feature extraction unit, the bidirectional gated recurrent unit layer, the dynamic feature selection self-attention layer and the multi-task collaborative prediction output layer, to achieve a clear reconstruction of the dynamic simulation prediction model framework, thereby completing the construction of a stable dynamic simulation prediction model; the dynamic simulation prediction model after the model parameters are solidified can be directly and clearly applied to the dynamic simulation prediction of real-time data without the need for re-training or parameter adjustment.

[0098] First, raw data from vibration, acoustic, temperature, electrical, visual, and communication modes are collected using a vibration sensor array, an acoustic sensor array, a temperature sensor array, a current sensor array, an image sensor array, and a wireless signal acquisition device. The sampling frequency ranges from 1000 Hz to 5000 Hz, and the acquisition period is 60 seconds. The raw data is filtered, denoised, and normalized to obtain standardized datasets for each modality. Subsequently, a Mamba deep state-space model is constructed for each standardized dataset of modal data. The model consists of one input mapping layer, three selective state-space layers, two residual connection layers, and one output mapping layer in series. The input mapping layer has 256 neurons, and the selective state-space layers have state-space equations of 256, 128, and 64 dimensions, respectively. Each selective state-space layer uses a linear state equation, with the hidden state vector initialized to all zeros, and the initial values ​​of each state-space parameter matrix are initialized from a random normal distribution. Next, a three-layer feedforward neural network was used to construct the parameter diffusion space of the adaptive multimodal diffusion model. The initial feature parameters of this diffusion space followed a standard Gaussian distribution, the diffusion step size was set to 1000 steps, and the diffusion variance adjustment coefficient ranged from 0.0001 to 0.02. During the feature parameter diffusion process, a forward diffusion process was used to generate a diffusion feature parameter sequence. A three-layer feedforward neural network was then used to predict noise and perform a reverse denoising process to obtain the real-time adaptive model parameters. The three-layer feedforward neural network was trained 400 times with a learning rate of 0.001. The real-time fusion weights of the long-term dependent feature sequences of each modality were then determined based on the real-time adaptive model parameters. The fused feature sequence was obtained through element-by-element weighted fusion and calculation of the fused sequence stability index until the fused sequence stability index was less than 0.01. The fused feature sequence is then input into the dynamic simulation prediction model framework, which consists of a multi-scale feature extraction unit, a bidirectional gated recurrent unit layer, a dynamic feature selection self-attention layer, and a multi-task collaborative prediction output layer. The multi-scale feature extraction unit consists of parallel one-dimensional convolution channels with kernel sizes of 3, 5, and 7, with 32 kernels per channel. The bidirectional gated recurrent unit layer consists of two layers, each with 128 hidden units. The self-attention layer consists of eight attention heads. The multi-task collaborative prediction output layer contains two parallel fully connected prediction sublayers for short-term and long-term prediction, respectively, with each sublayer containing 64 neurons. Finally, the fused prediction value is calculated through forward propagation and the mean squared error loss is calculated with the actual observation value. The model parameters are optimized through backpropagation using the stochastic gradient descent algorithm. The training iterations are 500 and the learning rate is 0.0005. The trained model parameters are obtained and fixed into the dynamic simulation prediction model framework, thus completing the construction of the dynamic simulation prediction model.

[0099] The present invention significantly improves the ability to accurately extract long-term dependent feature sequences in multimodal data through the Mamba deep state-space model, solving the problem that traditional models have difficulty in effectively capturing long-sequence features. By constructing an adaptive multimodal diffusion learning model, real-time dynamic optimization of model parameters is achieved, significantly improving the generalization performance and real-time adaptability of the model. Dynamic weight fusion is achieved through real-time adaptive model parameters, effectively improving the accuracy and stability of multimodal data fusion. By designing a prediction framework that includes multi-scale feature extraction, bidirectional sequence modeling, and dynamic feature attention selection, the prediction accuracy and generalization ability are further improved, thereby enhancing the practical application performance of the model in complex dynamic environments.

[0100] In this embodiment, S1 specifically includes:

[0101] S11. Using a vibration sensor array to collect vibration modal data, and performing filtering, noise reduction, and amplitude normalization on the vibration modal data to obtain a standardized vibration modal data set;

[0102] S12. Acoustic modal data is collected using a sound sensor array, and the acoustic modal data is subjected to denoising, feature filtering, and sampling rate adjustment to obtain a standardized acoustic modal data set.

[0103] S13, using a temperature sensor array to collect temperature modal data, and performing outlier removal, smoothing filtering, and scaling processing on the temperature modal data to obtain a standardized temperature modal data set;

[0104] S14, using a current sensor array to collect electrical modal data, and performing denoising filtering, feature extraction, and normalization processing on the electrical modal data to obtain a standardized electrical modal data set;

[0105] S15. Using an image sensor array to collect visual modality data, and performing image enhancement, deblurring, and image size normalization on the visual modality data to obtain a standardized visual modality dataset;

[0106] S16. Using a wireless signal acquisition device to collect communication modal data, and performing signal demodulation, filtering, and amplitude normalization processing on the communication modal data to obtain a standardized communication modal data set;

[0107] S17. When the data collection stage does not set any one or more of vibration modal data, acoustic modal data, temperature modal data, electrical modal data, visual modal data, and communication modal data to be collected, the corresponding modal data position that is not set to be collected is marked with an empty data marker for placeholder processing;

[0108] S18. When data is missing in the modalities that have been set for collection, the missing modal data are reconstructed using the historical modal data interpolation method to obtain a complete standardized data set;

[0109] S19: Perform timestamp alignment and data fusion processing on the standardized data sets of each modality obtained in steps S11 to S18 to obtain a standardized data set of each modality data that is time-synchronized.

[0110] When the present invention is implemented, multiple MEMS acceleration sensors of model ADXL335 are used to collect vibration modal data, with a sensor sensitivity of 300mV / g, and the vibration signal is passed through a low-pass filter with a cutoff frequency set to 500Hz. The threshold denoising method based on wavelet transform is used for denoising, and the amplitude normalization range is 0 to 1; the acoustic modal data is collected by multiple MEMS microphone sensors of model MP34DT01, with a sampling rate set to 44.1kHz, the denoising method adopts spectrum subtraction filtering, and the sampling rate adjustment method is resampling method; when collecting temperature modal data, multiple digital temperature sensors of model DS18B20 are used, with a measurement accuracy of ±0.5℃, the 3σ criterion is used to eliminate outliers, the sliding average method is used for smoothing filtering, and the Min-Max normalization method is used for scale transformation; the electrical modal data is collected by a Hall current sensor of model ACS712, and the sampling The sampling rate is set to 2000 Hz, and the fast Fourier transform (FFT) frequency domain analysis method is used for feature extraction; the visual modal data is collected using an array of multiple OV5640 image sensors, the image enhancement uses the histogram equalization method, the deblurring uses the Wiener filter, and the image size is uniformly scaled to 224×224 pixels; the communication modal data is collected using the NRF24L01 wireless module with a carrier frequency of 2.4 GHz, the signal demodulation uses the GFSK demodulation method, and the amplitude normalization range is 0 to 1; the position of the collected modal data is not set during the data collection stage, and an empty data matrix with a value of 0 is used to clearly occupy the position; for modes with missing data, the historical modal data interpolation method uses the linear interpolation method; the timestamp alignment method uses the highest frequency mode as the benchmark, and the remaining modal data are aligned and fused according to the benchmark mode by linear interpolation to obtain a standardized data set of time-synchronized multimodal data.

[0111] The present invention significantly improves the accuracy of data acquisition and the consistency of data quality through a specifically defined multimodal data acquisition and preprocessing method, ensuring the accuracy and reliability of subsequent feature extraction and model construction; through an interpolation and placeholder mechanism with clear modalities for missing data or uncollected data, it effectively solves the impact of missing data on the generalization performance of the model, enhancing the data processing capability and generalization adaptability of the present invention in actual complex environments.

[0112] In this embodiment, S2 specifically includes:

[0113] S21. Construct a Mamba deep state-space model consisting of one input mapping layer, three selective state-space layers, two residual connection layers, and one output mapping layer connected in series.

[0114] S22. Determine the number of input neurons of the input mapping layer to be 256 according to the data dimension of the standardized data set of each modal data, and use a linear mapping method to map the standardized data set of each modal data to an initial input feature vector;

[0115] S23. Initialize the hidden state vector in the first selective state space layer to an all-zero vector with a dimension of 256, construct a linear state space equation with a state dimension of 256, use the initial input feature vector as input to perform state update, and obtain the first state output vector;

[0116] S24, using the first residual connection layer to superimpose and fuse the first state output vector and the initial input feature vector element by element to obtain a first fused state feature representation, and using the first fused state feature representation as the input of the second selective state space layer;

[0117] S25. Initialize the hidden state vector in the second selective state space layer to the first state output vector with a dimension of 128, and construct a linear state space equation with a state dimension of 128. Use the first fused state feature representation as input to update the state to obtain the second state output vector. Then, use the second residual connection layer to superimpose and fuse the second state output vector with the first fused state feature representation element by element to obtain the second fused state feature representation. Use the second fused state feature representation as the input of the third selective state space layer.

[0118] S26. Initialize the hidden state vector in the third selective state space layer to the second state output vector with a dimension of 64, construct a linear state space equation with a state dimension of 64, use the second fused state feature representation as input to perform state update, and obtain the third state output vector. Then, map the third state output vector to a feature space with a dimension of 64 through the output mapping layer to obtain the sequence feature representation of each modal data.

[0119] S27. Using the sequence feature representation of each modal data as the training target, perform forward propagation and backpropagation iterative training on the Mamba deep state-space model using the standardized data set of each modal data. The number of iterative training times is 300, and the learning rate is 0.0005. The state space parameter matrices of the first to third selective state space layers are updated by the gradient descent algorithm to obtain a converged Mamba deep state-space model.

[0120] S28. Input the standardized data sets of each modality data into the converged Mamba deep state space model respectively to extract sequence features and obtain the long-term dependency feature sequence of each modality.

[0121] When the present invention is implemented, the input mapping layer is first determined based on the data dimension of the fused feature sequence. The input mapping layer uses a linear mapping method to map the fused feature sequence to an initial feature vector. The mapping method is implemented by matrix multiplication, and the initial value of the input mapping matrix is ​​initialized by a random normal distribution. Then, three selective state space layers are constructed. Each selective state space layer uses the following linear state space equation: h t+1 =Ah t +Bu t ,o t =Ch t Among them, the state vector h t The initial value is set to an all-zero vector, and the input vector is u t , the output vector is o t , the initial values ​​of the state space parameter matrices A, B, and C are initialized using the Xavier random initialization method; then two residual connection layers are constructed, the first residual connection layer realizes the element-by-element addition of the output of the first selective state space layer and the initial input feature vector, and the second residual connection layer realizes the element-by-element addition of the output of the second selective state space layer and the first fusion feature representation, and the residual connection process is implemented using element-by-element vector addition; when constructing the output mapping layer, a linear mapping method is used to map the feature vector output by the third selective state space layer to a feature space of a predetermined dimension; when training the model parameters, the feature representation of each modal data sequence is used as the training target, The predicted value is calculated by forward propagation of the standardized data set, and the mean square error loss function between the predicted value and the target value is calculated. The gradient of the model parameters is then calculated and updated by backpropagation of the stochastic gradient descent algorithm. The number of training iterations is set to 300, the learning rate is set to 0.0005, and the momentum optimization strategy is adopted in each iteration. The momentum parameter value is set to 0.9 to improve the convergence speed and stability. After the training is completed, the stably converged model parameters are solidified and saved in the pickle file format to the computer hard disk for subsequent call. Finally, a stable and reliable Mamba deep state space model is obtained, which is used for the accurate extraction of long-term dependent feature sequences of each modality.

[0122] The present invention achieves accurate modeling and effective capture of complex long sequence data in fusion feature sequences by constructing a multi-level Mamba deep state space model and coordinating it with residual connection technology, solving the problem of insufficient sequence feature capture capability of traditional models. By clearly setting the initial hidden state and initializing the parameters, the stability and efficiency of model training are improved, and the generalization ability and prediction accuracy of the model are significantly enhanced. At the same time, through strict training and parameter optimization strategies, the model converges more quickly and stably, further enhancing the application reliability of the model in actual complex environments.

[0123] In this embodiment, S3 specifically includes:

[0124] S31. Construct a parameter diffusion space with a dimension of 64 based on the long-term dependency feature sequence of each modality, wherein the initial probability distribution of any feature parameter in the parameter diffusion space is a standard Gaussian distribution;

[0125] S32. Construct a three-layer feedforward neural network for predicting diffusion process noise, wherein the input layer dimension is 64 neurons, the hidden layer dimension is 128 neurons, and the output layer dimension is 64 neurons. The neural network takes the diffusion characteristic parameter and the number of diffusion steps as input and uses the predicted noise as output;

[0126] S33. Set the diffusion step length of the parameter diffusion space to 1000 steps, and define the state transfer equation of the forward diffusion process as:

[0127]

[0128] Among them, z t represents the diffusion characteristic parameter of the tth step, ∈ represents the random noise sampled from the standard Gaussian distribution, β t It represents the diffusion variance adjustment coefficient of the t-th step and satisfies the linear adjustment method. The value range is 0.0001 to 0.02;

[0129] S34, taking the long-term dependency feature sequence of each mode as the initial input, performing step-by-step iteration through the forward diffusion process, and generating a diffusion feature parameter sequence evolving from low noise to high noise in the parameter diffusion space;

[0130] S35. For the diffusion characteristic parameter sequence, the state transfer equation of the reverse diffusion process is defined as:

[0131]

[0132] in, ∈ θ (z t ,t) is the noise of the t-th step predicted by the neural network, and the neural network parameters θ are obtained through training;

[0133] S36, iteratively optimize the neural network parameters θ in the back diffusion process, the number of training times is 400 times, and the optimization goal is to make the predicted noise ∈ θ (z t ,t) and the mean square error between the real noise ∈ added in the forward diffusion process is minimized;

[0134] S36, through the back diffusion process, the real added noise ∈ and the predicted noise ∈ θ (z t ,t) between the mean square error loss function, the gradient descent algorithm is used to iteratively train the parameters of the three-layer feedforward neural network, the number of training times is 400 times, the learning rate is 0.001, and the three-layer feedforward neural network with training convergence is obtained;

[0135] S37, applying the three-layer feedforward neural network after training convergence in the back diffusion process of the parameter diffusion space in real time to generate a feature parameter sequence after diffusion in real time;

[0136] S38. Extract the feature parameters of the first diffusion step from the feature parameter sequence after diffusion generated in real time as the feature parameters after diffusion.

[0137] When the present invention is implemented, a parameter diffusion space is first constructed based on the long-term dependent feature sequence of each mode. The initial values ​​of the feature parameters in the parameter diffusion space are randomly generated according to the standard Gaussian distribution, and the random seed is fixed to 42 to ensure repeatability; then a three-layer feedforward neural network is clearly established, and the network input is a joint feature vector composed of the diffusion feature parameters of the current step and the corresponding diffusion step number. The number of neurons in the input layer is fixed to 64, the number of hidden layers is fixed to 128, and the number of neurons in the output layer is fixed to 64. The activation function of each neuron adopts the ReLU function, and the initial value of each network parameter is determined by the Xavier initialization method; the value range of the diffusion variance adjustment coefficient β_t is fixed to 0.0001 to 0.02, and a linear adjustment method is used to determine each step in the 1000-step diffusion process; in the forward diffusion process, random noise ε is independently generated through a standard Gaussian distribution, The diffusion characteristic parameter z_t of each step is calculated through a clear state transfer equation; in the reverse diffusion process, the three-layer feedforward neural network constructed above is used to predict the noise value required for the diffusion characteristic parameter of each step, and the network parameters are trained through the mean square error loss function between the network parameters and the actual added random noise ε. The training optimization adopts the stochastic gradient descent algorithm, and the number of training iterations is clearly set to 400 times, the learning rate is clearly set to 0.001, and the momentum optimization parameter is clearly set to 0.9 during the training process to ensure the stability and rapid convergence of the training; the reverse diffusion process predicts the diffusion noise and updates the parameters through the trained neural network, and the final denoised characteristic parameters are obtained after 1000 steps of reverse diffusion; finally, the diffusion characteristic parameters of the first step are clearly taken as the characteristic parameters after diffusion and used as the input of the subsequent steps to ensure the stability and consistency of the diffusion characteristic parameters.

[0138] The present invention effectively realizes the efficient diffusion and precise reverse denoising convergence of diffusion feature parameters by clearly constructing the parameter diffusion space of the adaptive multimodal diffusion learning algorithm, thereby improving the accuracy and stability of real-time optimization of model feature parameters; the clear neural network structure and parameter optimization method improve the prediction accuracy and generalization performance of the diffusion model, effectively solves the problem of unstable parameter optimization of traditional diffusion methods, and significantly enhances the ability of model parameters to adapt to dynamic environmental changes in real time.

[0139] In this embodiment, the S4 specifically includes:

[0140] S41. Construct a reverse denoising convergence space with a dimension of 64 based on the characteristic parameters after diffusion, and initialize the convergence step size in the reverse denoising process to 1000 steps;

[0141] S42. Define the state transfer equation of the reverse denoising process as:

[0142]

[0143] in, Represents the characteristic parameters after the t-th step reverse denoising process. Initially is the characteristic parameter after diffusion, γ t It represents the denoising coefficient in the t-th step reverse denoising process, and its value range is 0.0001 to 0.02. represents the predicted noise of the t-th step reverse denoising;

[0144] S43. Construct a two-layer fully connected neural network for predicting the noise of the reverse denoising process, with an input layer dimension of 64 neurons and an output layer dimension of 64 neurons, taking the characteristic parameters of the reverse denoising and the number of denoising steps as input, and predicting the denoised noise;

[0145] S44. Using a standardized dataset of historically collected data from each modality, the mean square error between the actual added noise and the predicted denoised noise during the reverse denoising process was used as the loss function. The parameters of the two-layer fully connected neural network were iteratively trained using the stochastic gradient descent algorithm. The number of training times was 300 and the learning rate was 0.001. A converged two-layer fully connected neural network was obtained.

[0146] S45, using a two-layer fully connected neural network after training convergence to predict the noise of the reverse denoising process in real time, executing the reverse denoising process, and generating a denoising feature parameter sequence evolving from high noise to low noise in real time;

[0147] S46. Extracting the feature parameters of the first step of reverse denoising from the denoising feature parameter sequence generated in real time as real-time adaptive model parameters.

[0148] In the specific implementation of the present invention, a reverse denoising convergence space is first constructed according to the characteristic parameters after diffusion. The characteristic parameters in the initial reverse denoising process are explicitly initialized by the characteristic parameters after diffusion. A two-layer fully connected neural network is explicitly constructed as a denoising prediction model for the reverse denoising process. The number of neurons in the input layer and the output layer of the neural network is 64. No additional activation function is set between the hidden layer and the output layer. The initial parameters of the network are determined by the Xavier random initialization method, and the random seeds of the network weights and bias parameters are fixed to 42. A fixed step size of 1000 steps is explicitly adopted in the reverse denoising process, and the state equation is explicitly adopted in the state transfer process:

[0149]

[0150] The denoising coefficient γ tThe value range is clearly set to 0.0001 to 0.02, and the linear change method is used to determine the specific value of each step; the standardized data set of historically collected modal data is used in the training process of the neural network. The network training uses the mean square error between the real added noise and the network predicted noise as a clear loss function, and the stochastic gradient descent algorithm is used to optimize the network parameters. The number of training iterations is clearly set to 300 times, the learning rate is clearly set to 0.001, and the momentum optimization method is clearly used with a momentum parameter of 0.9 to ensure the stability and rapidity of the convergence of the training process; the two-layer fully connected neural network that has converged after training is applied to the reverse denoising process, and each step of the reverse denoising process explicitly uses the predicted noise to denoise and update the current feature parameters; after 1000 steps of denoising processing, the feature parameters of the first step of reverse denoising are explicitly extracted from the generated denoising feature sequence as the final real-time adaptive model parameters for explicit use in the subsequent fusion process.

[0151] The present invention effectively realizes accurate and efficient reverse denoising and convergence processing of the characteristic parameters after diffusion by constructing a clear reverse denoising convergence space and coordinating it with a stable two-layer fully connected neural network structure, thereby improving the real-time optimization accuracy of the adaptive model parameters; the clear state transition equation and parameter initialization method enhance the stability and reliability of the reverse denoising process, solves the problems of slow parameter convergence and poor noise removal effect in traditional methods, and significantly improves the real-time adaptability and prediction accuracy of the model in a dynamic data environment.

[0152] In this embodiment, the S5 specifically includes:

[0153] S51, determining a real-time fusion weight corresponding to each modal feature sequence based on real-time adaptive model parameters;

[0154] S52, using the real-time fusion weight to perform weighted processing on the long-term dependent feature sequence of each modality to obtain a weighted feature sequence of each modality;

[0155] S53, according to the timestamp information of the feature sequence, performing a fusion operation on the sequence elements of the weighted feature sequence of each modality according to the corresponding timestamp to obtain an initial fused feature sequence;

[0156] S54, calculating the stability index of the initial fusion feature sequence, and determining whether the fusion sequence meets the preset stability condition;

[0157] S55: If the fused sequence does not meet the preset stability condition, the real-time fusion weights of each modal feature sequence are updated based on the real-time adaptive model parameters, and the process returns to step S52 to perform dynamic weighted fusion of the feature sequence again.

[0158] S56. If the fused sequence meets the preset stability condition, the current fused feature sequence is output as the final fused feature sequence.

[0159] During the specific implementation of the present invention, the real-time fusion weight corresponding to each modal feature sequence is first determined according to the real-time adaptive model parameters. The calculation method of each fusion weight value is to perform a clear normalization calculation on the real-time adaptive model parameters using the Softmax function to ensure that the sum of the weights is 1; next, the real-time fusion weight obtained by the clear calculation is used to perform element-by-element weighting processing on the long-term dependent feature sequence of each modality. The weighting process is clearly implemented using the element-by-element multiplication method; then, the timestamps of the feature sequences of each modality are clearly used as the benchmark, and the weighted feature sequences of each modality are clearly fused element-by-element. The fusion process adopts the element-by-element addition method to obtain the initial fusion feature sequence. feature sequence; then, the variance of the fused sequence is clearly adopted as the stability indicator, and the specific calculation method of the variance is clearly defined as the average value of the square of the difference between each sequence element and the sequence mean; the clear threshold value of the fused sequence stability indicator is further determined to be 0.01, and it is clearly judged whether the fused sequence meets the stability condition; if the fused sequence does not meet the stability condition, the real-time fusion weights of each modal feature sequence are recalculated and updated based on the real-time adaptive model parameters, and the weighted fusion step is returned to be re-executed; when the fused sequence reaches the clear stability threshold, the current fused feature sequence is clearly output as a stable input sequence for subsequent dynamic simulation prediction model training and prediction of the present invention.

[0160] The present invention dynamically adjusts the multimodal data fusion weights based on clearly calculated real-time adaptive model parameters, thereby achieving accurate dynamic fusion of long-term dependent feature sequences, effectively overcoming the defect that traditional static fusion methods are difficult to adapt to changes in dynamic data environments; explicitly adopting the fusion sequence variance as a stability indicator to control the fusion process, effectively improving the stability and consistency of the fusion sequence, significantly improving the generalization performance and prediction accuracy of the dynamic simulation prediction model, and enhancing the application adaptability of the present invention in actual complex dynamic scenarios.

[0161] In this embodiment, S6 specifically includes:

[0162] S61, inputting the fused feature sequence into a multi-scale feature extraction unit, and performing a forward propagation operation of the one-dimensional convolution channel to obtain a multi-scale fused feature sequence;

[0163] S62: Input the multi-scale fusion feature sequence into a bidirectional gated recurrent unit layer, and obtain a dynamic feature sequence containing forward and backward depth sequence feature information through a forward propagation operation of the bidirectional gated recurrent unit;

[0164] S63. Initialize the dynamic feature selection self-attention layer using the real-time adaptive model parameters and determine the initial attention weight of each attention head in the multi-head self-attention mechanism;

[0165] S64: Input the dynamic feature sequence into the dynamic feature selection self-attention layer, calculate the attention weight of each sequence element through the multi-head self-attention mechanism, and obtain a dynamically weighted feature sequence;

[0166] S65, inputting the dynamically weighted feature sequence in parallel to the short-term prediction sublayer and the long-term prediction sublayer of the multi-task collaborative prediction output layer, performing linear mapping calculations on each prediction sublayer to obtain a short-term prediction value and a long-term prediction value;

[0167] S66: Input the short-term prediction value and the long-term prediction value into the fusion output layer for linear mapping fusion to obtain a final fusion prediction value;

[0168] S67. Compare the final fusion prediction value with the actual observation value, calculate the mean square error loss, and use the stochastic gradient descent algorithm to iteratively optimize the model parameters of the multi-scale feature extraction unit, the bidirectional gated recurrent unit layer, the dynamic feature selection self-attention layer, and the multi-task collaborative prediction output layer through back propagation. The number of iterative training times is 500 times, the learning rate is 0.0005, and the trained model parameters are obtained.

[0169] The preset dynamic simulation prediction model framework specifically includes:

[0170] A multi-scale feature extraction unit is constructed, which includes three parallel one-dimensional convolution channels with convolution kernel sizes of 3, 5, and 7, the number of convolution kernels is 32, and the convolution step size is 1. It is used to extract local features of different scales from the input fused feature sequence and obtain a multi-scale fused feature sequence by element-by-element splicing;

[0171] Constructing a bidirectional gated recurrent unit layer, including two layers of bidirectional gated recurrent units connected in series, each layer of bidirectional gated recurrent units containing 128 hidden units, for performing forward and backward deep sequence feature modeling on the multi-scale fusion feature sequence;

[0172] Construct a dynamic feature selection self-attention layer, including a multi-head self-attention mechanism with 8 self-attention heads and 128 dimensions. The attention weight of each sequence element is dynamically calculated based on the real-time adaptive model parameters to highlight the key features in the fused feature sequence that have an important impact on the dynamic prediction results.

[0173] Construct a multi-task collaborative prediction output layer, which includes two parallel fully connected prediction sublayers. The first prediction sublayer is used for short-term prediction and contains 64 neurons. The second prediction sublayer is used for long-term prediction and contains 64 neurons. Both prediction sublayers receive features output from the attention layer in parallel.

[0174] Constructing a fusion output layer, performing linear mapping fusion on the prediction results of the two prediction sub-layers in the multi-task collaborative prediction output layer to obtain the final fusion prediction numerical output;

[0175] The fused feature sequence and real-time adaptive model parameters are input into the dynamic simulation prediction model framework. Through forward propagation and backpropagation iterative optimization, the model parameters of the multi-scale feature extraction unit, bidirectional gated recurrent unit layer, dynamic feature selection self-attention layer and multi-task collaborative prediction output layer are trained to generate trained model parameters.

[0176] In the specific implementation of the present invention, the fused feature sequence is first input into the multi-scale feature extraction unit, and three parallel one-dimensional convolution channels are explicitly constructed. Each channel uses convolution kernel sizes of 3, 5 and 7, the number of convolution kernels is fixed to 32, and the convolution step size is explicitly set to 1. After the convolution operation, feature maps of different scales are obtained respectively, and then a multi-scale fusion feature sequence is explicitly obtained through an element-by-element splicing operation; then the multi-scale fusion feature sequence is input into a clearly constructed bidirectional gated recurrent unit layer, and the bidirectional gated recurrent unit layer explicitly includes two bidirectional GRU layers in series, the number of hidden units in each layer is fixed to 128, the initial hidden state of each bidirectional GRU unit is explicitly set to a zero vector, and the parameters of each GRU unit are explicitly determined using an orthogonal initialization method; next, the dynamic feature selection self-attention layer is initialized using the explicitly given real-time adaptive model parameters, the multi-head self-attention mechanism explicitly includes 8 attention heads, the dimension of each attention head is fixed to 128, and the self-attention mechanism is explicitly determined by a clear scaling point The attention weights are calculated by the product method to realize dynamic weighting of the input feature sequence; then, the dynamically weighted feature sequence is explicitly input into the short-term prediction sublayer and the long-term prediction sublayer in parallel. The two prediction sublayers are explicitly implemented by the linear mapping method, and the number of neurons is fixed at 64; further, the short-term prediction values ​​and the long-term prediction values ​​are explicitly input into the fusion output layer for explicit linear mapping fusion, and the fusion output layer mapping method is implemented by the fully connected layer; finally, by calculating the explicit mean square error loss function between the fusion prediction value and the actual observation value, the stochastic gradient descent algorithm is adopted, and the parameters of the multi-scale feature extraction unit, the bidirectional gated recurrent unit layer, the dynamic feature selection self-attention layer and the multi-task collaborative prediction output layer are iteratively optimized through explicit back propagation. The number of training iterations is explicitly set to 500 times, the learning rate is explicitly set to 0.0005, the momentum optimization strategy is explicitly adopted in the training process, and the momentum parameter is set to 0.9. Finally, the trained model parameters are explicitly obtained for subsequent dynamic simulation prediction.

[0177] The present invention achieves accurate modeling and dynamic prediction of multi-scale and long- and short-time features in fused feature sequences by clearly constructing a dynamic simulation prediction model framework of multi-scale feature extraction units, bidirectional sequence feature modeling, adaptive dynamic feature attention selection, and multi-task prediction output, effectively solving the problem that traditional prediction methods are difficult to simultaneously take into account different scale features and long- and short-time predictions; clear training optimization strategies and parameter settings further improve the model's training efficiency, prediction accuracy, and generalization performance, and significantly enhance the model's application effect and reliability in complex dynamic data environments.

[0178] Example 1:

[0179] In order to verify the feasibility of the present invention, the present invention is applied to the data monitoring and predictive management scenario of a large intelligent manufacturing enterprise. The enterprise has multiple automated production workshops, rich equipment types, complex operating environment, and the equipment operating status is closely related to production quality. In recent years, with the rapid development of intelligent manufacturing, the enterprise has faced huge challenges in equipment status monitoring and predictive maintenance, especially the various types of monitoring data, large data scale, and difficulty in data fusion. Traditional status monitoring and prediction methods are difficult to effectively process such a large scale of multimodal data, resulting in low accuracy of equipment abnormality warning, high false alarm rate and inability to achieve real-time adaptive optimization. For this reason, the enterprise decided to deploy the data-driven dynamic simulation prediction model construction method proposed in the present invention to improve the prediction accuracy and real-time monitoring of equipment status.

[0180] During implementation, multimodal data from production workshop equipment is first collected in real time using multiple different types of sensors, including equipment vibration data, sound signals, temperature data, electrical signals, visual data, and wireless communication signals. Specifically, vibration data is collected using MEMS accelerometers, sound signals using MEMS microphones, temperature data using digital temperature sensors, electrical signals using Hall effect current sensors, visual data using a high-precision image sensor array, and communication signals using wireless modules. Each modal data set undergoes filtering, denoising, normalization, and standardization preprocessing to ensure consistency and reliability of data quality. A standardized multimodal dataset is then constructed for subsequent processing.

[0181] Next, each modal data is input into the Mamba deep state space model designed by the present invention. When extracting long-term dependent features, this model can fully capture the long-term dynamic relationship between different modal data and generate a more representative long-term dependent feature sequence. In order to achieve real-time adaptive parameter optimization, the adaptive multimodal diffusion learning algorithm designed by the present invention is adopted to construct a parameter diffusion space based on the long-term dependent feature sequence. Through diffusion and reverse denoising convergence processing, high-quality adaptive model parameters are generated in real time. Using the real-time adaptive model parameters, the modal feature sequences are dynamically weighted and fused to generate a fused feature sequence. Thereafter, the fused feature sequence enters the dynamic simulation prediction model framework. The framework uses a multi-scale feature extraction unit for local feature extraction, a bidirectional gated recurrent unit to capture the previous and next sequence information, a self-attention mechanism to dynamically adjust the feature weights, and a multi-task collaborative prediction unit to achieve collaborative output of short-term and long-term predictions. During the training process, the model parameters are reversely optimized by the difference between the real-time observation data and the model prediction value. After 500 iterative training, the final model parameters are solidified and used for real-time prediction. After three months of implementation, the system showed a significant improvement in prediction effect. The detailed data is shown in Table 1:

[0182] Table 1: Comparison of equipment status prediction and maintenance performance before and after implementation

[0183]

[0184] A detailed comparison of the data in Table 1 shows that the accuracy, false alarm rate, number of downtimes, and response time of equipment anomaly predictions before and after the implementation of the present invention have been significantly improved. The accuracy of anomaly predictions increased from 82.6% before implementation to 97.8% after implementation, and the false alarm rate decreased from 6.4% to 1.3%. The number of monthly equipment downtimes due to failures also decreased significantly, from an average of 15 times per month to an average of 3 times per month. Furthermore, because the model uses a real-time adaptive optimization mechanism, the average maintenance response time has been significantly reduced, from 55 minutes before implementation to 7 minutes, greatly improving maintenance efficiency and reducing economic losses caused by production interruptions.

[0185] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: The following steps are involved: S1. Collecting multi-source heterogeneous multimodal raw data through a data acquisition device, and preprocessing the multimodal raw data to obtain a standardized data set of each modality data; S2. Use the Mamba deep state space model to extract sequence features from the standardized data sets of each modality data to obtain the long-term dependency feature sequence of each modality; S3. Establishing a parameter diffusion space using an adaptive multimodal diffusion model, inputting a long-term dependent feature sequence into the parameter diffusion space, performing real-time diffusion processing on the parameters, and obtaining characteristic parameters after diffusion; S4, performing reverse denoising and convergence processing on the diffused characteristic parameters to generate real-time adaptive model parameters; S5. Dynamically weighted fusion of the long-term dependent feature sequences of each modality is performed based on the real-time adaptive model parameters to obtain a fused feature sequence; S6. Inputting the fused feature sequence into a preset dynamic simulation prediction model framework, training the dynamic simulation prediction model framework in combination with real-time adaptive model parameters to generate trained model parameters; S7. Solidify the trained model parameters into the dynamic simulation prediction model framework to complete the construction of the dynamic simulation prediction model.

2. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: Said S1 specifically includes: S11. Using a vibration sensor array to collect vibration modal data, and performing filtering, noise reduction, and amplitude normalization on the vibration modal data to obtain a standardized vibration modal data set; S12. Acoustic modal data is collected using an acoustic sensor array, and denoising, feature filtering, and sampling rate adjustment are performed on the acoustic modal data to obtain a standardized acoustic modal data set; S13, using a temperature sensor array to collect temperature modal data, and performing outlier removal, smoothing filtering, and scaling processing on the temperature modal data to obtain a standardized temperature modal data set; S14, using a current sensor array to collect electrical modal data, and performing denoising filtering, feature extraction, and normalization processing on the electrical modal data to obtain a standardized electrical modal data set; S15. Using an image sensor array to collect visual modality data, and performing image enhancement, deblurring, and image size normalization on the visual modality data to obtain a standardized visual modality dataset; S16. Using a wireless signal acquisition device to collect communication modal data, and performing signal demodulation, filtering, and amplitude normalization processing on the communication modal data to obtain a standardized communication modal data set; S17. When the data collection stage does not set any one or more of vibration modal data, acoustic modal data, temperature modal data, electrical modal data, visual modal data, and communication modal data to be collected, the corresponding modal data position that is not set to be collected is marked with an empty data marker for placeholder processing; S18. When data is missing in the modalities that have been set for collection, the missing modal data are reconstructed using the historical modal data interpolation method to obtain a complete standardized data set; S19: Perform timestamp alignment and data fusion processing on the standardized data sets of each modality obtained in steps S11 to S18 to obtain a standardized data set of each modality data that is time-synchronized.

3. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: The S2 specifically includes: S21. Construct a Mamba deep state-space model consisting of one input mapping layer, three selective state-space layers, two residual connection layers, and one output mapping layer connected in series. S22. Determine the number of input neurons of the input mapping layer to be 256 according to the data dimension of the standardized data set of each modal data, and use a linear mapping method to map the standardized data set of each modal data to an initial input feature vector; S23. Initialize the hidden state vector in the first selective state space layer to an all-zero vector with a dimension of 256, construct a linear state space equation with a state dimension of 256, use the initial input feature vector as input to perform state update, and obtain the first state output vector; S24, using the first residual connection layer to superimpose and fuse the first state output vector and the initial input feature vector element by element to obtain a first fused state feature representation, and using the first fused state feature representation as the input of the second selective state space layer; S25. Initialize the hidden state vector in the second selective state space layer to the first state output vector with a dimension of 128, and construct a linear state space equation with a state dimension of 128. Use the first fused state feature representation as input to update the state to obtain the second state output vector. Then, use the second residual connection layer to superimpose and fuse the second state output vector with the first fused state feature representation element by element to obtain the second fused state feature representation. Use the second fused state feature representation as the input of the third selective state space layer. S26. Initialize the hidden state vector in the third selective state space layer to the second state output vector with a dimension of 64, construct a linear state space equation with a state dimension of 64, use the second fused state feature representation as input to perform state update, and obtain the third state output vector. Then, map the third state output vector to a feature space with a dimension of 64 through the output mapping layer to obtain the sequence feature representation of each modal data. S27. Using the sequence feature representation of each modal data as the training target, perform forward propagation and backpropagation iterative training on the Mamba deep state-space model using the standardized data set of each modal data. The number of iterative training times is 300, and the learning rate is 0.0005. The state space parameter matrices of the first to third selective state space layers are updated by the gradient descent algorithm to obtain a converged Mamba deep state-space model. S28. Input the standardized data sets of each modality data into the converged Mamba deep state space model respectively to extract sequence features and obtain the long-term dependency feature sequence of each modality.

4. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: The S3 specifically includes: S31. Construct a parameter diffusion space with a dimension of 64 based on the long-term dependency feature sequence of each modality, wherein the initial probability distribution of any feature parameter in the parameter diffusion space is a standard Gaussian distribution; S32. Construct a three-layer feedforward neural network for predicting diffusion process noise, wherein the input layer dimension is 64 neurons, the hidden layer dimension is 128 neurons, and the output layer dimension is 64 neurons. The neural network takes the diffusion characteristic parameter and the number of diffusion steps as input and uses the predicted noise as output; S33. Set the diffusion step length of the parameter diffusion space to 1000 steps, and define the state transfer equation of the forward diffusion process as: Among them, z t represents the diffusion characteristic parameter of the tth step, ∈ represents the random noise sampled from the standard Gaussian distribution, β t It represents the diffusion variance adjustment coefficient of the t-th step and satisfies the linear adjustment method. The value range is 0.0001 to 0.02; S34, taking the long-term dependency feature sequence of each mode as the initial input, performing step-by-step iteration through the forward diffusion process, and generating a diffusion feature parameter sequence evolving from low noise to high noise in the parameter diffusion space; S35. For the diffusion characteristic parameter sequence, the state transfer equation of the reverse diffusion process is defined as: in, ∈ θ (z t ,t) is the noise of the tth step predicted by the neural network, and the neural network parameters θ are obtained through training; S36, iteratively optimize the neural network parameters θ in the back diffusion process, the number of training times is 400 times, and the optimization goal is to make the predicted noise ∈ θ (z t ,t) and the mean square error between the real noise ∈ added in the forward diffusion process is minimized; S36, through the back diffusion process, the real added noise ∈ and the predicted noise ∈ θ (z t ,t) between the mean square error loss function, the gradient descent algorithm is used to iteratively train the parameters of the three-layer feedforward neural network, the number of training times is 400 times, the learning rate is 0.001, and the three-layer feedforward neural network with training convergence is obtained; S37, applying the three-layer feedforward neural network after training convergence in the back diffusion process of the parameter diffusion space in real time to generate a feature parameter sequence after diffusion in real time; S38. Extract the feature parameters of the first diffusion step from the feature parameter sequence after diffusion generated in real time as the feature parameters after diffusion.

5. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: The S4 specifically includes: S41. Construct a reverse denoising convergence space with a dimension of 64 based on the characteristic parameters after diffusion, and initialize the convergence step size in the reverse denoising process to 1000 steps; S42. Define the state transfer equation of the reverse denoising process as: in, Represents the characteristic parameters after the t-th step reverse denoising process. Initially is the characteristic parameter after diffusion, γ t It represents the denoising coefficient in the t-th step reverse denoising process, and its value range is 0.0001 to 0.

02. represents the predicted noise of the t-th step reverse denoising; S43. Construct a two-layer fully connected neural network for predicting the noise of the reverse denoising process, with an input layer dimension of 64 neurons and an output layer dimension of 64 neurons, taking the characteristic parameters of the reverse denoising and the number of denoising steps as input, and predicting the denoised noise; S44. Using a standardized dataset of historically collected data from each modality, the mean square error between the actual added noise and the predicted denoised noise during the reverse denoising process was used as the loss function. The parameters of the two-layer fully connected neural network were iteratively trained using the stochastic gradient descent algorithm. The number of training times was 300 and the learning rate was 0.

001. A converged two-layer fully connected neural network was obtained. S45, using a two-layer fully connected neural network after training convergence to predict the noise of the reverse denoising process in real time, executing the reverse denoising process, and generating a denoising feature parameter sequence evolving from high noise to low noise in real time; S46. Extracting the feature parameters of the first step of reverse denoising from the denoising feature parameter sequence generated in real time as real-time adaptive model parameters.

6. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: The S5 specifically includes: S51, determining a real-time fusion weight corresponding to each modal feature sequence based on real-time adaptive model parameters; S52, using the real-time fusion weight to perform weighted processing on the long-term dependent feature sequence of each modality to obtain a weighted feature sequence of each modality; S53, according to the timestamp information of the feature sequence, performing a fusion operation on the sequence elements of the weighted feature sequence of each modality according to the corresponding timestamp to obtain an initial fused feature sequence; S54, calculating the stability index of the initial fusion feature sequence, and determining whether the fusion sequence meets the preset stability condition; S55: If the fused sequence does not meet the preset stability condition, the real-time fusion weights of each modal feature sequence are updated based on the real-time adaptive model parameters, and the process returns to step S52 to perform dynamic weighted fusion of the feature sequence again. S56. If the fused sequence meets the preset stability condition, the current fused feature sequence is output as the final fused feature sequence.

7. A method for constructing a dynamic simulation prediction model based on data-driven, characterized in that: The preset dynamic simulation prediction model framework specifically includes: A multi-scale feature extraction unit is constructed, which includes three parallel one-dimensional convolution channels with convolution kernel sizes of 3, 5, and 7, the number of convolution kernels is 32, and the convolution step size is 1. It is used to extract local features of different scales from the input fused feature sequence and obtain a multi-scale fused feature sequence by element-by-element splicing; Constructing a bidirectional gated recurrent unit layer, including two layers of bidirectional gated recurrent units connected in series, each layer of bidirectional gated recurrent units containing 128 hidden units, for performing forward and backward deep sequence feature modeling on the multi-scale fusion feature sequence; Construct a dynamic feature selection self-attention layer, including a multi-head self-attention mechanism with 8 self-attention heads and 128 dimensions. The attention weight of each sequence element is dynamically calculated based on the real-time adaptive model parameters to highlight the key features in the fused feature sequence that have an important impact on the dynamic prediction results. Construct a multi-task collaborative prediction output layer, which includes two parallel fully connected prediction sublayers. The first prediction sublayer is used for short-term prediction and contains 64 neurons. The second prediction sublayer is used for long-term prediction and contains 64 neurons. Both prediction sublayers receive features output from the attention layer in parallel. Constructing a fusion output layer, performing linear mapping fusion on the prediction results of the two prediction sub-layers in the multi-task collaborative prediction output layer to obtain the final fusion prediction numerical output; The fused feature sequence and real-time adaptive model parameters are input into the dynamic simulation prediction model framework. Through forward propagation and backpropagation iterative optimization, the model parameters of the multi-scale feature extraction unit, bidirectional gated recurrent unit layer, dynamic feature selection self-attention layer and multi-task collaborative prediction output layer are trained to generate trained model parameters.

8. A method for constructing a dynamic simulation prediction model based on data driving, characterized in that: The S6 specifically includes: S61, inputting the fused feature sequence into a multi-scale feature extraction unit, and obtaining a multi-scale fused feature sequence through a forward propagation operation of the one-dimensional convolution channel; S62: Input the multi-scale fusion feature sequence into a bidirectional gated recurrent unit layer, and obtain a dynamic feature sequence containing forward and backward depth sequence feature information through a forward propagation operation of the bidirectional gated recurrent unit; S63. Initialize the dynamic feature selection self-attention layer using the real-time adaptive model parameters and determine the initial attention weight of each attention head in the multi-head self-attention mechanism; S64: Input the dynamic feature sequence into the dynamic feature selection self-attention layer, calculate the attention weight of each sequence element through the multi-head self-attention mechanism, and obtain a dynamically weighted feature sequence; S65, inputting the dynamically weighted feature sequence in parallel to the short-term prediction sublayer and the long-term prediction sublayer of the multi-task collaborative prediction output layer, performing linear mapping calculations on each prediction sublayer to obtain a short-term prediction value and a long-term prediction value; S66: Input the short-term prediction value and the long-term prediction value into the fusion output layer for linear mapping fusion to obtain a final fusion prediction value; S67. Compare the final fusion prediction value with the actual observation value, calculate the mean square error loss, and use the stochastic gradient descent algorithm to iteratively optimize the model parameters of the multi-scale feature extraction unit, the bidirectional gated recurrent unit layer, the dynamic feature selection self-attention layer, and the multi-task collaborative prediction output layer through back propagation. The number of iterative training times is 500 times, the learning rate is 0.0005, and the trained model parameters are obtained.

Images