A heterogeneous dual-flow timing signal classification method and system based on MPI closed-form optimization

Abstract

This invention discloses a heterogeneous dual-stream time-series signal classification method and system based on MPI closed-form optimization. Addressing the issues of gradient descent's tendency to reach local optima, skip connection noise contamination, and negative migration in heterogeneous dual-stream classification, this invention inputs the preprocessed one-dimensional time-series signal into two parallel feature extraction streams. The gradients of the first and second feature extraction streams are independent, and their feature paradigms are heterogeneous, extracting local structural features and multi-scale time-series features respectively. Under the condition that the gradients of both streams are truncated, they are fused to obtain a joint feature matrix. This joint feature matrix is ​​then input into a multi-layer classifier. Each layer uses a randomly orthogonally initialized and never-updated hidden layer mapping matrix. The classification weights are solved layer by layer using MPI closed-form optimization and accumulated for output, with no gradient calculation throughout the classification process. This invention overcomes three core defects simultaneously at the paradigm optimization level, significantly improving classification accuracy and reproducibility.

Description

Technical Field

[0001] This invention belongs to the field of time-series signal processing technology, and in particular relates to a heterogeneous dual-stream time-series signal classification method and system based on MPI closed-loop optimization. Background Technology

[0002] One-dimensional time-series signal classification is a core task in fields such as human-computer interaction, medical diagnosis, and motion recognition. Typical applications include emotion recognition based on EEG (Electroencephalogram), arrhythmia detection based on ECG (Electrocardiogram), and human activity recognition based on IMU (Inertial Measurement Unit). These signals are characterized by high noise, high dimensionality, and non-stationarity, which places stringent demands on the robustness and generalization ability of classification models.

[0003] The existing technical solutions mainly have the following three core defects: Defect 1: Gradient descent optimization gets stuck in local optima. Traditional deep neural networks (such as 1D-CNN and DeepConvLSTM) rely on gradient descent algorithms like Adam for end-to-end parameter updates. Because the error surface for time-series signal classification tasks is highly non-convex, the classification head is prone to convergence to local optima, leading to unstable classification performance. Furthermore, the results of the same network under different random initializations vary significantly, resulting in poor reproducibility.

[0004] Defect 2: U-Net Skip Connections Introduce Noise Pollution. The U-Net architecture uses skip connections to directly concatenate shallow features from each layer of the encoder to the corresponding layer of the decoder to preserve multi-scale temporal context information. However, in one-dimensional high-noise temporal signals, shallow features contain a large amount of high-frequency transient noise. Direct concatenation will indiscriminately transmit this noise to the deep semantic space, leading to feature degradation. At the same time, deep gradients propagate back to the shallow layers through skip connections, destroying the physical representation of shallow features and interfering with the encoder's multi-scale feature learning process.

[0005] Defect 3: Negative Transfer Caused by Heterogeneous Two-Stream Feature-Level Fusion. For time-series signals containing both local structural features (such as the QRS waveform topology of ECG) and multi-scale temporal context features, using a two-stream network for heterogeneous feature extraction is a natural choice. However, the two feature streams differ significantly in physical meaning, statistical distribution, and feature dimensionality: local structural features originate from spatial topology modeling via two-dimensional convolution, while multi-scale temporal features originate from sequence context modeling via one-dimensional convolution. If feature-level concatenation is directly performed and a classifier is trained using gradient methods, the gradients from the two heterogeneous features interfere with each other, leading to "negative feature transfer"—the fused performance is actually lower than the optimal single-stream performance. This is the fundamental reason why current technologies have been forced to abandon feature-level fusion and instead adopt decision-level fusion.

[0006] In summary, existing one-dimensional time-series signal classification techniques suffer from several drawbacks. Gradient descent optimization is prone to getting trapped in local optima, leading to performance instability. U-Net skip connections introduce noise pollution and interfere with shallow feature learning, while heterogeneous dual-stream feature-level fusion suffers from negative transfer problems, resulting in fused performance that is actually worse than single-stream fusion. Therefore, there is an urgent need for a time-series signal classification method and system that can simultaneously overcome these shortcomings and improve classification accuracy and reproducibility. Summary of the Invention

[0007] To address the above technical problems, this invention provides a heterogeneous dual-stream timing signal classification method and system based on MPI closed-loop optimization.

[0008] The technical solution adopted by this invention to solve its technical problem is: A heterogeneous dual-stream time-series signal classification method based on MPI closed-loop optimization includes the following steps: S100: Acquire a one-dimensional time-series signal and preprocess it to obtain standardized input samples; S200: The preprocessed signal is input in parallel to the first feature extraction stream and the second feature extraction stream to extract the local structural feature matrix and the multi-scale temporal feature matrix, respectively; wherein, the two feature extraction processes maintain gradient independence and the feature extraction paradigms of the two feature extraction streams are heterogeneous. S300: Under the condition that the gradient calculation graphs of the local structural feature matrix and the multi-scale temporal feature matrix have been truncated, the local structural feature matrix and the multi-scale temporal feature matrix are fused after standardization to obtain the joint feature matrix. S400: The joint feature matrix is ​​input into a multi-layer classifier. Each layer of the multi-layer classifier contains a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. The classification weights are solved layer by layer using the closed-form solution of MPI. The classification result is obtained by accumulating the prediction outputs of each layer. The classification process of the multi-layer classifier is a gradient-free calculation process.

[0009] Preferably, the first feature extraction stream in S200 employs a two-stage decoupling optimization strategy to extract the local structural feature matrix. In the first stage, a deep convolutional neural network is trained using a gradient descent algorithm. The encoder of the deep convolutional network consists of multiple Conv2DTranspose convolutional layers, used to progressively unfold the one-dimensional input into a two-dimensional latent space. The decoder consists of a single Conv2D convolutional layer plus a global average pooling layer, used to compress the two-dimensional representation back into a fixed-dimensional feature vector. In the second stage, the weights of the trained network convolutional layers are frozen and the gradient computation graph is truncated. Under the premise of fixed feature representation, the optimization of the classification head is transformed into a ridge regression convex optimization problem. The globally optimal closed-form solution is obtained by solving the Moore-Penrose inverse matrix, which serves as the globally optimal classification weight matrix. .

[0010] Preferably, in the second stage, the objective function of the ridge regression convex optimization problem is defined as: (1) in, For the characteristic matrix, Let be the classification weight matrix to be determined. For the real label matrix, This is the regularization penalty coefficient. For Tikhonov regularization terms; for formula (1) with respect to Find the partial derivative and set it to zero: (2) Simplify formula (2) and extract... : (3) Multiplying both sides by the inverse matrix on the left yields the globally optimal closed-form solution: (4) in It is the identity matrix. Through grid search Optimize within the given range.

[0011] Preferably, the second feature extraction stream adopts a one-dimensional U-shaped network with closed adaptive skip connections. The U-shaped network includes multiple encoders and corresponding number of decoders. Each encoder is followed by max pooling for downsampling, and each decoder is followed by transposed convolution for upsampling. The corresponding encoding and decoding layers are connected by MPI closed adaptive skip connections. Each skip connection is equipped with a local closed projection operator based on Moore-Penrose inverse, and gradient backpropagation is actively truncated at the skip connection.

[0012] Preferably, the closed adaptive skip connection is implemented through the following steps: The shallow feature tensor output by the encoder With the deep feature tensor output by the decoder Flattening along the batch and sequence dimensions yields a two-dimensional matrix. and : (5) (6) in, , Where N is the batch size, The sequence length of the l-th layer is... and These represent the number of channels for shallow features and deep features, respectively. by For input, For the projection target, the linear mapping from shallow features to the deep semantic space is transformed into a ridge regression convex optimization problem. The Moore-Penrose inverse is used to obtain the optimal linear least squares projection weight matrix under the ridge regularization constraint for the current batch. : (7) in, This is the regularization penalty coefficient. It is the identity matrix; Accumulate data from each batch using an exponential moving average. To track global statistics across the entire dataset. : (8) in The EMA attenuation coefficient; During the reasoning phase, fixed methods are used. shallow features Perform channel linear mapping to obtain the projected features after noise filtering and channel alignment: (9) Projected shallow features with deep features The features are concatenated along the channel dimension and then fed into the next layer decoder. (10).

[0013] Preferably, the fusion in S300 is a splicing fusion, resulting in a joint feature matrix. : (11) in, For local structural feature matrix, It is a multi-scale time series feature matrix.

[0014] Preferably, the multi-layer classifier in S400 is a multi-layer residual closed-form classifier, which consists of N cascaded sub-networks. Each sub-network contains a hidden layer mapping matrix that is randomly orthogonally initialized. Each layer of the multi-layer residual closed-form classifier performs the following steps: For the first layer, the joint feature matrix As input, for the nth layer (n≥2), the target residual of the previous layer is used as input and mapped to the random orthogonal latent space through a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. ,in, For hidden layer dimensions; For the target residual of the previous level Initial residual , Given the true label matrix, the objective function for the nth layer is defined as: (12) Regarding formula (12) By taking the partial derivative and setting it to zero, we can obtain the unique globally optimal closed-form solution to this convex optimization problem: (13) The residuals are then updated layer by layer according to the following formula as the learning objective for the next layer: (14) in, and These are nonlinear activation functions and their inverse transforms, respectively. After repeating the above process for a total of N layers, the predicted outputs of each layer are accumulated into the final classification result, with no gradient calculation throughout the process.

[0015] A heterogeneous dual-stream time-series signal classification system based on MPI closed-loop optimization is provided for executing a heterogeneous dual-stream time-series signal classification method based on MPI closed-loop optimization, comprising: The signal acquisition and preprocessing module is used to acquire one-dimensional time-series signals and perform preprocessing to obtain standardized input samples. The heterogeneous dual-stream parallel feature extraction module is used to input the preprocessed signal in parallel to the first feature extraction stream and the second feature extraction stream, and extract the local structural feature matrix and the multi-scale temporal feature matrix, respectively; wherein, the two feature extraction processes maintain gradient independence, and the feature extraction paradigms of the two feature extraction streams are heterogeneous to each other; The gradient isolation feature-level fusion module is used to fuse the local structural feature matrix and the multi-scale temporal feature matrix after standardization, under the condition that the gradient computation graphs of both the local structural feature matrix and the multi-scale temporal feature matrix have been truncated, to obtain a joint feature matrix. The multi-layer residual closed-form classification module is used to input the joint feature matrix into a multi-layer classifier. Each layer of the multi-layer classifier contains a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. The classification weights are solved layer by layer using the MPI closed-form solution, and the classification result is obtained by accumulating the prediction outputs of each layer. The classification process of the multi-layer classifier is a gradient-free process.

[0016] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a heterogeneous dual-stream timing signal classification method based on MPI closed-loop optimization.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a heterogeneous dual-stream timing signal classification method based on MPI closed-loop optimization.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. Global Optimality and Reproducibility in the Classification Stage under Fixed Feature Representation: This invention employs MPI (Moore-Penrose Inverse) closed-form solutions in both the DCNet classifier head replacement and the multi-layer residual closed-form classifier. By truncating the gradient and fixing the feature representation, the non-convex optimization of the classification layer is transformed into a ridge regularized convex optimization problem, thus obtaining the unique global optimal solution to this closed-form subproblem. This makes the model independent of random initialization and learning rate, and the classification results are reproducible, overcoming the tendency to get trapped in local optima when using gradient descent for the classifier head. 2. Solving noise pollution from U-Net skip connections at the architecture level: This invention proposes MPI closed adaptive skip connections, introducing a closed projection operator at the skip connection to analytically solve the optimal channel mapping in one step, and actively truncating gradient backpropagation. This mechanism simultaneously eliminates the direct transmission of shallow high-frequency noise to deeper layers, as well as the inverse interference of deep gradients on shallow encoders, fundamentally solving the feature degradation problem of U-Net in high-noise time-series signals.

[0019] 3. Breaking the bottleneck of negative transfer in heterogeneous dual-stream systems, enabling the first practical application of feature-level fusion: This invention performs feature-level concatenation under the condition that the gradients of both heterogeneous features have been truncated. Combined with a multi-layer residual closed-loop classifier with gradient-free computation, it completely eliminates negative transfer caused by mutual interference between heterogeneous gradients. Experiments show that the feature-level fusion performance of this scheme is significantly better than traditional decision-level fusion, fully realizing the complementary potential of heterogeneous dual-stream systems for the first time. Attached Figure Description

[0020] Figure 1 This is a general framework diagram of a heterogeneous dual-stream time-series signal classification method based on closed-loop optimization in one embodiment of the present invention; Figure 2 This is a flowchart of a heterogeneous dual-stream time-series signal classification method based on closed-loop optimization in one embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] See Figure 1 The system architecture of this invention consists of three functional modules connected in series: ① Local structured flow feature extraction module (DCNet two-stage decoupling optimization); ② Multi-scale temporal flow feature extraction module (1D U-Net with MPI closed adaptive skip connections); ③ Heterogeneous feature fusion and classification module (gradient-isolated feature-level concatenation + multi-layer residual closed classifier). The two feature extraction processes are completely independent, and concatenation and fusion are performed under gradient isolation conditions. The entire system classification stage has no end-to-end gradient backpropagation.

[0023] The two streams exhibit fundamental heterogeneity in their feature extraction paradigms: the first feature stream, Stream1 (DCNet), upscales the one-dimensional input to a two-dimensional latent space and extracts local topological features through two-dimensional convolution; the second feature stream, Stream2 (MPI-UNet), maintains the one-dimensional convolution paradigm and extracts sequence context features through a multi-scale encoding and decoding structure. This heterogeneity ensures strong complementarity between the features of the two streams.

[0024] See Figure 2 The overall processing flow of this invention is as follows: (I) Signal Acquisition and Preprocessing S100: Acquire a one-dimensional time-series signal and perform preprocessing to obtain standardized input samples.

[0025] Specifically, one-dimensional time-series signals are acquired, including but not limited to EEG (electroencephalogram) signals, ECG (electrocardiogram) signals, IMU (inertial measurement unit) signals, speech signals, industrial sensor signals, or financial time-series signals. The acquired raw signals undergo necessary filtering, framing, and normalization preprocessing to obtain a standardized input sample set.

[0026] (II) Heterogeneous Dual-Stream Parallel Feature Extraction S200: The preprocessed signal is input in parallel to the first feature extraction stream and the second feature extraction stream to extract the local structural feature matrix and the multi-scale temporal feature matrix, respectively; wherein, the two feature extraction processes maintain gradient independence, and the feature extraction paradigms of the two feature extraction streams are heterogeneous.

[0027] Specifically, traditional convolutional neural networks (such as DCNet) rely on gradient descent algorithms like Adam for end-to-end parameter updates. Because the error surface corresponding to time-series signal classification is highly non-convex, the fully connected classification head at the end is prone to converge to local optima, leading to unstable classification performance. Furthermore, the results of the same network under different random initializations vary significantly, resulting in poor reproducibility. To overcome this optimization bottleneck, this invention proposes a two-stage decoupled optimization strategy: the first stage retains the deep feature learning capability of gradient descent, and the second stage replaces the classification head with an MPI closed-form solution, transforming the non-convex optimization problem into a convex problem with a unique analytical solution, fundamentally eliminating the risk of local optima. Furthermore, DCNet can be replaced with other convolutional neural networks (such as ResNet-1D and InceptionTime), as long as the gradient is truncated in the second stage and the MPI closed-form solution is applied to replace the classification head.

[0028] In one embodiment, the first feature extraction stream in S200 employs a two-stage decoupling optimization strategy to extract the local structural feature matrix. In the first stage, a deep convolutional neural network is trained using the gradient descent algorithm. The encoder of the deep convolutional network consists of multiple Conv2DTranspose convolutional layers (with 64, 128, and 256 filters respectively, a stride of 2, BatchNorm normalization, and SELU activation function), used to transform the one-dimensional input from... Unfolding progressively to a two-dimensional latent space, the decoder consists of a single Conv2D convolutional layer followed by a global average pooling layer. This compresses the two-dimensional representation back into a fixed-dimensional feature vector. The Adam optimizer then performs end-to-end updates to the network weights via standard backpropagation, outputting a feature matrix. (N is the number of samples); In the second stage, the weights of the trained network convolutional layers are frozen and the gradient computation graph is truncated. The optimization of the classification head is transformed into a ridge regression convex optimization problem. The globally optimal closed-form solution is obtained by solving the Moore-Penrose inverse matrix, which serves as the globally optimal classification weight matrix. With the feature representation fixed, this step transforms the non-convex classification optimization problem, which originally relied on stochastic gradient descent, into a convex subproblem with a unique analytical solution under ridge regularity constraints, fundamentally eliminating the risk of local optima in the classification head.

[0029] In one embodiment, in the second stage, the objective function of the ridge regression convex optimization problem is defined as: (1) in, For the characteristic matrix, Let be the classification weight matrix to be determined. For the real label matrix, This is the regularization penalty coefficient. For Tikhonov regularization terms; for formula (1) with respect to Find the partial derivative and set it to zero: (2) Simplify formula (2) and extract... : (3) Multiplying both sides by the inverse matrix on the left yields the globally optimal closed-form solution: (4) in It is the identity matrix. Through grid search The solution obtained by formula (4) is the globally unique optimal solution in a mathematical sense. The result is completely reproducible and does not depend on random initialization and learning rate settings. It should be noted that the search for the regularization coefficient C can be replaced by Bayesian optimization (such as the Optuna framework) instead of grid search, which reduces the search cost when the hyperparameter space is large.

[0030] In one embodiment, the second feature extraction stream employs a one-dimensional U-shaped network with closed adaptive skip connections. The U-shaped network includes multiple encoders and corresponding decoders. Each encoder is followed by max pooling for downsampling, and each decoder is followed by transposed convolution for upsampling. The corresponding encoding and decoding layers are connected by MPI closed adaptive skip connections. Each skip connection is equipped with a local closed projection operator based on the Moore-Penrose inverse, and gradient backpropagation is actively truncated at the skip connection.

[0031] Specifically, the U-Net architecture, with its symmetrical encoder-decoder structure and skip connections, has a natural advantage in multi-scale feature fusion, making it suitable for extracting multi-level contextual features from local details to global semantics in one-dimensional time-series signals. However, the skip connections in traditional U-Net directly concatenate shallow encoder features to the corresponding decoder layer, resulting in two fundamental drawbacks in noisy time-series signals: Noise pass-through problem: Shallow features contain a lot of high-frequency transient noise. Directly splicing these features transmits the noise to deeper layers without filtering, polluting the semantic feature space.

[0032] Gradient back-propagation contamination problem: Deep gradients propagate back to shallow layers through skip connections, disrupting the physical morphology learning of the shallow layers of the encoder and interfering with the multi-scale feature learning process.

[0033] To address this, this invention proposes an MPI closed adaptive skip connection mechanism. At each skip connection, a local closed projection operator based on the Moore-Penrose inverse is introduced, and gradient backpropagation is actively stopped (Stop-Gradient) at this point. This mechanism transforms the channel mapping from shallow features to the deep semantic space into a linear least squares problem under ridge regularization constraints. The optimal projection weight matrix for the current batch is obtained in one step through analytical computation, achieving adaptive closed filtering of shallow, high-noise features and realizing "noise isolation" at the physical computation graph level.

[0034] MPI-UNet is constructed using standard 1D convolutions (Conv1D), forming a fundamental heterogeneity with the 2D convolutions of Stream 1: Stream 1 is responsible for 2D local topology reconstruction, while Stream 2 is responsible for 1D multi-scale temporal context capture. The network consists of 4 encoder layers and corresponding 4 decoder layers, with the number of channels as follows: (Encoding direction) Each encoder layer is followed by max pooling for downsampling, and each decoder layer is preceded by transposed convolution for upsampling. Encoder and decoder layers are connected via MPI closed adaptive skip connections. Finally, a multi-scale temporal feature matrix is ​​output through global average pooling (GAP). .

[0035] Unlike the independent two-stage training of Stream 1, MPI-UNet uses a gradient method to train the encoder-decoder backbone weights as a whole, and introduces closed projection and truncates gradients only at skip connections to ensure that the integrity of backbone feature learning and the purity of skip connections are independent of each other.

[0036] In one embodiment, a closed adaptive skip connection is implemented through the following steps: Feature flattening: flattening the shallow feature tensors output by the encoder. With the deep feature tensor output by the decoder Flattening along the batch and sequence dimensions yields a two-dimensional matrix. and : (5) (6) in, , Where N is the batch size, The sequence length of the l-th layer is... and These represent the number of channels for shallow features and deep features, respectively; due to the expansion of the sample size... The matrix has an extremely large number of rows, which fully guarantees the full rank condition and the numerical stability of MPI inversion.

[0037] Training-period batch closed-form solution: For input, For the projection target, the linear mapping from shallow features to the deep semantic space is transformed into a ridge regression convex optimization problem. The Moore-Penrose inverse is used to obtain the optimal linear least squares projection weight matrix under the ridge regularization constraint for the current batch. : (7) in, This is the regularization penalty coefficient. It is the identity matrix; Formula (7), given the current batch features, solves for the optimal linear channel mapping under ridge regularization constraints, effectively filtering high-frequency noise components orthogonal to deep semantic directions in shallow features. It is worth emphasizing that this invention embeds the MPI closed-loop solver into the underlying architecture of the U-Net skip connections, fundamentally different from existing MPI implementations that only replace the end-classifier head in transfer learning. This invention addresses the problem of shallow high-frequency noise permeating layer by layer during multi-scale feature transfer, representing a novel extension of the MPI mechanism in deep feature fusion pathways, rather than a simple transfer of technical methods. It should be noted that the MPI closed-loop adaptive skip connections can be selectively applied only to shallow high-resolution skip connections (such as the first two layers) to reduce computational overhead in scenarios with less noise; or applied to all layers to maximize the protection range.

[0038] EMA cumulative curing: This involves the continuous accumulation of curing from each batch using exponential moving average (EMA). To track global statistics across the entire dataset. : (8) in This is the EMA attenuation factor, with a default value of 0.99, initially. It is a zero matrix; The core functions of the EMA mechanism are threefold: First, it approximates the characteristic distribution statistics of the entire dataset with a finite batch sample size, making... First, it gradually converges towards the global optimum; second, it suppresses sampling noise in a single batch through historical weighted smoothing, preventing abrupt changes in the projection matrix due to the distribution shift of individual batches; third, it provides implicit regularization to enhance... Numerical stability. It should be noted that EMA accumulates a weighted average of the local optima of each batch, providing an asymptotic estimate of the global statistics, rather than guaranteeing a mathematically optimal global solution. Its advantages lie in engineering feasibility and numerical stability. Furthermore, EMA accumulation can be replaced by solving the entire training set all at once after training (offline mode), suitable for scenarios with a fixed amount of training data and no need for online updates.

[0039] Gradient Truncation and Forward Inference: During the forward propagation phase of training, the gradient computation graph of shallow features is actively truncated at skip connections (Stop-Gradient), and then a fixed gradient is used. shallow features Perform a channel linear mapping (equivalent to an unbiased 1×1 convolution or channel linear mapping) to obtain noise-filtered and channel-aligned projected features: (9) Gradient truncation completely blocks the backpropagation of deep gradients to shallow gradients at the physical computation graph level, eliminating the gradient backpropagation contamination problem. The inference phase directly uses the fixed... Performing a matrix multiplication incurs the same computational cost as a regular linear layer, with no additional overhead.

[0040] Channel alignment stitching: This involves aligning and stitching the projected shallow features. with deep features The features are concatenated along the channel dimension and then fed into the next layer decoder. (10).

[0041] Although this mechanism is applied to all skip connection layers, its adaptive filtering effect is particularly significant in shallow high-resolution features, effectively blocking the transmission of high-frequency transient noise to the deep semantic space. Furthermore, since the projection matrices of each layer are solved independently, this mechanism adaptively learns the optimal feature alignment strategy at each scale, essentially making it a general closed-loop adaptive feature fusion operator.

[0042] It should be noted that while both MPI closed adaptive skip connections and ordinary learnable linear layers (such as 1×1 convolutions) involve matrix multiplication in their inference forms, they differ fundamentally in their optimization paradigms and gradient behaviors: The optimization paradigms differ: a typical 1×1 convolutional layer relies on backpropagation and gradient descent for iterative solutions, with the optimization process dependent on the learning rate and initialization. This poses a risk of local optima under noisy signals, and the training results are random. The MPI gate, on the other hand, treats channel projection as a linear least squares problem under ridge regularization constraints, solving it analytically in one step per batch. Given the features of the current batch, the result is determined, without relying on iteration or initialization. This difference in paradigms is the essential distinction between the two, rather than merely a difference in the number of parameters.

[0043] Gradient isolation prevents contamination: Ordinary 1×1 convolutional layers do not truncate gradients, and deep gradients will be propagated back to shallow layers along skip connections, interfering with the encoder's feature learning; at the MPI closed gate, Stop-Gradient is actively applied, and the transmission of gradient information between the backbone encoder and skip connections is completely decoupled, realizing physical isolation between feature transmission and parameter optimization at the computation graph level.

[0044] (III) Gradient-isolated feature-level fusion S300: Under the condition that the gradient calculation graphs of the local structural feature matrix and the multi-scale temporal feature matrix have been truncated, the local structural feature matrix and the multi-scale temporal feature matrix are fused after standardization to obtain the joint feature matrix. In one embodiment, the fusion in S300 is a splicing fusion, resulting in a joint feature matrix. : (11) in, For local structural feature matrix, It is a multi-scale time series feature matrix.

[0045] Specifically, during the stitching process, the gradient calculation maps of both feature paths were truncated. By freezing and truncating in the second phase of DCNet By independently training and truncating MPI-UNet, there are no gradient interactions, completely eliminating the root cause of negative feature transfer at the computational graph level. This is one of the key mechanisms by which this invention achieves feature-level fusion performance comparable to decision-level fusion in a heterogeneous two-stream framework for the first time.

[0046] As an alternative, the fusion method of the two-stream features can also be changed from splicing to weighted summation. As long as gradient isolation is guaranteed (both feature computation graphs are truncated before fusion), negative migration can be effectively suppressed.

[0047] (iv) Multilevel residual closed-form classification S400: The joint feature matrix is ​​input into a multi-layer classifier. Each layer of the multi-layer classifier contains a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. The classification weights are solved layer by layer using the closed-form solution of MPI. The classification result is obtained by accumulating the prediction outputs of each layer. The classification process of the multi-layer classifier is a gradient-free calculation process.

[0048] Specifically, obtaining joint features Subsequently, if gradient descent is directly used to train the classifier, two fundamental problems exist: First, the gradient will flow back to the two feature representation spaces via the joint features, triggering negative transfer caused by the inconsistent distribution of heterogeneous features, thus negating the effect of the aforementioned gradient isolation design; second, the error surface corresponding to the high-dimensional joint features is still highly non-convex, and gradient descent still faces the risk of local optima. To address these issues, this invention designs a multi-layer residual closed-form classifier: the classification weights are solved layer by layer using the MPI closed-form solution, with zero gradient throughout, which completely blocks gradient backflow and transforms the optimization of each layer into a convex problem with a unique analytical solution.

[0049] In one embodiment, the multi-layer classifier in S400 is a multi-layer residual closed-form classifier, which consists of N cascaded sub-networks. Each sub-network contains a hidden layer mapping matrix that is randomly orthogonally initialized. Each layer of the multi-layer residual closed-form classifier performs the following steps: Hidden layer mapping: For the first layer, using the joint feature matrix As input, for the nth layer (n≥2), the target residual of the previous layer is used as input and mapped to the random orthogonal latent space through a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. ,in, For hidden layer dimensions; Closed-form solution outputs weights: based on the target residual of the previous layer. Initial residual , Given the true label matrix, the objective function for the nth layer is defined as: (12) Regarding formula (12) By taking the partial derivative and setting it to zero, we can obtain the unique globally optimal closed-form solution to this convex optimization problem: (13) Residual update: The residuals are updated layer by layer according to the following formula as the learning objective for the next layer: (14) in, and These are nonlinear activation functions and their inverse transforms, respectively. After repeating the above process for a total of N layers, the predicted outputs of each layer are accumulated into the final classification result, with no gradient calculation throughout the process.

[0050] Specifically, the first layer is... The input is the true label matrix Y, and the learning objective is the initial residual. The output weights are solved using MPI closed-form calculation. The second layer is Continue to solve for the new goal ...and so on, layer by layer, each layer focuses on fitting the residual part that has not yet been captured by the previous layer, and finally the prediction outputs of N layers are accumulated to approximate the true label.

[0051] For the nth layer, if a nonlinear activation function is directly included in the error function... This would violate the analytical derivative property of the objective function, making it impossible to obtain a closed-form solution. Therefore, this invention proposes a linearized objective reconstruction strategy: applying an inverse transformation of the activation function to the objective residual. (For example, when) When it is Sigmoid, This is a Logit transformation that maps the nonlinear error approximation problem back to the linear feature space before activation. This operation ensures the strictly convex optimization properties of the objective function, making it perfectly compatible with the MPI-based closed-loop solution framework.

[0052] The following are some specific application examples: Example 1: Human Activity Recognition Based on IMU Inertial Signals In smart wearable device scenarios, the triaxial accelerometer and gyroscope signals collected by the IMU sensor built into a smartphone or wristband are used as input (S100). After heterogeneous dual-stream parallel feature extraction (S200), gradient isolation feature fusion (S300), and closed classification (S400), the output results of activities such as walking, running, going upstairs, going downstairs, sitting, and standing are obtained. Stream 1 (DCNet) captures the spatial topological features of motion, while Stream 2 (MPI-UNet) captures multi-scale temporal dynamic features. The complementary fusion of the two stream features effectively improves the recognition accuracy of complex motion patterns.

[0053] Example 2: ECG arrhythmia detection In wearable ECG monitoring scenarios, a single-lead ECG signal is used as input (S100). After dual-stream feature extraction (S200), gradient isolation fusion (S300), and closed-loop classification (S400), the output results for various arrhythmias, including normal sinus rhythm and premature ventricular contractions, are obtained. Stream 1 (DCNet) extracts local topological morphological features of the QRS waveform, and Stream 2 (MPI-UNet) extracts multi-scale temporal contextual features of RR interval changes. The two are fused under gradient isolation conditions, effectively avoiding negative transfer caused by joint training of heterogeneous features. The effectiveness of the method is verified on the MIT-BIH arrhythmia database (strict cross-patient assessment protocol).

[0054] Example 3: EEG Emotion Recognition In the context of brain-computer interfaces and emotion computing, differential entropy (DE) features extracted from multi-channel EEG signals are used as input (S100). After dual-stream parallel feature extraction (S200), gradient isolation fusion (S300), and closed classification (S400), the output results are positive, neutral, and negative emotion recognition. Stream 1 (DCNet) models the spatial topological relationship between channels through two-dimensional convolution, while Stream 2 (MPI-UNet) captures the temporal hierarchical features between frequency bands through multi-scale one-dimensional convolution. The complementarity of the heterogeneous features of the two streams is fully reflected in the emotion recognition task.

[0055] Compared with the prior art, the present invention has the following advantages: I. Closed-form solution of key sub-problems, reproducible results in classification stage: This invention achieves closed-form solution for two key sub-problems: (1) the ridge regular linear classification sub-problem of the DCNet classification head under the premise of fixed feature representation; (2) the layer-by-layer residual fitting sub-problem of each layer of the multi-layer residual closed-form classifier. Both of the above sub-problems are strictly convex optimization problems with unique analytical solutions, making the optimization results in the classification stage completely reproducible and independent of the initialization and learning rate of stochastic gradient descent. It should be noted that the backbone feature extraction network still uses gradient descent training, and the entire system does not make an overall global optimal declaration; the reproducibility advantage of this invention is mainly reflected in the classification optimization stage, which has a significant improvement in stability compared with end-to-end gradient training of the classification head.

[0056] II. For the first time, the MPI mechanism is implemented at the skip connections of deep networks, addressing noise pollution from the underlying architecture: Existing MPI-related works apply the closed-form solution of MPI to the replacement of the end-classifier head in transfer learning, which is a post-processing modification of the output layer. This invention implements the MPI mechanism at the skip connections of U-Net, a low-level feature transfer path, targeting the novel problem of shallow high-frequency noise gradually penetrating layer by layer during multi-scale feature fusion. An analytical projection operator is introduced at each skip connection, and gradients are actively truncated. This simultaneously solves the problems of noise pass-through and gradient inverse pollution at the computational graph level. It is a novel application and extension of the MPI mechanism in the internal connection structure of deep networks, and is fundamentally different from existing MPI end-classifier head replacement work.

[0057] III. Experimental Verification – Comparative Analysis of Multiple Ablation Mechanisms Demonstrates the Irreplaceability of Each Innovative Point: To demonstrate the technical effectiveness of each innovative mechanism in this invention, a systematic ablation comparative experiment was conducted on ECG arrhythmia detection (MIT-BIH, strict cross-patient protocol) and IMU activity recognition (UCI HAR) tasks. The results are shown in Table 1: Table 1 Comparison of classification performance of different methods on the UCI HAR and MIT-BIH datasets. The ablation experiments in Table 1 visually demonstrate the technical advantages of this invention: Closed-form optimization avoids local optima (support point 1): DCNet (Adam+MPI) consistently outperforms DCNet (Adam) trained by pure gradient, proving that the closed-form solution of MPI effectively overcomes the defect of gradient method being prone to getting trapped in local optima.

[0058] Heterogeneous dual-stream complementary advantages (support and protection point 2): The complete form of this invention (dual-stream + MPI) achieves optimal performance (96.66% and 90.84%), significantly surpassing classic baselines such as single-stream architecture and 1D-CNN, verifying that two-dimensional spatial morphology and one-dimensional multi-scale temporal features have strong complementarity.

[0059] Table 2 Performance Comparison of Feature-Level Fusion and Decision-Level Fusion Table 2 verifies the breakthrough of this invention in the dual-stream feature fusion mechanism: breaking the negative transfer bottleneck (supporting protection point three): the feature-level fusion scheme proposed in this invention significantly outperforms traditional decision-level fusion (Late Fusion). This directly proves that the skip connections and gradient truncation before fusion combined with the multi-layer residual closed classifier of this invention successfully cut off the gradient interference of heterogeneous features at the computation graph level, enabling feature-level fusion to realize its true potential in complex time-series tasks for the first time.

[0060] Fourth, the two streams exhibit fundamental heterogeneity and complementarity: Stream 1 uses two-dimensional convolution to extract local topological features, while Stream 2 uses one-dimensional convolution to extract multi-scale temporal context features. The two streams differ fundamentally in feature extraction paradigms, receptive field morphology, and information dimensions, ensuring strong feature complementarity. The stronger the heterogeneity, the greater the benefit of gradient-isolated feature-level fusion.

[0061] V. Clear physical meaning of parameters and insensitive hyperparameters: The core hyperparameters of this invention are only the regularization coefficient C (which controls the trade-off between model complexity and empirical risk, and has a clear physical meaning) and the EMA decay coefficient α (which controls the update rate of the projection matrix). There is no need to finely adjust the learning rate scheduling strategy, and the invention maintains stable performance within a large parameter range.

[0062] In one embodiment, a heterogeneous dual-stream time-series signal classification system based on MPI closed-loop optimization is also provided, for performing a heterogeneous dual-stream time-series signal classification method based on closed-loop optimization, including: The signal acquisition and preprocessing module is used to acquire one-dimensional time-series signals and perform preprocessing to obtain standardized input samples. The heterogeneous dual-stream parallel feature extraction module is used to input the preprocessed signal in parallel to the first feature extraction stream and the second feature extraction stream, and extract the local structural feature matrix and the multi-scale temporal feature matrix, respectively; wherein, the two feature extraction processes maintain gradient independence, and the feature extraction paradigms of the two feature extraction streams are heterogeneous to each other; The gradient isolation feature-level fusion module is used to fuse the local structural feature matrix and the multi-scale temporal feature matrix after standardization, under the condition that the gradient computation graphs of both the local structural feature matrix and the multi-scale temporal feature matrix have been truncated, to obtain a joint feature matrix. The multi-layer residual closed-form classification module is used to input the joint feature matrix into a multi-layer classifier. Each layer of the multi-layer classifier contains a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. The classification weights are solved layer by layer using the MPI closed-form solution, and the classification result is obtained by accumulating the prediction outputs of each layer. The classification process of the multi-layer classifier is a gradient-free process.

[0063] For specific limitations regarding the heterogeneous dual-stream timing signal classification system based on MPI closed-loop optimization, please refer to the limitations of the heterogeneous dual-stream timing signal classification method based on MPI closed-loop optimization mentioned above, which will not be repeated here. Each module in the aforementioned heterogeneous dual-stream timing signal classification system based on MPI closed-loop optimization can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0064] In one embodiment, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a heterogeneous dual-stream timing signal classification method based on MPI closed-loop optimization.

[0065] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements a heterogeneous dual-stream timing signal classification method based on MPI closed-loop optimization.

[0066] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0067] The above provides a detailed description of the heterogeneous dual-stream timing signal classification method and system based on MPI closed-loop optimization provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention, and the descriptions of the embodiments are merely for the purpose of helping to understand the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A heterogeneous dual-stream time-series signal classification method based on MPI closed-loop optimization, characterized in that, Includes the following steps: S100: Acquire a one-dimensional time-series signal and preprocess it to obtain standardized input samples; S200: The preprocessed signal is input in parallel to the first feature extraction stream and the second feature extraction stream to extract the local structural feature matrix and the multi-scale temporal feature matrix, respectively; wherein, the two feature extraction processes maintain gradient independence and the feature extraction paradigms of the two feature extraction streams are heterogeneous. S300: Under the condition that the gradient calculation graphs of the local structural feature matrix and the multi-scale temporal feature matrix have been truncated, the local structural feature matrix and the multi-scale temporal feature matrix are fused after standardization to obtain the joint feature matrix. S400: The joint feature matrix is ​​input into a multi-layer classifier. Each layer of the multi-layer classifier contains a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. The classification weights are solved layer by layer using the closed-form solution of MPI. The classification result is obtained by accumulating the prediction outputs of each layer. The classification process of the multi-layer classifier is a gradient-free calculation process.

2. The method according to claim 1, characterized in that, In S200, the first feature extraction stream employs a two-stage decoupled optimization strategy to extract local structural feature matrices. In the first stage, a deep convolutional neural network is trained using gradient descent. The encoder of this deep convolutional network consists of multiple Conv2DTranspose convolutional layers, used to progressively unfold the one-dimensional input into a two-dimensional latent space. The decoder consists of a single Conv2D convolutional layer plus a global average pooling layer, used to compress the two-dimensional representation back into a fixed-dimensional feature vector. In the second stage, the weights of the trained network's convolutional layers are frozen, and the gradient computation graph is truncated. Under the premise of fixed feature representation, the optimization of the classification head is transformed into a ridge regression convex optimization problem. The globally optimal closed-form solution is obtained by solving the Moore-Penrose inverse matrix, which serves as the globally optimal classification weight matrix. .

3. The method according to claim 1, characterized in that, In the second stage, the objective function of the ridge regression convex optimization problem is defined as: （1） in, For the characteristic matrix, Let be the classification weight matrix to be determined. For the real label matrix, This is the regularization penalty coefficient. For Tikhonov regularization terms; for formula (1) with respect to Find the partial derivative and set it to zero: （2） Simplify formula (2) and extract... : （3） Multiplying both sides by the inverse matrix on the left yields the globally optimal closed-form solution: （4） in It is the identity matrix. Through grid search Optimize within the given range.

4. The method according to claim 1, characterized in that, The second feature extraction stream uses a one-dimensional U-shaped network with closed adaptive skip connections. The U-shaped network contains multiple encoders and corresponding decoders. Each encoder is followed by max pooling for downsampling, and each decoder is followed by transposed convolution for upsampling. The corresponding encoding and decoding layers are connected by MPI closed adaptive skip connections. Each skip connection is equipped with a local closed projection operator based on Moore-Penrose inverse, and gradient backpropagation is actively truncated at this skip connection.

5. The method according to claim 4, characterized in that, Closed adaptive skip connections are implemented through the following steps: The shallow feature tensor output by the encoder With the deep feature tensor output by the decoder Flattening along the batch and sequence dimensions yields a two-dimensional matrix. and : （5） （6） in, , Where N is the batch size, The sequence length of the l-th layer is... and These represent the number of channels for shallow features and deep features, respectively. by For input, For the projection target, the linear mapping from shallow features to the deep semantic space is transformed into a ridge regression convex optimization problem. The Moore-Penrose inverse is used to obtain the optimal linear least squares projection weight matrix under the ridge regularization constraint for the current batch. : （7） in, This is the regularization penalty coefficient. It is the identity matrix; Accumulate data from each batch using an exponential moving average. To track global statistics across the entire dataset. : （8） in The EMA attenuation coefficient; During the reasoning phase, fixed methods are used. shallow features Perform channel linear mapping to obtain the projected features after noise filtering and channel alignment: （9） Projected shallow features with deep features The features are concatenated along the channel dimension and then fed into the next layer decoder. （10）。 6. The method according to claim 5, characterized in that, The fusion in S300 is a splicing fusion, resulting in a joint feature matrix. : （11） in, For local structural feature matrix, It is a multi-scale time series feature matrix.

7. The method according to claim 6, characterized in that, The multi-layer classifier in S400 is a multi-layer residual closed-form classifier, which consists of N cascaded subnetworks. Each subnetwork contains a hidden layer mapping matrix that is randomly orthogonally initialized. Each layer of the multi-layer residual closed-form classifier performs the following steps: For the first layer, the joint feature matrix As input, for the nth layer (n≥2), the target residual of the previous layer is used as input and mapped to the random orthogonal latent space through a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. ,in, For hidden layer dimensions; For the target residual of the previous level Initial residual , Given the true label matrix, the objective function for the nth layer is defined as: （12） Regarding formula (12) By taking the partial derivative and setting it to zero, we can obtain the unique globally optimal closed-form solution to this convex optimization problem: （13） The residuals are then updated layer by layer according to the following formula as the learning objective for the next layer: （14） in, and These are nonlinear activation functions and their inverse transforms, respectively. After repeating the above process for a total of N layers, the predicted outputs of each layer are accumulated into the final classification result, with no gradient calculation throughout the process.

8. A heterogeneous dual-stream time-series signal classification system based on MPI closed-loop optimization, characterized in that, For performing the method as described in any one of claims 1 to 7, comprising: The signal acquisition and preprocessing module is used to acquire one-dimensional time-series signals and perform preprocessing to obtain standardized input samples. The heterogeneous dual-stream parallel feature extraction module is used to input the preprocessed signal in parallel to the first feature extraction stream and the second feature extraction stream, and extract the local structural feature matrix and the multi-scale temporal feature matrix, respectively; wherein, the two feature extraction processes maintain gradient independence, and the feature extraction paradigms of the two feature extraction streams are heterogeneous to each other; The gradient isolation feature-level fusion module is used to fuse the local structural feature matrix and the multi-scale temporal feature matrix after standardization, under the condition that the gradient computation graphs of both the local structural feature matrix and the multi-scale temporal feature matrix have been truncated, to obtain a joint feature matrix. The multi-layer residual closed-form classification module is used to input the joint feature matrix into a multi-layer classifier. Each layer of the multi-layer classifier contains a hidden layer mapping matrix that is randomly orthogonally initialized and never updated during training. The classification weights are solved layer by layer using the MPI closed-form solution, and the classification result is obtained by accumulating the prediction outputs of each layer. The classification process of the multi-layer classifier is a gradient-free process.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images