A power unit efficiency degradation prediction system and method fusing deep learning

Abstract

This invention relates to the field of computer system technology based on specific computational models, and discloses a power unit efficiency degradation prediction system and method integrating deep learning. The system includes: a data access module distributing multi-dimensional time-series operating data; parallel first and second feature extraction modules outputting first and second feature tensors respectively; an orthogonal constraint module generating an orthogonal penalty term based on the absolute value of the dot product of the two tensors and the product of their L2 norms, and accumulating it to the prediction loss function; an asymmetric causal gating module generating an axial bias vector based on the variance of the first feature tensor and unidirectionally accumulating it to the second feature tensor to block its backpropagation gradient update; and a prediction output module outputting an efficiency degradation trend curve. This invention constrains the feature space to be orthogonal to cut off the penetration of high-frequency feature gradients, solving the problem of high-frequency operating condition disturbances submerging slowly varying degradation gradients, and improving the prediction accuracy of long-period variation trends.

Description

Technical Field

[0001] This invention relates to a power unit efficiency degradation prediction system and method that integrates deep learning, belonging to the field of computer system technology based on specific computational models. Background Technology

[0002] Currently, recurrent neural networks and time-series convolutional network architectures are used to process continuous observation sequences acquired by multi-channel sensors, thereby constructing a high-dimensional feature space and completing the state prediction of industrial equipment. This constitutes a processing method in the field of computational model technology. Such time-series prediction models transform the spatial manifold of multi-dimensional input signals through the weight matrix of multi-layer network nodes, thereby extracting nonlinear representations.

[0003] Based on the principle of error backpropagation, the update of network weights depends on the gradient flow partial derivatives of the error loss function with respect to the parameters of each layer. When high-amplitude transient operating condition fluctuations and weak, slow-changing deterioration trends coexist in the input sequence, the high-variance high-frequency feature components dominate the backpropagation channel, causing the update vector of the shared weight matrix to be highly biased towards fitting short-term operating condition disturbances. For example, Chinese invention patent CN121434663B discloses a compressor performance degradation analysis and water washing recommendation method based on LSTM neural network and boundary constraint optimization. It uses a single-flow neural network structure to fit the performance degradation trajectory and non-ideal operating conditions such as deep peak shaving of the unit. In this context, the input sequence is synchronous with high-amplitude transient operating condition fluctuations and weak, slow-changing deterioration trends. High-variance, high-frequency feature components dominate the backpropagation channel. In the monitoring environment of large power equipment that has been in service for a long time, the system is constantly under complex operating conditions of variable load peak shaving and non-stationary disturbances. The existing processing method is to maintain the absolute accuracy of the time series data fitting in the short term, so that the neural network weights are updated with high-frequency response transient load pulses, resulting in cumulative gradient coverage. This causes the weak gradient signal representing the long-term deterioration trend to attenuate and be submerged in the deep transmission, so that the calculation model cannot precipitate the long-term equipment degradation characterization in the potential feature space, leading to the degradation of the long-term prediction capability.

[0004] Therefore, how to construct a parallel logic topology that isolates mixed frequency gradient contamination and establish an asymmetric causal feature unidirectional transmission mechanism while maintaining the independent convergence of the long-term slowly varying decay feature space has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A power unit efficiency degradation prediction system integrating deep learning, comprising:

[0006] The data access module is used to acquire multi-dimensional time-series runtime data and distribute it to the deep feature processing module;

[0007] The deep feature processing module includes a first feature extraction module, a second feature extraction module, an orthogonal constraint module, and an asymmetric causal gating module: the first feature extraction module outputs a first feature tensor; the second feature extraction module is connected in parallel with the first feature extraction module and outputs a second feature tensor; the orthogonal constraint module is connected to both the first and second feature extraction modules and calculates the absolute value of the dot product of the first and second feature tensors, dividing it by the product of the L2 norms of the first and second feature tensors to generate an orthogonal penalty term, which is then accumulated into the prediction loss function; the asymmetric causal gating module is connected to both the first and second feature extraction modules and calculates the variance of the first feature tensor and generates an axial bias vector accordingly, which is then unidirectionally accumulated into the second feature tensor to block the backpropagation gradient update of the second feature extraction module.

[0008] The prediction output module, connected to the asymmetric causal gating module, is used to output the efficiency decay trend curve.

[0009] Preferably, the deep feature processing module further includes a feature manifold boundary self-detection module; the feature manifold boundary self-detection module is connected to the second feature extraction module, and is used to generate a numerical upper bound based on the variance envelope of the multidimensional time-series running data, and determine whether the L2 norm of the first feature tensor exceeds the numerical upper bound; when it exceeds the numerical upper bound, the feature manifold boundary self-detection module distributes a truncation signal to the second feature extraction module, and the second feature extraction module responds to the truncation signal and keeps the module weight parameters at their current values ​​in the current update cycle.

[0010] Preferably, the first feature extraction module includes a multi-layer one-dimensional convolution module, used to slide and scan the multi-dimensional time-series running data along the time axis to extract high-frequency feature components that are internal constituent elements of the first feature tensor.

[0011] Preferably, the second feature extraction module includes a series of dilated convolution modules and gated recurrent unit modules; the dilated convolution module is used to expand the dilation rate scan of the multidimensional time-series running data to extract long-period feature components as the basis for modeling the second feature tensor; the gated recurrent unit module is used to model the long-period feature components to generate the second feature tensor.

[0012] Preferably, the asymmetric causal gating module includes a causal masking module and a cross-attention mechanism fusion module; the causal masking module is used to mask the feature elements after the current time in the first feature tensor; the cross-attention mechanism fusion module is used to associate the first feature tensor after being masked by the causal masking module with the second feature tensor, and the output after association is used as the input feature of the prediction output module.

[0013] Preferably, the orthogonal constraint module adds the orthogonal penalty term as a regularization term to the prediction loss function to construct a joint optimization objective function; the deep feature processing module also includes a parameter optimization module, which is connected to the orthogonal constraint module and is used to calculate the gradient value based on the joint optimization objective function and adjust the module weight parameters in the first feature extraction module and the second feature extraction module.

[0014] Preferably, the first feature extraction module, the second feature extraction module, and the asymmetric causal gating module constitute a long-short-term feature cascade module; the data access module inputs the mean component and variance component separated from the multi-dimensional time-series running data into the long-short-term feature cascade module in parallel, and the prediction output module receives the output of the long-short-term feature cascade module and maps it to obtain the efficiency decay trend curve.

[0015] Preferably, the multidimensional time-series operational data acquired by the data access module is a dimensionless digital sequence after being scaled by mean and standard deviation; the dimensionless digital sequence includes load characteristic value sequences, temperature characteristic value sequences and pressure characteristic value sequences at the same numerical scale.

[0016] Preferably, the efficiency degradation trend curve output by the prediction output module includes a first prediction curve and a second prediction curve aligned on the same time dimension axis; the first prediction curve is the equipment degradation component curve, and the second prediction curve is the efficiency acceleration degradation trend component curve.

[0017] A method for predicting the efficiency degradation of power generating units by incorporating deep learning is implemented through a power generating unit efficiency degradation prediction system that incorporates deep learning, and includes the following steps:

[0018] Step 101: The data access module acquires multi-dimensional time-series runtime data and distributes it to the deep feature processing module;

[0019] Step 102: The first feature extraction module outputs the first feature tensor;

[0020] Step 103: The second feature extraction module operates in parallel with the first feature extraction module to output the second feature tensor;

[0021] Step 104: The orthogonal constraint module calculates the absolute value of the dot product of the first feature tensor and the second feature tensor and divides it by the product of the L2 norm of the first feature tensor and the L2 norm of the second feature tensor to generate an orthogonal penalty term, and adds the orthogonal penalty term to the prediction loss function.

[0022] Step 105: The asymmetric causal gating module calculates the variance of the first feature tensor and generates an axial bias vector accordingly. The axial bias vector is then unidirectionally accumulated onto the second feature tensor to block the backpropagation gradient update of the second feature extraction module.

[0023] Step 106: The prediction output module receives the decoupled input features fused by the asymmetric causal gating module and maps the output efficiency decay trend curve.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. In the prediction of power unit efficiency degradation, by setting up a first feature extraction network and a second feature extraction network in parallel logical topology after the data input interface, the temperature, pressure and unit load command streams synchronously collected by the distributed control system are distributed to the dual-stream channel. The spatial cosine similarity between the first feature tensor and the second feature tensor in the vector space is calculated using orthogonal constraint units, generating an orthogonal penalty term and accumulating it as an independent component into the main loss function. During the error backpropagation process, the weight update direction of the two feature extraction networks is constrained, so that the first feature tensor and the second feature tensor converge to an orthogonal state in the vector space. This mechanism cuts off the penetration path of high-frequency operating condition feature gradients into the hidden layer of the second feature extraction network, changing the situation where the high variance component dominates the gradient backpropagation when the conventional model uses a single channel to process mixed time-series signals, resulting in the high-frequency operating condition features drowning out the long-term slow-changing degradation gradient. This allows the second feature extraction network to continuously extract pure physical degradation features without stopping the machine and in a variable load environment, improving the accuracy of the prediction model in extracting long-term performance change trends.

[0026] 2. By setting an asymmetric causal gating unit, the variance statistics of the first feature tensor within a preset sliding time window are extracted. When the variance statistics exceed the static baseline threshold, a one-dimensional compensation bias vector is generated using the built-in mapping matrix. This one-dimensional compensation bias vector is then unidirectionally accumulated onto the second feature tensor in the prediction output layer. The computation graph node of the unidirectional accumulation operation is configured to block gradient backpropagation, so that the reconstructed tensor containing the impact of drastic load changes only completes the forward decay trend prediction output without updating the feature extraction weight baseline of the second feature extraction network's base layer. This unit establishes a unidirectional parameter compensation boundary that conforms to the physical causal law that operating condition fluctuations accelerate fatigue degradation, resolving the contradiction between the absolute mathematical orthogonality of the feature space and the real physical interaction. It reflects the intensifying effect of operating condition fluctuations on lifespan reduction in forward inference while maintaining the long-term memory stability of the feature extraction weights of the second feature extraction network.

[0027] 3. By embedding an information bottleneck processing layer at the front end of the second feature extraction network, the information entropy of each sensor feature channel within the input time window is calculated. When the comprehensive information entropy is lower than a set threshold, dynamic pooling step size adjustment logic is triggered to increase the step size parameter of the pooling operation. When the unit is in steady-state operation and no new physical loss information is generated, this processing layer forcibly compresses the feature transmission dimension of redundant data within the model, so that the second feature tensor retains only the deterioration information spanning long periods, generating synergistic gain with the orthogonal constraint unit, reconstructing the feature density at the source of information flow transmission, avoiding the dilution of weak attenuation signals by highly redundant steady-state operating data, and improving the stability of the long-period prediction network in state recognition under constant operating conditions. Attached Figure Description

[0028] Figure 1 This is a module connection diagram of the power unit efficiency degradation prediction system integrating deep learning according to the present invention;

[0029] Figure 2 This is a flowchart illustrating the operational state transition of the power unit efficiency degradation prediction system integrating deep learning, as described in this invention.

[0030] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0032] A power unit efficiency degradation prediction system integrating deep learning, comprising:

[0033] The data access module is used to acquire multi-dimensional time-series runtime data and distribute it to the deep feature processing module;

[0034] The deep feature processing module includes a first feature extraction module, a second feature extraction module, an orthogonal constraint module, and an asymmetric causal gating module: the first feature extraction module outputs a first feature tensor; the second feature extraction module is connected in parallel with the first feature extraction module and outputs a second feature tensor; the orthogonal constraint module is connected to both the first and second feature extraction modules and calculates the absolute value of the dot product of the first and second feature tensors, dividing it by the product of the L2 norms of the first and second feature tensors to generate an orthogonal penalty term, which is then accumulated into the prediction loss function; the asymmetric causal gating module is connected to both the first and second feature extraction modules and calculates the variance of the first feature tensor and generates an axial bias vector accordingly, which is then unidirectionally accumulated into the second feature tensor to block the backpropagation gradient update of the second feature extraction module.

[0035] The prediction output module, connected to the asymmetric causal gating module, is used to output the efficiency decay trend curve.

[0036] Preferably, the deep feature processing module further includes a feature manifold boundary self-detection module; the feature manifold boundary self-detection module is connected to the second feature extraction module, and is used to generate a numerical upper bound based on the variance envelope of the multidimensional time-series running data, and determine whether the L2 norm of the first feature tensor exceeds the numerical upper bound; when it exceeds the numerical upper bound, the feature manifold boundary self-detection module distributes a truncation signal to the second feature extraction module, and the second feature extraction module responds to the truncation signal and keeps the module weight parameters at their current values ​​in the current update cycle.

[0037] Preferably, the first feature extraction module includes a multi-layer one-dimensional convolution module, used to slide and scan the multi-dimensional time-series running data along the time axis to extract high-frequency feature components that are internal constituent elements of the first feature tensor.

[0038] Preferably, the second feature extraction module includes a series of dilated convolution modules and gated recurrent unit modules; the dilated convolution module is used to expand the dilation rate scan of the multidimensional time-series running data to extract long-period feature components as the basis for modeling the second feature tensor; the gated recurrent unit module is used to model the long-period feature components to generate the second feature tensor.

[0039] Preferably, the asymmetric causal gating module includes a causal masking module and a cross-attention mechanism fusion module; the causal masking module is used to mask the feature elements after the current time in the first feature tensor; the cross-attention mechanism fusion module is used to associate the first feature tensor after being masked by the causal masking module with the second feature tensor, and the output after association is used as the input feature of the prediction output module.

[0040] Preferably, the orthogonal constraint module adds the orthogonal penalty term as a regularization term to the prediction loss function to construct a joint optimization objective function; the deep feature processing module also includes a parameter optimization module, which is connected to the orthogonal constraint module and is used to calculate the gradient value based on the joint optimization objective function and adjust the module weight parameters in the first feature extraction module and the second feature extraction module.

[0041] Preferably, the first feature extraction module, the second feature extraction module, and the asymmetric causal gating module constitute a long-short-term feature cascade module; the data access module inputs the mean component and variance component separated from the multi-dimensional time-series running data into the long-short-term feature cascade module in parallel, and the prediction output module receives the output of the long-short-term feature cascade module and maps it to obtain the efficiency decay trend curve.

[0042] Preferably, the multidimensional time-series operational data acquired by the data access module is a dimensionless digital sequence after being scaled by mean and standard deviation; the dimensionless digital sequence includes load characteristic value sequences, temperature characteristic value sequences and pressure characteristic value sequences at the same numerical scale.

[0043] Preferably, the efficiency degradation trend curve output by the prediction output module includes a first prediction curve and a second prediction curve aligned on the same time dimension axis; the first prediction curve is the equipment degradation component curve, and the second prediction curve is the efficiency acceleration degradation trend component curve.

[0044] A method for predicting the efficiency degradation of power generating units by incorporating deep learning includes the following steps:

[0045] Step 101: The data access module acquires multi-dimensional time-series runtime data and distributes it to the deep feature processing module;

[0046] Step 102: The first feature extraction module outputs the first feature tensor;

[0047] Step 103: The second feature extraction module operates in parallel with the first feature extraction module to output the second feature tensor;

[0048] Step 104: The orthogonal constraint module calculates the absolute value of the dot product of the first feature tensor and the second feature tensor and divides it by the product of the L2 norm of the first feature tensor and the L2 norm of the second feature tensor to generate an orthogonal penalty term, and adds the orthogonal penalty term to the prediction loss function.

[0049] Step 105: The asymmetric causal gating module calculates the variance of the first feature tensor and generates an axial bias vector accordingly. The axial bias vector is then unidirectionally accumulated onto the second feature tensor to block the backpropagation gradient update of the second feature extraction module.

[0050] Step 106: The prediction output module receives the decoupled input features fused by the asymmetric causal gating module and maps the output efficiency decay trend curve.

[0051] Example 1: In the scenario of monitoring power unit operation data including variable load peak shaving conditions, the multi-dimensional time-series operation data collected by the distributed control system simultaneously contains high-amplitude transient operating condition fluctuation characteristics and weak, slowly changing attenuation trend characteristics. Because the temperature, pressure, and unit load command changes caused by grid dispatching instructions have high-frequency and large-variance characteristics, while the efficiency attenuation caused by equipment material wear or thermal structural fouling has low-frequency and weak signal characteristics, when these mixed-frequency time-series signals are synchronously input into a general deep learning model, the model training relies on minimizing the overall prediction error, resulting in the high-variance high-frequency operating condition characteristics generating a dominant gradient in the backpropagation channel, producing an increasing gradient overlay. This submerges the weak low-frequency gradient signal representing physical attenuation in the deep propagation, causing the update vector of the shared weight matrix to be highly biased towards fitting short-term operating condition disturbances. Consequently, the computational model cannot precipitate long-term equipment attenuation characteristics in the potential feature space, leading to the risk of overfitting by degrading long-term prediction capabilities and misjudging short-term load declines as permanent efficiency attenuation.

[0052] To eliminate the entanglement and contamination of multi-scale features during gradient evolution, the power unit efficiency degradation prediction system integrating deep learning includes a data access module, a deep feature processing module, and a prediction output module. The data access module acquires dimensionless digital sequences after mean-standard deviation scaling as multi-dimensional time-series operating data, and simultaneously distributes load feature value sequences, temperature feature value sequences, and pressure feature value sequences at the same numerical scale to the first feature extraction module and the second feature extraction module in the deep feature processing module. The first feature extraction module uses a multi-layer one-dimensional convolution module to slide and scan the multi-dimensional time-series operating data along the time axis to extract high-frequency feature components, and outputs a first feature tensor representing the transient operating state of the unit. The second feature extraction module, connected in parallel, uses a series of dilated convolution modules and gated recurrent unit modules to expand the dilation rate scan of multidimensional time-series operating data to extract long-period feature components and model them, outputting a second feature tensor characterizing the long-term deterioration state of the unit. According to Shannon's information theory, the continuous observation sequence generated by a steady-state physical system has low information entropy and high redundancy. Before the multi-dimensional time series operation data is input into the second feature extraction module, the front-end embedded information bottleneck processing layer is used to calculate the joint Shannon information entropy of the load feature value, temperature feature value and pressure feature value sequence within the preset sampling point time window. When the joint Shannon information entropy calculation value is lower than 0.5 for three consecutive sliding time windows, it is determined that the unit is in steady-state operation condition. The step size adjustment instruction is distributed to the expanding convolution module to increase the pooling operation step size from the reference value 1 to the preset value 3.

[0053] The aforementioned threshold of 0.5 is established through empirical upper bound calculations derived from the joint information entropy distribution characteristics of tens of thousands of benchmark time-series data points obtained from offline statistical analysis of the target unit within its 100% full-load rated operating range. When the real-time information entropy measurement is below 0.5, probability theory can be used to definitively prove that the proportion of new physical feature increments caused by load disturbances in the current multi-sensor sequences is below the minimum critical value, and the overall data exhibits extremely high periodic static redundancy. Simultaneously, increasing the pooling operation step size to a preset value of 3 is based on the engineering discrete calibration results of the Nyquist sampling theorem in a slowly varying low-frequency signal downsampling scenario. This extraction step size can reduce the calculation length of the time-domain feature map to one-third, significantly saving throughput resources, while ensuring that the low-frequency main envelope component representing the aging trend of long-cycle equipment does not suffer information truncation or aliasing distortion. It also adjusts the action to reduce the dimension of redundant data feature transmission within the model, maintaining the second feature tensor. To obtain the sparsity of degraded information over long periods, the orthogonal constraint module is connected to the first feature extraction module and the second feature extraction module respectively, and receives the first feature tensor. With the second feature tensor The spatial cosine similarity between the two in the latent feature space is calculated to obtain the orthogonal penalty term. Orthogonal penalty term The calculation formula is as follows: ,in, For the dimensionless orthogonal penalty term component, The dimensionless first feature tensor output by the first feature extraction module represents the transient operating state of the unit. The dimensionless second feature tensor output by the second feature extraction module represents the long-term deterioration state of the unit. The ratio of the absolute value of its dot product to the product of the second norm maintains the high-dimensional dimensionless equivalence and self-consistency of the dimensions on both sides of the formula.

[0054] The orthogonal constraint module will include orthogonal penalty terms. As a regularization term, it is accumulated into the prediction loss function to establish the joint optimization objective function. The parameter optimization module calculates the gradient value based on the joint optimization objective function and adjusts the module weight parameters in the first and second feature extraction modules. During error backpropagation, it constrains the weight update direction of the two modules, making the first feature tensor... With the second feature tensor The vector space converges to an orthogonal state, thus providing a mathematical barrier within the neural network to isolate high-frequency operating condition disturbances from low-frequency physical degradation. This cuts off the penetration path of high-frequency operating condition feature gradients into the hidden layer of the second feature extraction module, enabling the second feature extraction module to continuously extract pure physical degradation features without shutdown and under varying load environments. This increases the accuracy of the prediction model in extracting long-term performance variation trends. This isolation barrier does not directly violate or change the backpropagation differential chain solution system fundamental to deep learning, but rather optimizes the dynamic geometric constraints of the objective to make the parameters representing transient high-frequency disturbances... The gradient descent direction and the weight update descent direction of the extraction device with long-period decay gradually approach a 90-degree perpendicular angle within the deep, high-dimensional parametric manifold space. According to the physical projection law reflected by the vector dot product, when the two-stream network space converges completely orthogonally, no matter how violently the high-frequency error penalty amplitude of the input oscillates, its mathematical projection component on the other low-frequency steady-state network weight update direction will approach a zero vector. Thus, through absolute mutual separation at the direction level, the substantial cutoff and immunity of the interference gradient of the underlying weight update is achieved within the shared global joint computation framework.

[0055] Since absolute mathematical orthogonality would sever the physical relationship that objectively accelerates the fatigue degradation process of the unit under high-frequency variable operating conditions, the deep feature processing module is further equipped with an asymmetric causal gating module connected to the first feature extraction module and the second feature extraction module respectively. The asymmetric causal gating module uses a causal masking module to mask the first feature tensor. Calculate the first feature tensor after occlusion from the feature elements after the current time step. The local variance value within a preset sliding time window is used to characterize the severity of load changes. Based on this, a one-dimensional axial offset vector is calculated using a built-in mapping matrix. According to the principle of cumulative material fatigue damage, the fatigue damage degree of pressure-bearing components is positively correlated with the variance of alternating stress amplitude. The built-in mapping matrix contains static weight coefficients, which are predetermined through an offline calibration process: extracting rotor temperature and steam pressure data covering the entire overhaul cycle from the unit's historical operating database; using the least squares method to fit a scatter plot of the correlation between the local variance values ​​of historical time-series data and the fatigue crack propagation rate calibration values ​​obtained from physical flaw detection in the unit's historical inspection records; extracting the slope parameter of the fitted relationship curve and solidifying it as a weight benchmark; and extracting the current first feature tensor during the real-time inference stage. The local variance statistic is linearly multiplied with the mapping matrix to generate a one-dimensional axial bias vector. This reduces the dimensionless control flow variance to a quantized physical acceleration attenuation compensation value. During data flow interaction, the extracted local variance statistic is numerically represented as a one-dimensional time-series numerical sequence with a length consistent with the depth of the input time sliding window. The aforementioned pre-defined built-in mapping matrix is ​​initialized as a horizontal weight vector with one row and the same number of columns as the number of hidden layer feature channels in the second feature extraction module. When the two execution components trigger the operation, the tensor outer product operation logic at the bottom of the framework performs fully connected divergence in the feature direction, expanding the single variance scalar sequence into a two-dimensional bias supplementary matrix with multi-channel feature width attributes. This ensures that the transformed axial bias vector achieves a complete structural size match with the second feature tensor in both time and channel scales. The cross-attention mechanism fusion module then unidirectionally accumulates the one-dimensional axial bias vector into the second feature tensor. Furthermore, the computation graph nodes of the one-way accumulation operation are used to block gradient backpropagation, so that the reconstructed tensor containing the effects of load drastic changes is only used for the forward decay trend prediction output, and does not propagate gradients to the backpropagation channel to maintain the feature extraction weight benchmark at the bottom layer of the second feature extraction module.

[0056] The prediction output module is connected to the asymmetric causal gating module. It receives the fused decoupled input features and maps them to the output efficiency decay trend curve. The efficiency decay trend curve includes a first prediction curve and a second prediction curve aligned on the same time dimension axis. The first prediction curve is the equipment degradation component curve, and the second prediction curve is the efficiency acceleration decay trend component curve. This protects the long-term memory stability of the underlying decay feature extraction network while reflecting the physical exacerbation effect of operating condition fluctuations on lifespan reduction. The prediction curve output by the prediction output module is stripped of transient load interference and directly reflects the degradation trend of the unit's underlying thermal structure. This solves the overfitting deviation problem of conventional deep learning models in long-cycle equipment performance prediction. To cope with abnormal input mutation noise caused by extreme operating conditions or severe external power grid impacts, the deep feature processing module also includes a feature manifold boundary self-detection module connected to the second feature extraction module. The feature manifold boundary self-detection module calculates the numerical upper bound based on the variance envelope of multi-dimensional time-series operating data and determines the first feature tensor. Whether the L2 norm exceeds the upper bound; when it exceeds the upper bound, the feature manifold boundary self-detection module distributes a truncation signal to the second feature extraction module. The second feature extraction module responds to the truncation signal and keeps the module weight parameters at their current values ​​within the current update cycle, thereby pausing the gradient backpropagation update of the second feature extraction module for the current batch of data. This provides a protective isolation barrier against external shocks or fault data, preventing transient distortions from destroying the long-term accumulated decay feature weights, and enabling the computational model to maintain its original long-term memory and stably output the efficiency decay trend curve when facing external abnormal working condition samples.

[0057] Example 2: When the system monitors the operating data of power generating units and verifies the predictive stability of the calculation model under variable load peak shaving conditions, the time-series data required for the experiment comes from a real-time monitoring platform of a distributed control system of a 600MW supercritical thermal power generating unit. The data resolution of the temperature measurement points is set to 0.1℃, the acquisition accuracy of the pressure sensor is maintained at 0.01MPa, and the sampling frequency of load and active power is set to 1Hz. To simulate the high-frequency complex load regulation disturbances caused by actual power grid dispatch, Gaussian white noise with a signal-to-noise ratio of 20dB and 50Hz power frequency high-order harmonic noise components are actively superimposed onto the load characteristic value sequence, temperature characteristic value sequence, and pressure characteristic value sequence at the input end to construct a test matrix with non-ideal dissipation characteristics. The sliding sampling window length in the depth feature processing module is determined. At that time, due to the main technical constraints of the transient load transition rate of the generator set and the transient throughput bandwidth of the feature register of the processing chip, the essence of its technical trade-off lies in balancing the timeliness of capturing the timing characteristics of the variable operating conditions with the smooth stability of suppressing random measurement noise. Accordingly, a window size selection rule is established. That is, when the transient fluctuation gradient amplitude of the signal to be monitored is in the preset high dynamic oscillation range, in order to prevent local feature phase aliasing from occurring when the convolution operator slides along the time axis, the sliding sampling window length is determined. To improve the fine-grained sensitivity of time series feature analysis, the sliding sampling window length is approached towards the lower limit of its defined range. Therefore, in this experiment, the sliding sampling window length is... The time-domain window containing 512 consecutive data points was locked as an engineering example of applying this determination rule under the condition of high frequency and large variance of multidimensional time series running data.

[0058] During the verification process of feature decoupling and model training, multi-dimensional time-series operational data containing composite interference and with a matrix dimension of 512 x 3 was synchronously distributed to a parallel feature extraction network. The first feature extraction module used multi-layer one-dimensional convolution operators to slide and scan the time-series data, generating a dimensionless first feature tensor with a dimension of 512 x 64 representing the transient operating state of the unit in the internal hidden layer register. The second feature extraction module, which maintains an independent topological connection with the main module, utilizes a series of three dilated convolution operators with dilation rates of 1, 2, and 4, along with a gated recurrent unit operator containing 128 hidden state units, to perform global dependency modeling on long-period slowly varying components. The output is a dimensionless second feature tensor with dimensions of 512 x 64, representing the long-term deterioration state of the unit. To explicitly demonstrate the system-level synergistic effect and the rationality of the boundary range of orthogonalization constraints in the feature space, this experiment constructs a stress test system containing multidimensional control groups. Specifically, this includes the present invention sample group using a complete decoupling scheme, a first-part missing control group with the orthogonal constraint module completely removed from the network architecture, a second-part missing control group with the asymmetric causal gating module completely removed, and an out-of-range control group where the regularization constraint weight parameter in the orthogonal constraint module is set to 1.5 and exceeds the upper limit of the target value window. To compensate for the second-order update delay and computational drift loss caused by high-frequency sliding truncation in the floating-point addition unit of the high-dimensional tensor, a first-order time-difference correction operator is injected at the unidirectional accumulation node of the asymmetric causal gating module. First-order time difference correction operator As shown in the following formula: ,in, This is a dimensionless first-order computational delay compensation operator. This is the preset time constant adjustment coefficient. To characterize the dimensionless first feature tensor of the unit's transient operating condition, its time derivative term is aligned with the first-order time-series difference feature in terms of homogeneity of physical property dimensions. Based on the feedforward differential compensation mechanism of automatic control principles, the derivative component of the system input signal is extracted to perform lead correction to offset the phase delay caused by the inertia of the data processing pipeline. Facing high-frequency load pulses from the external power grid, the high-dimensional tensor slicing and splicing and floating-point multiplication and addition operations within the calculation model generate millisecond-level computational delays, causing the extracted feature gradient to lag behind the actual physical transients. In each data batch inference stage, the dimensionless first feature tensor of the current sampling time and the previous sampling time are calculated. The difference between the L2 norm values ​​is divided by the data acquisition time step to obtain the real-time time derivative. The time derivative is multiplied by a preset constant of 0.05 to determine the current differential slope. Substituting this into the calculation node yields the first-order time difference correction operator. The real-time compensation value is synchronously summed with the bias vector in the asymmetric causal gating module, physically smoothing out the microscopic computational delay and the slow-varying response time difference of the unit. This smoothing does not refer to directly offsetting the macroscopic physical fatigue wear of equipment that has lasted for months or years with millisecond-level internal feature calculation variables, but rather to aligning the time axis within the model to address the data feedforward inference phase difference generated when the calculation model processes high-frequency load changes. System computing power delay will cause the timestamp of the feature tensor reflecting the current severe load fluctuation to lag behind the timestamp of the weak, slowly varying degradation signal originally carried in the underlying data. By applying this first-order time-series differential compensation at the prediction end, the additional cumulative damage exacerbation component caused by load fluctuation can be mapped ahead of time in the computational flow graph, ensuring that each transient condition impact can be accurately projected and summed onto the corresponding degradation feature tensor slice at the same time. This achieves synchronous fusion and matching of external transient state updates and long-period base state evolution within the virtual feature manifold, thereby utilizing the first-order time difference correction operator. Real-time correction of gradient update lag bias caused by drastic load fluctuations.

[0059] To test the nonlinear response and generalization performance of the model under complex industrial interference environments, the experiment actively set up three problem intensity gradient control systems with clearly defined levels, consisting of low-variance perturbation, medium-variance perturbation, and high-variance perturbation. As the interference power spectral density increases regularly at the time-domain input, the root mean square error (RMSE) of the first and second missing control groups (not using the present invention) shows a steep, nonlinear exponential upward trend. This indicates that the strong gradient flow corresponding to high-frequency, high-variance load conditions covers and penetrates the degradation gradient of low-frequency equipment during backpropagation. In contrast, the RMSE of the present invention's sample group stabilizes in a flat, low-level range of 0.012, 0.015, and 0.018 under the three problem intensity gradients, respectively, indicating significant suppression of noise perturbation. Furthermore, a nonlinear performance degradation inflection point is observed in the out-of-range control group. When the regularization constraint weight parameter is adjusted to the overload region exceeding the upper limit of the working window, due to the spatial cosine similarity of the potential feature space... The degree penalty term constrains the mapping degrees of freedom of the hidden layer tensor manifold, causing the second prediction curve output by the prediction output module to produce numerical saturation and physical deformation when characterizing the accelerated decline trend of efficiency. Its prediction root mean square error becomes 0.145. This performance inflection point directly provides data support for establishing the value boundary. The test data of this multidimensional gradient comparison experiment confirms that the physical combination of the orthogonal constraint of the feature space and the unidirectional blocking unit of the causal gradient in the asymmetric causal gating module produces a synergistic anti-noise effect that exceeds the independent superposition of each unit. Under the premise of fully ensuring that the high-frequency load change action has a physical stimulating effect on the life loss, the equipment degradation component corresponding to the first prediction curve is directly mapped to the slow performance degradation trend of the unit's underlying thermal structure. Inside the calculation model, the original distributed control system time-series signal full of high-amplitude random disturbances is converted into a high-fidelity efficiency degradation trend curve output with a clear physical parameter chain, so as to achieve the accurate determination of the health benchmark of the power unit in the long-cycle load change industrial environment.

[0060] Example 3: This example combines Figures 1 to 2 This paper describes a power unit efficiency degradation prediction system and method that integrates deep learning, such as... Figure 1As shown, the multi-dimensional time-series running data at the top points to the data access module distributing multi-dimensional time-series running data. This data is then directed to the first feature extraction module (outputting the first feature tensor) and the second feature extraction module (outputting the second feature tensor). The first feature extraction module outputs the first feature tensor, which is then fed to the orthogonal constraint module to generate an orthogonal penalty term, which is accumulated in the prediction loss function. Similarly, the second feature extraction module outputs the second feature tensor, which is also fed to the orthogonal constraint module to generate an orthogonal penalty term, which is accumulated in the prediction loss function. Finally, the orthogonal constraint module generates an orthogonal penalty term, which is accumulated in the prediction loss function and outputs the orthogonal penalty term. The penalty term is applied to the prediction loss function indicated by the dashed box. Simultaneously, the first feature extraction module outputs the first feature tensor and its variance value to the asymmetric causal gating module to generate an axial bias vector that blocks backpropagation gradient updates. The second feature extraction module outputs the second feature tensor and its variance value to the asymmetric causal gating module to generate an axial bias vector that blocks backpropagation gradient updates. The asymmetric causal gating module generates an axial bias vector that blocks backpropagation gradient updates and outputs the fused decoupled input features to the prediction output module to map the efficiency decay trend. The prediction output module maps the efficiency decay trend and ultimately points to the efficiency decay trend curve.

[0061] like Figure 2 As shown, the process starts from a hollow circle node, migrates and deploys data access, and points to the initial calibration state. In the initial calibration state, the target unit's benchmark dataset is collected, and the adaptive correction factor and stable backpropagation gradient initial polarity are calculated. After updating the feature scaling parameters, the process points to the dual-stream feature parallel extraction state. In this state, a sliding scan is performed, outputting the dual-stream feature tensor. The cosine similarity of the latent space and the orthogonality penalty term are calculated and accumulated into the prediction loss function, constraining the feature space to converge to an orthogonal state. When the local variance exceeds the static benchmark threshold, the process points to the causal-gated unidirectional compensation state. In this state, the local variance of the sliding time window is extracted, and a one-dimensional axial bias vector is generated from the image. This vector is then reconstructed using forward unidirectional accumulation to maintain the stability of the underlying feature extraction weight memory. Gradient backpropagation is blocked, and the output is... The trend curve returns to the parallel extraction state of dual-stream features. When the measured value of the current L2 norm exceeds the upper limit, it points to the manifold boundary truncation protection state. The manifold boundary truncation protection state triggers the dynamic truncation protection loop and distributes the truncation signal to the gating update enable terminal, forcibly blocking the backpropagation update of the current batch and maintaining the current value of the hidden layer module weight parameters. After the input data is restored to within the scheduling range, it returns to the parallel extraction state of dual-stream features. When the full running cycle is reached and the cumulative dot product mean is greater than the calibration threshold, it points to the periodic parameter adjustment state. The periodic parameter adjustment state automatically starts the time window automatic adjustment and determines that feature aliasing has occurred in the hidden layer space, lowers the backpropagation learning rate to a low rate, and suppresses the transient gradient flow from scouting the feature hidden layer. After maintaining the low rate state for 3 update cycles, it returns to the parallel extraction state of dual-stream features.

[0062] Example 4: When the system faces a severe peak-shaving condition caused by high-frequency grid dispatch commands or a transient thermal stress shock caused by a sudden change in fuel calorific value, the multi-dimensional time-series operating data distributed by the data access module contains nested high-amplitude non-steady-state sudden noise. At this time, the feature manifold boundary self-detection module in the deep feature processing module is used to determine when to activate the update truncation mechanism to protect long-term deteriorated state features. However, if the upper bound of the value on which this module depends is in an uncalibrated state, it will cause erroneous truncation during normal operating condition switching or missed truncation under extreme external shocks. Specifically, when the upper bound of the value is set below the normal range, the first feature tensor flowing out during the rapid load change process of the unit will be affected. The L2 norm will frequently cross this boundary, causing the feature manifold boundary self-detection module to erroneously distribute truncated signals to the second feature extraction module. This causes the second feature extraction module to frequently pause the hidden layer weight update, resulting in the loss of continuous equipment health degradation trend information in the time domain, leading to component deviations in long-term efficiency decay prediction. Conversely, if the upper bound of the value is set higher than the normal range, the high-amplitude abnormal distortion gradient generated by external power grid faults or extreme load steps will penetrate unimpeded into the back propagation channel of the second feature extraction module, causing the long-accumulated slow-varying physical degradation characterization weights to be covered, resulting in distortion of the historical operating benchmark of the calculation model and divergence of the long-term prediction trend due to update anomalies.

[0063] To eliminate the setting deviation of the numerical upper bound, the system establishes the numerical boundary through a specific parameter determination procedure during the offline calibration phase. The parameter determination procedure is applied to the historical steady-state operation dataset of the unit, which is accessed by the distributed control system, to calculate the first characteristic tensor under the historical steady-state operating conditions. The statistical mean and standard deviation of the dimensionless L2 norm sequence within a sliding time window are used to determine the dimensionless numerical upper bound for boundary protection through a weighted axiom mapping. Dimensionless upper bound The calculation formula is as follows: ,in, This is a dimensionless upper bound locked by the internal register of the characteristic manifold boundary self-detection module. The first feature tensor under the historical steady-state operation sample set The dimensionless L2 norm mean, The first feature tensor under the historical steady-state operation sample set The dimensionless L2 standard deviation is used to maintain the homogeneity equivalence of physical property units on both sides of the formula through dimensionless scalar operations. This historical steady-state operation dataset is essentially composed of raw multi-dimensional time-series operation data sequences synchronously collected by the front-end distributed control system. The random fluctuation variance envelope of these physical quantities objectively defines the extreme boundary of the underlying data variation of the unit under normal operating conditions. Based on this fixed front-end raw variance envelope benchmark, the multi-dimensional time-series operation data within the corresponding time period is positively input into the first feature extraction module, whose parameters have been frozen. The data is then processed by the one-dimensional convolution operator within the module. By performing deterministic linear multiply-add and nonlinear activation mapping operations, the variance envelope distribution of the data in the three-dimensional physical space is projected proportionally and stably into the hidden feature space. This allows for the explicit derivation of the high-dimensional spatial statistical mean and standard deviation of the first feature tensor directly associated with its mapping. During real-world engineering replication, the processor reads the aforementioned offline calibration parameters stored in non-volatile memory. When real-time multidimensional time-series data acquired by field sensors is input into the depth feature processing module, the feature manifold boundary self-detection module receives the real-time first feature tensor output by the first feature extraction module. The feature manifold boundary self-detection module calculates the real-time first feature tensor within the current time domain window. The current measured value of the L2 norm and the dimensionless upper bound Perform point-by-point numerical comparisons. In the emergency load shedding test condition of the generator set, the real-time first characteristic tensor... The current measured value of the L2 norm suddenly increased to 4.85 due to the transient shock, and this value exceeded the pre-calculated and locked upper bound of the dimensionless value. If the specific value is 3.25, the feature manifold boundary self-detection module generates a level cutoff signal and distributes it to the gate update enable terminal of the second feature extraction module. In response to the signal, the second feature extraction module multiplies the weight gradient matrix product term of the current period by 0 to forcibly block the current backpropagation update, so that the feature extraction weights in the hidden layer state unit of the second feature extraction module remain locked in the original state, blocking the penetration of abnormal gradient flow.

[0064] When the transient impact of the generator set ends and the input data from the field sensors returns to within the dispatch range, the real-time first feature tensor output by the first feature extraction module... The current measured value of the L2 norm converges to 1.80, because it is less than the dimensionless upper bound. With a lock value of 3.25, the feature manifold boundary self-detection module automatically cancels the level cutoff signal, enabling the second feature extraction module to resume normal gradient updates for long-period slowly deteriorating components without delay. In this way, the dynamic cutoff protection loop established by the feature manifold boundary self-detection module within the vector space eliminates the impact of external random abnormal shocks on the long-term hidden layer feature weights. This ensures that the first and second prediction curves output by the prediction output module consistently correspond to the underlying unit structure physical damage and accelerated physical efficiency decay trend throughout the entire life cycle operation. This enables the computational model to extract and predict the long-period efficiency decay law when facing industrial noise.

[0065] Example 5: When the system faces a multi-dimensional time-series operation data monitoring environment involving the migration and deployment of power generating units, due to differences in the auxiliary equipment consumption parameters of the target units and inconsistencies in sensor bias, manifold dislocations occur between the newly accessed data baseline and the parameters of the original model. This causes the output parameters of the first and second feature extraction modules to cross predetermined boundaries. To adjust the model to an initial state that adapts to the current environment, initial calibration is performed before the prediction output module is activated. The data access module collects a benchmark dataset containing 2048 sampling points of the target unit under rated operating conditions. The deep feature processing module reads this dataset, calculates and outputs the dimensionless initial transient feature tensor. With dimensionless initial degradation feature tensor The parameter optimization module calculates the mapping relationship to update the feature scaling parameters and baseline adaptive correction factor. The calculation formula is as follows: ,in, This is a dimensionless baseline adaptive correction factor. Let the dimensionless initial transient characteristic tensor be... The dimensionless initial degradation feature tensor is adjusted using a baseline adaptive correction factor. Adjust the scale of the input data to maintain the initial polarity direction of the backpropagation gradient flow.

[0066] When the system is running in a continuous prediction cycle, in order to control the trend drift and weight parameter changes caused by component wear, the deep feature processing module performs automatic parameter adjustment of the time window, and the parameter optimization module, after every 168-hour running cycle, adjusts the dimensionless first feature tensor retained in the current cycle. The cumulative dot product mean is used as the pre-input. If, during continuous variable load operation, the cumulative dot product mean is found to be monotonically increasing and all 24 consecutive sampling points are greater than the calibration threshold of 0.45, it is determined that feature aliasing has occurred in the hidden layer space. At this time, the asymmetric causal gating module lowers the backpropagation learning rate to 0.1 times the original set factor and maintains a low rate state in the subsequent 3 model update cycles to suppress the transient gradient flow from eroding the hidden layer of feature extraction. The system restores the preset parameter adjustment state, so that the tensor manifold inside the calculation model is restored to the orthogonal steady state. Finally, the first prediction curve and the second prediction curve output by the prediction output module remain consistent with the physical degradation components of the unit under long-term operating conditions.

[0067] Example 6: When the system faces a complex deployment situation where the initial weight hyperparameter topology space of the prediction model is undetermined, causing oscillations or convergence stagnation in the backpropagation gradient flow, the system invokes a high-dimensional multi-manifold topology calibration and loss balance coefficient determination procedure to construct a standardized hidden layer processing path for the computational model. In the processor-controlled offline training environment, a training dataset containing 10,000 dimensionless standardized samples is input, distributed by the data access module. The number of one-dimensional convolutional layers inside the first feature extraction module is set to 3, each containing 64 dimensionless feature channel operators, with a convolution kernel size of 5. The number of gated recurrent unit layers inside the second feature extraction module is set to 2, and the hidden layer dimension is set to 128. To balance the technical problem between the fitting error of the unit's transient operating conditions and the orthogonal constraints of potential features, a dimensionless loss balance coefficient is introduced to perform compound regularization superposition on the objective function. The resulting joint optimization objective function formula is as follows: ,in, For the dimensionless joint optimization objective function term, This is the dimensionless basic loss term for the prediction task. This is a dimensionless loss balance coefficient. The dimensionless orthogonal penalty term components are used to maintain the homogeneity of mathematical dimensions on both sides of the formula through scalar accumulation of the dimensionless components. The calibration system discretizes the loss balance coefficients in the range of 0.01 to 0.50 with a gradient step size of 0.05, and performs 50 cycles of forward calculation and backpropagation parameter updates by substituting each gradient value into the discretized components. The first feature tensor is extracted in real time from the hidden layer register. With the second feature tensor To monitor the separation state of the potential vector space.

[0068] When the loss balance coefficient is set to less than 0.05, the gradient weight of the latent feature orthogonal constraint in the joint optimization objective function is too low, which fails to cut off the penetration of high-frequency variable load feature gradients into low-frequency hidden layers within the neural network, resulting in the first feature tensor... With the second feature tensor When the vector orthogonality between the two is less than 70%, the long-term root mean square error of the computational model on the independent test set degrades to 0.125. However, when the loss balance coefficient is set greater than 0.25, the system generates a nonlinear overfit suppression inflection point. Excessive penalty weights suppress the mapping degrees of freedom of the deep feature processing module to the basic decay components, leading to manifold degradation of the hidden layer weight matrix, numerical divergence in the basic loss term of the prediction task, and an increase in the test root mean square error to 0.188. Correspondingly, the calibration procedure locks the loss balance coefficient at a parameter operating point of 0.10. Under this specific numerical constraint, the hidden layer feature weights converge smoothly after 35 computation cycles, and the first feature tensor... With the second feature tensor The orthogonality in the feature space reaches 94.2%. Finally, the first and second prediction curves output by the prediction output module stably track the performance degradation trend of the real power generation unit in the entire time domain, and the calculation model enters a steady state of high generalization and error convergence.

[0069] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A power unit efficiency degradation prediction system integrating deep learning, characterized in that, include: The data access module is used to acquire multi-dimensional time-series runtime data and distribute it to the deep feature processing module; The deep feature processing module includes a first feature extraction module, a second feature extraction module, an orthogonal constraint module, and an asymmetric causal gating module: the first feature extraction module is used to output a first feature tensor; The second feature extraction module is connected in parallel with the first feature extraction module and is used to output the second feature tensor. The orthogonal constraint module is connected to the first feature extraction module and the second feature extraction module respectively. It is used to calculate the absolute value of the dot product of the first feature tensor and the second feature tensor and divide it by the product of the L2 norm of the first feature tensor and the L2 norm of the second feature tensor to generate an orthogonal penalty term, and accumulate the orthogonal penalty term into the prediction loss function. The asymmetric causal gating module is connected to the first feature extraction module and the second feature extraction module respectively. It is used to calculate the variance value of the first feature tensor and generate an axial bias vector accordingly. The axial bias vector is unidirectionally accumulated onto the second feature tensor to block the backpropagation gradient update of the second feature extraction module. The prediction output module, connected to the asymmetric causal gating module, is used to output the efficiency decay trend curve.

2. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The deep feature processing module also includes a feature manifold boundary self-detection module; The feature manifold boundary self-detection module is connected to the second feature extraction module. It is used to generate a numerical upper bound based on the variance envelope of multidimensional time-series running data and to determine whether the L2 norm of the first feature tensor exceeds the numerical upper bound. When the numerical upper bound is exceeded, the feature manifold boundary self-detection module distributes a truncation signal to the second feature extraction module. The second feature extraction module responds to the truncation signal and keeps the module weight parameters at their current values ​​during the current update cycle.

3. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The first feature extraction module includes a multi-layer one-dimensional convolution module, which is used to slide and scan multi-dimensional time-series running data along the time axis to extract high-frequency feature components that are internal constituent elements of the first feature tensor.

4. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The second feature extraction module includes a series of dilated convolution modules and gated recurrent unit modules; the dilated convolution module is used to expand the dilation rate scan of multidimensional time-series running data to extract long-period feature components as the basis for modeling the second feature tensor. The gated loop unit module is used to model long-period feature components and generate a second feature tensor.

5. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The asymmetric causal gating module includes a causal masking module and a cross-attention mechanism fusion module; the causal masking module is used to mask feature elements in the first feature tensor after the current time step. The cross-attention mechanism fusion module is used to associate the first feature tensor after being masked by the causal masking module with the second feature tensor, and the associated output is used as the input feature of the prediction output module.

6. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The orthogonal constraint module adds the orthogonal penalty term as a regularization term to the prediction loss function to construct a joint optimization objective function; the deep feature processing module also includes a parameter optimization module, which is connected to the orthogonal constraint module and is used to calculate the gradient value based on the joint optimization objective function and adjust the module weight parameters in the first feature extraction module and the second feature extraction module.

7. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The first feature extraction module, the second feature extraction module, and the asymmetric causal gating module constitute a long-short-term feature cascade module; the data access module inputs the mean component and variance component separated from the multi-dimensional time-series running data into the long-short-term feature cascade module in parallel; the prediction output module receives the output of the long-short-term feature cascade module and maps it to obtain the efficiency decay trend curve.

8. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The multidimensional time-series operational data acquired by the data access module is a dimensionless digital sequence after being scaled by mean and standard deviation. The dimensionless digital sequence includes load characteristic value sequences, temperature characteristic value sequences, and pressure characteristic value sequences at the same numerical scale.

9. The power unit efficiency degradation prediction system integrating deep learning according to claim 1, characterized in that, The efficiency degradation trend curves output by the prediction output module include a first prediction curve and a second prediction curve aligned on the same time dimension axis; the first prediction curve is the equipment degradation component curve, and the second prediction curve is the efficiency acceleration degradation trend component curve.

10. A method for predicting the efficiency degradation of power generating units by incorporating deep learning, which is implemented through the power generating unit efficiency degradation prediction system by incorporating deep learning as described in claim 1, characterized in that... Includes the following steps: Step 101: The data access module acquires multi-dimensional time-series runtime data and distributes it to the deep feature processing module; Step 102: The first feature extraction module outputs the first feature tensor; Step 103: The second feature extraction module operates in parallel with the first feature extraction module to output the second feature tensor; Step 104: The orthogonal constraint module calculates the absolute value of the dot product of the first feature tensor and the second feature tensor and divides it by the product of the L2 norm of the first feature tensor and the L2 norm of the second feature tensor to generate an orthogonal penalty term, and adds the orthogonal penalty term to the prediction loss function. Step 105: The asymmetric causal gating module calculates the variance of the first feature tensor and generates an axial bias vector accordingly. The axial bias vector is then unidirectionally accumulated onto the second feature tensor to block the backpropagation gradient update of the second feature extraction module. Step 106: The prediction output module receives the decoupled input features fused by the asymmetric causal gating module and maps the output efficiency decay trend curve.

Images