A feature fusion method, device, equipment and medium

Abstract

The application discloses a feature fusion method, device, equipment and medium, and belongs to the technical field of machine learning. The method comprises the following steps: acquiring a first type and a second type of feature subsets of a target object; taking an initial equal-weight fusion result as input, and inputting the initial equal-weight fusion result into a machine learning model based on a symmetric decision tree to obtain an initial prediction result; using a SHAP value analysis method, calculating SHAP contribution values of each feature in each feature subset to the prediction result, and aggregating to obtain total SHAP contribution intensities of each subset; according to the relative sizes of the total SHAP contribution intensities, nonlinearly adjusting fusion weights of each feature subset in the next round of training, so that the weight of the subset with a higher contribution intensity is larger; repeating the training and feedback process until the model performance is stable, and obtaining an optimal fusion strategy. Through the closed-loop feedback driven by data, the dynamic self-adaptive optimization of the feature fusion weight is realized, and the model performance is improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a feature fusion method, apparatus, device, and medium. Background Technology

[0002] In various machine learning and data-driven applications, such as industrial equipment fault diagnosis, medical image analysis, and financial risk control, it is often necessary to process and analyze data from multiple sensors or multiple sources. To make full use of this information, data from different sources is usually extracted into different feature subsets, and then these feature subsets are fused as input to downstream machine learning models.

[0003] Existing feature fusion methods typically employ simple concatenation or preset fixed weights. For example, time-domain and frequency-domain features extracted from different sensor signals are directly concatenated into a longer feature vector, or fixed importance weights are assigned to different types of features based on domain expert experience. However, these methods have significant drawbacks: First, fixed-weight fusion is static and cannot be dynamically adjusted based on the actual contribution of features in a specific task, potentially leading to insufficient utilization of key information and inadequate suppression of irrelevant or noisy features, thus affecting the final performance of the model. Second, the feature fusion process and the basis for model decisions are often opaque, forming a "black box." Researchers find it difficult to understand the roles and importance of different feature subsets in model decisions, posing challenges to model optimization, debugging, and reliable deployment.

[0004] Therefore, there is an urgent need for a feature fusion method that can automatically and dynamically optimize the fusion weights of different feature subsets and provide reasonable explanations for fusion strategies and model decisions. Summary of the Invention

[0005] This invention provides a feature fusion method, apparatus, device, and medium, which solves the problems of related technologies where feature fusion weights typically rely on empirical settings or use fixed weights, lacking a dynamic adaptive adjustment mechanism, resulting in suboptimal fusion effects and a lack of interpretability in the model decision-making process. The technical solution is as follows: On the one hand, a feature fusion method is provided, the method comprising: Obtain the first and second subsets of features of the target object; The initial equal-weighted fusion result of the first type of feature subset and the second type of feature subset is used as the input feature vector and input into the machine learning model based on the symmetric decision tree to obtain the initial prediction result for the target task. Based on the initial prediction results and the corresponding machine learning model, the SHAP value analysis method is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction results. Based on the SHAP contribution value of each feature, the contributions of all features in the first feature subset and the second feature subset are aggregated to obtain the total SHAP contribution intensity of each feature subset. Based on the relative magnitude of their respective total SHAP contribution strengths, the fusion weights of the first and second feature subsets used to combine and form the input feature vector in the next round of training are adjusted in real time and non-linearly, so that the feature subset with the higher total SHAP contribution strength will have a greater increase in its fusion weight in subsequent training. The updated fusion weights are used to recalculate the input feature vector for the next round of training and are again used as input to the machine learning model. Repeat the above training and feedback process until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit; The final converged fusion weights are used as the optimal feature fusion strategy and applied to the actual prediction stage. A weighted combination of the first and second category feature subsets of the real-time input target object is performed and input into the trained and converged machine learning model to obtain the final prediction result. On the other hand, a feature fusion device based on symmetric decision trees and SHAP value feedback is provided. The device includes: an acquisition module for acquiring the first and second category feature subsets of the target object; The initial fusion module is used to take the initial equal-weight fusion result of the first type of feature subset and the second type of feature subset as the input feature vector; The model training module has a built-in machine learning model based on a symmetric decision tree, which is used to obtain prediction results for the target task by taking the input feature vector as input. The SHAP contribution calculation module is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction result based on the prediction result and the corresponding machine learning model, using the SHAP value analysis method. The contribution aggregation module is used to aggregate all feature contributions in the first and second feature subsets based on the SHAP contribution values ​​of each feature, so as to obtain the total SHAP contribution intensity of each feature subset. The weight update module is used to adjust the fusion weights of the first and second feature subsets in the next round of training in real time and non-linearly according to the relative magnitude of their total SHAP contribution intensity. This makes the feature subset with higher total SHAP contribution intensity have a greater increase in fusion weight in subsequent training. The iteration control module is used to control the model training module, the SHAP contribution calculation module, the contribution aggregation module and the weight update module to repeatedly perform iterative training until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit. The online prediction module is used to take the final converged fusion weights as the optimal feature fusion strategy, weight the first and second feature subsets of the real-time input test object, and input them into the machine learning model after training and convergence to obtain the final prediction result.

[0006] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for executing the computer program stored in the memory to implement the steps of the feature fusion method described above.

[0007] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the feature fusion method described above.

[0008] The technical solution provided by this invention can bring at least the following beneficial effects: By quantifying the contribution of different feature subsets to the prediction results using SHAP values, and adjusting the fusion weights based on these contributions in a feedback manner, a dynamic adaptive feature fusion mechanism is realized. Through iterative training, it can automatically discover and strengthen feature subsets that contribute more to the current task, thereby optimizing the quality of the input feature vector and improving the prediction performance of the machine learning model. This invention not only optimizes the fusion weights but also provides interpretability for the feature fusion process through the calculation and aggregation of SHAP contribution values. By analyzing the final converged fusion weights or the weight changes in each iteration, it is clear which type of feature subset plays a more crucial role in a specific task, providing an objective and quantitative basis for feature engineering optimization and model decision understanding. By combining SHAP value analysis with the iterative training process to form a closed-loop feedback, the adjustment of fusion weights becomes data-driven and goal-oriented, avoiding the blindness and complexity of manual parameter tuning and improving the efficiency and automation level of model development. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the overall process of a feature fusion method according to the present invention; Figure 2 This is a schematic diagram of an equivalent circuit model for fitting the electrochemical impedance spectrum of a battery according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a fusion device according to an embodiment of the present invention; Figure 4 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0011] As mentioned above, the present invention provides a feature fusion method and apparatus that can solve the problems that feature fusion weights in related technologies usually rely on experience to set or use fixed weights, lack a dynamic adaptive adjustment mechanism, resulting in suboptimal fusion effects and a lack of interpretability in the model decision-making process.

[0012] The following describes the specific implementation of the above concept.

[0013] Please refer to Figure 1 The present invention provides a feature fusion method, which includes: Step 100: Obtain the first type of feature subset and the second type of feature subset of the target object; Step 101: The initial equal-weighted fusion result of the first type of feature subset and the second type of feature subset is used as the input feature vector and input into the machine learning model based on the symmetric decision tree to obtain the initial prediction result for the target task. Step 102: Based on the initial prediction results and the corresponding machine learning model, the SHAP value analysis method is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction results. Step 103: Based on the SHAP contribution value of each feature, aggregate all feature contributions in the first feature subset and the second feature subset respectively to obtain the total SHAP contribution intensity of each feature subset. Step 104: Based on the relative magnitude of their respective total SHAP contribution strengths, adjust the fusion weights of the first and second feature subsets used to combine and form the input feature vector in the next round of training in real time and non-linearly, so that the feature subset with the higher total SHAP contribution strength will have a greater increase in its fusion weight in subsequent training. Step 105: Use the updated fusion weights to recalculate the input feature vector for the next round of training, and use it again as the input to the machine learning model; Step 106: Repeat the above training and feedback process until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit; Step 107: The final converged fusion weights are used as the optimal feature fusion strategy and applied to the actual prediction stage. The first and second feature subsets of the real-time input test object are weighted and combined, and then input into the machine learning model after training and convergence to obtain the final prediction result.

[0014] This application discloses a feature fusion method aimed at solving the problems of difficulty in weight allocation, reliance on human experience, and inability to adaptively adjust weights when fusing multi-source heterogeneous data in existing technologies. Especially in complex fault diagnosis or state assessment tasks, the contribution of features from different sources to the final decision often changes dynamically, making it difficult for fixed-weight fusion strategies to achieve optimal results. This invention introduces a closed-loop feedback mechanism, utilizing the interpretability analysis results of the model to dynamically guide the feature fusion process, thereby achieving an intelligent and adaptive feature fusion strategy.

[0015] The following description Figure 1 The execution method for each step is shown.

[0016] Step 100: Obtain the first type of feature subset and the second type of feature subset of the target object; This step first obtains two heterogeneous feature subsets of the target object, such as the first feature subset and the second feature subset.

[0017] Specifically, these two types of feature subsets can come from different sensors, different data modes, or features extracted from the same original signal using different methods. In this embodiment, a feature subset refers to a set of features extracted or organized from the original sensing data or basic data based on some common attribute (such as physical source, data mode, frequency domain range, or physical meaning).

[0018] In one specific embodiment, the first type of feature subset is a set of time-domain statistical features extracted from the first physical signal source of the target object, and the second type of feature subset is a set of frequency-domain spectral features extracted from the second physical signal source of the target object. The first physical signal source and the second physical signal source are different types of sensors or different sampling frequency channels.

[0019] Step 101: The initial equal-weighted fusion result of the first type of feature subset and the second type of feature subset is used as the input feature vector and input into the machine learning model based on the symmetric decision tree to obtain the initial prediction result for the target task. In the initial stage, due to the lack of prior knowledge to determine which type of feature is more important, the system adopts the fairest initial strategy, that is, to fuse the two feature subsets with equal weights and combine them into an initial input feature vector.

[0020] Next, the initial equally weighted fused input feature vector is fed into a pre-built machine learning model based on a symmetric decision tree. This model is trained for a specific target task (such as fault classification, performance prediction, etc.) and outputs an initial prediction result. The symmetric decision tree model was chosen because of its regular structure, high computational efficiency, and good compatibility with interpretability analysis methods such as SHAP, which facilitates the rapid and accurate calculation of feature contributions.

[0021] The symmetric decision tree in this embodiment is a complete binary tree structure, where nodes at the same level use the same features and the same splitting threshold for data partitioning to ensure that the decision path of the machine learning model has symmetry and structural interpretability. It is commonly used in gradient boosting algorithms such as CatBoost to improve computational efficiency, reduce model complexity, and effectively prevent overfitting, making it particularly suitable for deployment on embedded systems or edge devices with limited computing resources.

[0022] Step 102: Based on the initial prediction results and the corresponding machine learning model, the SHAP value analysis method is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction results. SHAP (Shapley Additive Explanations) value analysis is a feature attribution method used to interpret the prediction results of machine learning models. Its principle is as follows: each feature in the input feature vector is used as an attribution unit for the prediction contribution; the deviation of the machine learning model's prediction output for the current sample from the average prediction output of all samples is used as the total contribution to be assigned; by calculating the average marginal contribution of each specific feature to the model's prediction output in all possible permutations of the feature set, this average marginal contribution is determined as the contribution allocation value of that feature to the current prediction result; the positive or negative sign of the contribution allocation value indicates the positive or negative direction of the feature's contribution to the prediction result, and the absolute value of the contribution allocation value indicates the strength of the feature's contribution to the prediction result.

[0023] After obtaining the initial prediction result, the core feedback loop of this invention is initiated. The system utilizes the SHAP value analysis method to analyze the machine learning model that has just completed the prediction. Specifically, for the initial prediction result, the SHAP contribution value of each original feature constituting the input feature vector is calculated. This contribution value quantifies the magnitude and direction of each feature's contribution to pushing the prediction result from the average level to the current specific value. Through this step, this method makes the internal decision-making process of the model transparent, revealing the degree to which the model depends on different features in the current state.

[0024] Step 103: Based on the SHAP contribution value of each feature, aggregate all feature contributions in the first feature subset and the second feature subset respectively to obtain the total SHAP contribution intensity of each feature subset. In this step, the system categorizes and aggregates the calculated SHAP contribution values ​​of a massive number of individual features. The contribution values ​​of all features from the first feature subset are aggregated together to form the total SHAP contribution intensity of the first feature subset; similarly, the contribution values ​​of all features from the second feature subset are also aggregated to obtain the total SHAP contribution intensity of the second feature subset. This aggregation operation allows for a holistic comparison of the overall influence of different feature subsets on the model's current decision, without focusing on the role of individual features.

[0025] In one specific embodiment, the sum of the absolute values ​​of the SHAP contribution values ​​of all features within the first type of feature subset is calculated as the total SHAP contribution intensity of the first type of feature subset; the sum of the absolute values ​​of the SHAP contribution values ​​of all features within the second type of feature subset is calculated as the total SHAP contribution intensity of the second type of feature subset.

[0026] Step 104: Based on the relative magnitude of their respective total SHAP contribution strengths, adjust the fusion weights of the first and second feature subsets used to combine and form the input feature vector in the next round of training in real time and non-linearly, so that the feature subset with the higher total SHAP contribution strength will have a greater increase in its fusion weight in subsequent training. After entering the weight adjustment phase, the fusion weights of the two feature subsets in the next round of training are determined based on the relative magnitudes of the total SHAP contribution intensity of the first type of feature subset and the total SHAP contribution intensity of the second type of feature subset. The total SHAP contribution strength of the feature subset is positively correlated with the degree of dependence of the machine learning model on the feature subset. That is, the greater the total SHAP contribution strength, the greater the amount of effective information contributed by the feature subset in the model decision in the current round. Based on this, a nonlinear mapping function is used to adjust the fusion weights, so that the feature subsets with greater total SHAP contribution intensity can obtain a larger weight increase, thereby enabling the feature subsets with a large amount of effective information to obtain higher fusion weights in subsequent iterative training.

[0027] Step 105: Use the updated fusion weights to recalculate the input feature vector for the next round of training, and use it again as the input to the machine learning model; After the weights are updated, the system uses the updated fusion weights to perform a weighted combination of the first type of feature subset and the second type of feature subset to generate an updated input feature vector. The updated input feature vector is used as the new input to the machine learning model to perform the next round of iterative training and obtain the corresponding prediction result. The iterative process includes, in sequence: model training, prediction result acquisition, SHAP contribution value calculation, total SHAP contribution intensity aggregation, and fusion weight update. In each iteration, the fusion weight is adaptively adjusted based on the feedback information from the previous model prediction, so that the feature fusion method is gradually optimized as the iteration progresses, and finally converges to the optimal feature fusion strategy that is conducive to improving the prediction performance of the target task.

[0028] Step 106: Repeat the above training and feedback process until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit; This iterative process does not continue indefinitely, but continues until a preset stopping condition is met; In one specific embodiment, the preset stopping condition may be that, in N consecutive iterations, the improvement in classification performance of the machine learning model on the validation set is less than a preset threshold γ; or, between two adjacent iterations, the sum of the absolute values ​​of the changes in the fusion weights of the first and second feature subsets is less than a preset weight threshold δ. When the system reaches a stable state, it determines that the optimal fusion strategy has been found, and the iteration process terminates. Furthermore, an upper limit on the number of iterations can also be set to prevent excessively long training times.

[0029] Step 107: The final converged fusion weights are used as the optimal feature fusion strategy and applied to the actual prediction stage. The first and second feature subsets of the real-time input test object are weighted and combined, and then input into the machine learning model after training and convergence to obtain the final prediction result.

[0030] Finally, after the iterative process converges, the final set of fusion weights is considered the optimal feature fusion strategy for the specific task and is fixed. In actual online applications or prediction phases, when new test data is input, the system directly uses this optimal fusion weight set to weight and combine the first and second category feature subsets acquired in real time. The result is then input into the trained and converged machine learning model to quickly and accurately obtain the final prediction result. In this way, the present invention transforms the complex feature fusion problem, which requires expert knowledge, into an automated, data-driven optimization process, significantly improving the model's performance and generalization ability.

[0031] In one optional implementation, the feature fusion method is applied to fault diagnosis scenarios of large rotating machinery; The target object is a wind turbine gearbox; The first type of feature subset is obtained by installing an accelerometer on the gearbox housing as the first physical signal source, collecting vibration signals during the operation of the gearbox, and extracting a set of time-domain statistical features from the vibration signals. The set of time-domain statistical features includes root mean square value, kurtosis, margin, and peak factor. The second type of feature subset is obtained by installing an acoustic emission sensor on the main bearing of the gearbox as a second physical signal source, collecting stress wave signals during the operation of the gearbox, and extracting a frequency domain spectral feature set from the stress wave signals. The frequency domain spectral feature set includes specific frequency band energy, spectral entropy, and power spectral density peak values. The vibration signal is used to characterize the impact failure of the gearbox, including pitting corrosion on the tooth surface. The acoustic emission signal is used to characterize microcracks inside the material of the gearbox, including bearing inner ring cracks. The time-domain statistical features and the frequency-domain spectral features are complementary in characterizing different fault types. Based on the first type of feature subset and the second type of feature subset, the iterative training and weight feedback steps of the above method are executed, and the fusion weights of the two types of feature subsets are adaptively driven by the contribution intensity of the SHAP to converge towards the optimal direction. In the online diagnosis stage, the vibration signal features and acoustic emission signal features acquired in real time are weighted and combined with the converged optimal fusion weights, and input into the machine learning model after training convergence, and the fault diagnosis results of the gearbox are output. This implementation avoids the subjectivity of manually adjusting feature weights for different fault types through data-driven adaptive weight adjustment, and is especially suitable for the identification of early and weak faults.

[0032] In a specific application scenario, the feature fusion method is applied to fault diagnosis of lithium-ion batteries; First, execute steps 100 and 101 above, where the target object is a lithium-ion battery; The first type of feature subset is constructed as follows: An AC excitation signal is applied to the lithium-ion battery, and electrochemical impedance spectroscopy data of the lithium-ion battery are collected in the full frequency range. The electrochemical impedance spectroscopy data is divided into three frequency bands according to the frequency range: the first frequency band is from 1 kHz to 10 kHz, the second frequency band is from 1 Hz to 1 kHz, and the third frequency band is from 0.01 Hz to 1 Hz. The real part of impedance, the imaginary part of impedance, the magnitude of impedance, and the phase angle of impedance are extracted from the impedance data of the first frequency band, the second frequency band, and the third frequency band, respectively. The mean, variance, kurtosis, and integral value of each extracted parameter are calculated to form a high-frequency feature subset, a mid-frequency feature subset, and a low-frequency feature subset, respectively. The union of the high-frequency feature subset, the mid-frequency feature subset, and the low-frequency feature subset is taken as the first type of feature subset; Specifically, the frequency range of electrochemical impedance spectroscopy (EIS) is divided into high-frequency, mid-frequency, and low-frequency bands: in, Indicates high frequency band, Indicates mid-frequency band, Indicates low frequency band.

[0033] In one implementation, the frequency band can be divided into:

[0034] For each frequency band, statistical features of the real part of impedance, the imaginary part of impedance, impedance amplitude, and phase angle are extracted, including maximum value, minimum value, mean, variance, median, skewness, kurtosis, differential, and integral features.

[0035] The average real part of the impedance is: The mean of the imaginary part of the impedance is: The impedance amplitude integral characteristic is: In the discrete case, numerical integration can be used for calculation.

[0036] The phase angle change characteristics are as follows: in, and These represent the maximum and minimum frequencies within frequency band B, respectively.

[0037] This yields the first subset of features, including high-frequency features, mid-frequency features, and low-frequency features: ; The high-frequency feature subset is used to characterize the ohmic internal resistance state of the lithium-ion battery and the contact state between the electrode and the current collector. The associated fault types include connection abnormalities and solder joint defects. The mid-frequency feature subset is used to characterize the state of the SEI (Solid Electrolyte Interphase) film on the electrode surface of the lithium-ion battery and the state of the charge transfer process. The associated fault types include lithium plating and abnormal thickening of the SEI film. The low-frequency feature subset is used to characterize the state of the lithium-ion diffusion process inside the electrode material of the lithium-ion battery, and the associated fault types include battery polarization and thermal runaway precursors; at the same time, the above-mentioned second type of feature subset is specified as the equivalent circuit model parameter feature subset. like Figure 2 As shown, by fitting the electrochemical impedance spectroscopy data with an equivalent circuit model, a set of electrochemical parameters is identified, which constitutes the feature subset of the equivalent circuit model parameters. The equivalent circuit model is a second-order RC model, which includes the ohmic resistance Rs, the SEI film resistance Rsei, the charge transfer resistance Rct, the SEI film capacitance Qsei, the SEI film capacitance index nsei, the double-layer capacitance Qdl, the double-layer capacitance index ndl, and the Warburg impedance coefficient σW. Wherein, the ohmic resistance Rs is used to characterize the ohmic polarization state of the lithium-ion battery, the charge transfer resistance Rct is used to characterize the electrochemical polarization state of the lithium-ion battery, and the Warburg impedance coefficient σW is used to characterize the concentration polarization state of the lithium-ion battery. Further, please refer to Figure 2 In one specific embodiment, the equivalent circuit model may employ a model that includes an ohmic internal resistance. 201. SEI film impedance 202 and SEI film non-ideal capacitance 203 Parallel circuit, charge transfer impedance 204 and non-ideal double-layer capacitance 205 parallel circuit, and Warburg diffusion impedance A structure consisting of 206 series connections. This model, namely... It can effectively simulate the impedance behavior of typical lithium-ion batteries.

[0038] Among them, ohmic internal resistance 201 represents the combined resistance of the current collector, electrode material, diaphragm, and electrolyte. SEI film impedance. 202 and SEI film non-ideal capacitance The parallel circuit composed of 203 elements simulates the impedance characteristics of the SEI film on the negative electrode surface. Charge transfer impedance. 204 and double-layer nonideal capacitance The parallel circuit consisting of 205 elements characterizes the kinetic impedance during electrochemical reactions at the electrode / electrolyte interface. Warburg diffusion impedance. Parameter 206 describes the diffusion process of lithium ions within the particles of the electrode active material. By identifying these parameters, the performance state of different components within the battery can be directly quantified.

[0039] constant phase element The impedance is expressed as: in, Here, CPE is the coefficient, and n is the diffusion index. ω=2πf ω is the angular frequency.

[0040] Warburg diffusion impedance can be expressed as: in, for Warburg Diffusion coefficient.

[0041] The equivalent circuit model (ECM) total impedance model is obtained as follows: ECM parameters are identified by minimizing the error between the measured impedance and the model impedance. The objective function can be expressed as: 。

[0042] in, The set of parameters to be identified, i.e. The final second subset of features obtained: Electrochemical mechanism eigenvector .

[0043] The first and second feature subsets correspond to data-driven multi-band statistical features and model-driven electrochemical mechanism features, respectively. The two types of features are complementary in characterizing battery state. Based on the first type of feature subset and the second type of feature subset, the above iterative training and weight feedback steps are performed to obtain the optimal feature fusion strategy for lithium-ion battery fault diagnosis.

[0044] Next, the high-frequency features, mid-frequency features, low-frequency features, and electrochemical mechanism feature vectors will be fused to construct a heterogeneous feature set: .

[0045] Next, step 102 above is executed. Based on the initial prediction result and the corresponding machine learning model, the SHAP value analysis method is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction result. In one specific implementation, the multi-band weighted EIS feature vector enhanced by ECM is used as input, and the battery fault category label is used as output to construct a CatBoost multi-classification model.

[0046] The CatBoost model consists of multiple symmetric decision trees, and its model function is expressed as follows: in, For the number of decision trees, For learning rate, For the first A symmetric decision tree.

[0047] Symmetric decision trees use the same splitting features and splitting thresholds for all nodes at the same depth, which makes the model highly efficient for inference and suitable for deployment in vehicle-mounted BMS or edge computing units.

[0048] The above initial feature vectors: Input the CatBoost classification model for initial training, and calculate the SHAP value of each feature based on the trained CatBoost model. S .

[0049] Next, step 104 is executed. Based on the relative magnitude of their respective total SHAP contribution strengths, the fusion weights of the first and second feature subsets used to combine and form the input feature vector in the next round of training are adjusted in real time and non-linearly. This makes the feature subset with the higher total SHAP contribution strength have a greater increase in its fusion weight in subsequent training. During the weight adjustment phase, after the r-th round of training, the total SHAP contribution intensity S of the high-frequency feature subset is calculated respectively. H (r), the total SHAP contribution intensity S of the mid-frequency feature subsetM (r), the total SHAP contribution intensity S of the low-frequency feature subset L (r) and the total SHAP contribution intensity S of the equivalent circuit model parameter feature subset. E (r); The total SHAP contribution intensity is obtained by summing the absolute values ​​of the SHAP contribution values ​​of each feature within the corresponding feature subset; Each total SHAP contribution intensity is increased by a preset smoothing factor ε, where ε > 0, to obtain the corrected contribution intensity S. H (r)+ε、S M (r)+ε、S L (r)+ε and S E (r)+ε; The smoothing factor ε is used to prevent abnormal weight calculation when the total SHAP contribution strength is zero, and to retain basic weights for feature subsets with low contribution strength, so that they still have the possibility of participating in the competition in subsequent iterations.

[0050] The intensity of each correction contribution is normalized and used as the fusion weight for the next round (r+1th round) of training. Its expression is: in, Let g represent the fusion weights of the g-th feature subset in the (r+1)-th training round, where g∈{H,M,L,E}. This represents the total SHAP contribution strength of the g-th feature subset after the r-th training round. ε>0 is a preset smoothing factor to prevent the denominator from being zero, and C is a normalization constant; it is usually set to the number of subsets to ensure that the sum of all weights is a fixed value. For example, if there are 4 feature subsets, C=4 can be set so that the initial sum of weights is 4. Using this formula, the feature subset with the larger contribution in the previous round will have a higher weight calculated in the next round. The larger the value, the more non-linear adaptive updating of the weights is achieved.

[0051] Step 105: Use the updated fusion weights to recalculate the input feature vector for the next round of training, and use it again as the input to the machine learning model; Step 106: Repeat the above training and feedback process until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit; In one specific embodiment, during N consecutive iterations, the improvement in classification performance of the machine learning model on the validation set is less than a preset threshold γ; specifically, during N consecutive iterations, the classification performance of the machine learning model on the validation set satisfies the inequality: ;in, Let be the classification accuracy measured on the validation set before the termination of the t-th iteration. Let N be the classification accuracy measured on the validation set after the (t+k)th iteration, N be the preset number of consecutive monitoring rounds, and γ be the preset performance improvement threshold, where N≥3 and γ∈(0,0.01]. When the above inequality holds, it is determined that the classification performance of the machine learning model has reached saturation, and the iteration process terminates.

[0052] In one specific embodiment, the sum of the absolute values ​​of the changes in the fusion weights of the first type of feature subset and the second type of feature subset between two adjacent iterations. Less than the preset weight threshold δ; where, Once the system reaches a stable state, it determines that it has found the optimal fusion strategy, and the iterative process terminates. Alternatively, an upper limit on the number of iterations can be set to prevent excessively long training times.

[0053] Finally, step 107 is executed, where the final converged fusion weights are used as the optimal feature fusion strategy and applied to the actual prediction stage. The first and second feature subsets of the real-time input test object are weighted and combined, and then input into the machine learning model after training and convergence to obtain the final prediction result.

[0054] Finally, the SHAP dynamically weighted ECM-enhanced multi-band EIS feature vector is obtained: ; in, The high-frequency feature weights after convergence; The mid-frequency feature weights after convergence The low-frequency feature weights after convergence; These are the feature weights of the converged ECM parameters.

[0055] In this embodiment, an adaptive feedback mechanism based on SHAP values ​​is introduced during the training process of the fault category model. Specifically, the feature contribution is iteratively evaluated and fed back based on the SHAP values, and the fusion weights of the four sets of features—high-frequency, mid-frequency, low-frequency statistical features, and electrochemical mechanism feature vectors—are dynamically adjusted according to the evaluation results.

[0056] In one specific implementation, in order to improve the interpretability of fault classification results, this invention calculates SHAP values ​​while the CatBoost model outputs fault categories and probabilities, to explain the contribution of each EIS feature and ECM parameter to the classification results.

[0057] For the For class-specific faults, the output of the CatBoost model can be represented as: ; The features are divided into high-frequency feature sets according to their source. Mid-frequency feature set Low-frequency feature set and ECM parameter feature set And calculate the contribution strength of different feature groups to the target fault category: ； ； ； ； in, , , , These represent the effects of high-frequency, mid-frequency, low-frequency, and ECM parameters on the first... The contribution intensity of fault identification.

[0058] Furthermore, these mapping rules can be specified as a series of IF-THEN logical judgments. For example, a diagnostic rule base can be established as follows: If the high-frequency band SHAP contribution value High, and the average real part of the impedance in the high-frequency band is... Ohmic resistance If the SHAP contribution value of the real part of the high-frequency impedance is significantly higher than other characteristics, the system will determine that the fault is related to abnormal connection, increased ohmic impedance or abnormal collector. If the mid-frequency band SHAP contribution value and ECM parameter SHAP contribution value And SEI film impedance Charge transfer impedance or SEI membrane CPE parameters Q seihe n sei If the SHAP contribution values ​​all show a significant positive correlation, then the fault is determined to be related to excessive SEI film thickening or lithium plating on the negative electrode. If the contribution value of low-frequency band SAP The Warburg diffusion coefficient is relatively high. If the SHAP contribution values ​​are all significant, the fault is determined to be related to the obstruction of lithium-ion diffusion in the solid phase or an early precursor to thermal runaway. Using this set of rules, the final diagnostic report can be presented to the user in a clear and physically intuitive way.

[0059] Please see Figure 3This application also provides a feature fusion device 800 based on symmetric decision trees and SHAP value feedback. This device 800 can be a dedicated hardware device or a software system running on a general-purpose computer. In one hardware implementation, the device 800 includes a processor 810, a memory 820, a communication interface 830, and a sensor interface 840.

[0060] The memory 820 is used to store computer program instructions and various data generated during processing, such as raw sensor data, extracted feature subsets, intermediate fusion weights, and trained machine learning models. The processor 810 is the computing core of the device, responsible for executing the program instructions stored in the memory 820 to implement the various functional modules of the method of the present invention.

[0061] Specifically, when the processor 810 executes a program, it implements the following functional modules: The acquisition module 801 is connected to one or more external sensors (such as an EIS measuring instrument for batteries, a vibration sensor for machinery, etc.) via a sensor interface 840. It is used to acquire a first subset of features and a second subset of features of the target object. Specifically, the acquisition module 801 acquires the original measurement data of the target object in real time or in batches, and extracts the first type of feature subset and the second type of feature subset from it.

[0062] The initial fusion module 802 is used to take the initial equal-weight fusion result of the first type of feature subset and the second type of feature subset as the input feature vector; Specifically, this module receives two subsets of features from the acquisition module 801, and in the initial stage of iterative training, it fuses them with equal weights to generate an initial input feature vector. In subsequent iterations, it receives new weights from the weight update module 806 and performs weighted fusion accordingly.

[0063] The model training module 803 has a built-in machine learning model based on a symmetric decision tree, which is used to obtain prediction results for the target task by taking the input feature vector as input. Specifically, this module internally embeds or loads a machine learning model (such as the CatBoost model) built on a symmetric decision tree. It uses the input feature vector generated by the initial fusion module 802 as input to train or predict the model and obtain prediction results for the target task. This module then passes the prediction results and the current model state to the SHAP contribution calculation module 804.

[0064] The SHAP contribution calculation module 804 is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction result based on the prediction result and the corresponding machine learning model, using the SHAP value analysis method. Specifically, this module receives the output of the model training module 803 and uses the SHAP value analysis algorithm to calculate the contribution value of each feature in the input feature vector to the current prediction result.

[0065] The contribution aggregation module 805 is used to aggregate all feature contributions in the first feature subset and the second feature subset based on the SHAP contribution values ​​of each feature, so as to obtain the total SHAP contribution intensity of each feature subset. Specifically, this module classifies and summarizes the contribution values ​​of a large number of individual features output by the SHAP contribution calculation module 804. It aggregates all feature contribution values ​​originating from the first type of feature subset to obtain the total contribution intensity of the first type of subset; similarly, it also calculates the total contribution intensity of the second type of subset and passes these two intensity values ​​to the weight update module 806.

[0066] The weight update module 806 is used to adjust the fusion weights of the first type of feature subset and the second type of feature subset in the next round of training in real time and non-linearly according to the relative magnitude of their total SHAP contribution intensity, so that the feature subset with higher total SHAP contribution intensity will have a greater increase in its fusion weight in subsequent training. Specifically, based on the total contribution intensity of the two subsets provided by the contribution aggregation module 805, the module calculates the new fusion weights to be used in the next iteration in real time and non-linearly, and feeds back the updated weights to the initial fusion module 802 to form a closed loop.

[0067] The iteration control module 807 is used to control the model training module, the SHAP contribution calculation module, the contribution aggregation module and the weight update module to repeatedly perform iterative training until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit. Specifically, this module is responsible for monitoring the entire iterative training process. It checks whether the performance of the model training module 803 has converged, or whether the weight changes in the weight update module 806 have stabilized. Once the preset stopping conditions are met, the iteration control module 807 terminates the iteration process and instructs the system to solidify the final fused weights and the trained model.

[0068] The online prediction module 808 is used to take the final converged fusion weight as the optimal feature fusion strategy, weight the first and second feature subsets of the real-time input test object, and input them into the machine learning model after training and convergence to obtain the final prediction result.

[0069] Specifically, after iterative training is completed, the module enters the working state. It receives real-time test data through the acquisition module 801, uses the final fusion weights fixed by the iterative control module 807 to perform weighted combination of features, then calls the converged model trained in the model training module 803 to perform fast prediction, and outputs the final prediction result through the communication interface 830.

[0070] This application also provides a computer device, please refer to... Figure 4 The computer device can be a server, a personal computer, an embedded controller, etc. Its internal structure includes at least one processor 810 and a memory 820. The memory 820 stores executable computer programs, which, when executed by the processor 810, can fully implement all the steps of the feature fusion method described in any of the foregoing embodiments.

[0071] This application also provides a computer-readable storage medium, which can be non-volatile, such as read-only memory (ROM), flash memory, hard disk (HDD), or solid-state drive (SSD), or volatile, such as random access memory (RAM). The storage medium stores a series of computer program instructions, which, when loaded and executed by a processor, can implement all the steps of the feature fusion method described in any of the foregoing embodiments.

[0072] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform any of the feature fusion methods described in the above embodiments.

[0073] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0074] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0075] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A feature fusion method, characterized in that, include: Obtain the first and second subsets of features of the target object; The initial equal-weighted fusion result of the first type of feature subset and the second type of feature subset is used as the input feature vector and input into the machine learning model based on the symmetric decision tree to obtain the initial prediction result for the target task. Based on the initial prediction results and the corresponding machine learning model, the SHAP value analysis method is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction results. Based on the SHAP contribution value of each feature, the contributions of all features in the first feature subset and the second feature subset are aggregated to obtain the total SHAP contribution intensity of each feature subset. Based on the relative magnitude of their respective total SHAP contribution strengths, the fusion weights of the first and second feature subsets used to combine and form the input feature vector in the next round of training are adjusted in real time and non-linearly, so that the feature subset with the higher total SHAP contribution strength will have a greater increase in its fusion weight in subsequent training. The updated fusion weights are used to recalculate the input feature vector for the next round of training and are again used as input to the machine learning model. Repeat the above training and feedback process until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit; The final converged fusion weights are used as the optimal feature fusion strategy and applied to the actual prediction stage. The first and second feature subsets of the real-time input test object are weighted and combined, and then input into the machine learning model after training and convergence to obtain the final prediction result.

2. The method according to claim 1, characterized in that, The first type of feature subset is a set of time-domain statistical features extracted from the first physical signal source of the target object, and the second type of feature subset is a set of frequency-domain spectral features extracted from the second physical signal source of the target object. The first physical signal source and the second physical signal source are different types of sensors or different sampling frequency channels.

3. The method according to claim 1, characterized in that, The aggregation of all feature contributions within the first and second feature subsets specifically includes: calculating the sum of the absolute values ​​of all feature SHAP contribution values ​​within the first feature subset as the total SHAP contribution intensity of the first feature subset; and calculating the sum of the absolute values ​​of all feature SHAP contribution values ​​within the second feature subset as the total SHAP contribution intensity of the second feature subset.

4. The method according to claim 1, characterized in that, The symmetric decision tree is a complete binary tree structure, in which nodes at the same level use the same features and the same splitting threshold to partition data, so as to ensure that the decision path of the machine learning model has symmetry and structural interpretability.

5. The method according to claim 1, characterized in that, The first type of feature subset includes a high-frequency feature subset, a mid-frequency feature subset, and a low-frequency feature subset, while the second type of feature subset includes an equivalent circuit model (ECM) parameter feature subset.

6. The method according to claim 5, characterized in that, The process of adjusting the fusion weights of the first and second feature subsets in real time and non-linearly in the next round of training, based on the relative magnitude of their respective total SHAP contribution strengths, specifically includes: Calculate the total SHAP contribution intensity S of the high-frequency feature subset, mid-frequency feature subset, low-frequency feature subset, and ECM parameter feature subset after the r-th training round. H (r), S M (r), S L (r) and S E (r); Add a smoothing factor ε to the total SHAP contribution intensity of each feature subset to obtain the corrected contribution intensity S. H (r)+ε、S M (r)+ε、S L (r)+ε and S E (r)+ε; The intensity of each correction contribution is normalized and used as the fusion weight for the next round of training. Its expression is: in, Let g represent the fusion weights of the g-th feature subset in the (r+1)-th training round, where g∈{H,M,L,E}. denoted as the total SHAP contribution intensity of the g-th feature subset after the r-th round of training, ε>0 is a preset smoothing factor used to prevent the denominator from being zero, and C is a normalization constant.

7. The method according to claim 1, characterized in that, The criteria for determining that the performance of the machine learning model tends to be stable are: in N consecutive iterations, the improvement of the classification performance of the machine learning model on the validation set is less than a preset threshold γ; or, between two adjacent iterations, the sum of the absolute values ​​of the changes in the fusion weights of the first feature subset and the second feature subset is less than a preset weight threshold δ.

8. A feature fusion device based on symmetric decision tree and SHAP value feedback, characterized in that, include: The acquisition module is used to acquire the first type of feature subset and the second type of feature subset of the target object; The initial fusion module is used to take the initial equal-weight fusion result of the first type of feature subset and the second type of feature subset as the input feature vector; The model training module has a built-in machine learning model based on a symmetric decision tree, which is used to obtain prediction results for the target task by taking the input feature vector as input. The SHAP contribution calculation module is used to calculate the SHAP contribution value of each feature in the first feature subset and each feature in the second feature subset to the prediction result based on the prediction result and the corresponding machine learning model, using the SHAP value analysis method. The contribution aggregation module is used to aggregate all feature contributions in the first and second feature subsets based on the SHAP contribution values ​​of each feature, so as to obtain the total SHAP contribution intensity of each feature subset. The weight update module is used to adjust the fusion weights of the first and second feature subsets in the next round of training in real time and non-linearly according to the relative magnitude of their total SHAP contribution intensity. This makes the feature subset with higher total SHAP contribution intensity have a greater increase in fusion weight in subsequent training. The iteration control module is used to control the model training module, the SHAP contribution calculation module, the contribution aggregation module and the weight update module to repeatedly perform iterative training until the performance of the machine learning model tends to stabilize or reaches the preset iteration limit. The online prediction module is used to take the final converged fusion weights as the optimal feature fusion strategy, weight the first and second feature subsets of the real-time input test object, and input them into the machine learning model after training and convergence to obtain the final prediction result.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-7.

Images