Equipment sensor data processing method and device for generalized fault diagnosis

Abstract

The present invention provides a method and apparatus for processing equipment sensor data for generalized fault diagnosis, which relates to the field of data processing technology. It determines standardized calculation parameters for multiple source domains based on multiple equipment types or operating conditions, and performs standardized processing on the sensor data of the target domain equipment accordingly, eliminating the scale differences caused by the inherent characteristics of the equipment and the sensor configuration. Furthermore, a preset autoencoder is combined with a dynamic feature decoupling matrix to decouple the time-frequency features in the mixed signal, thereby improving the interpretability and separability of the feature space. In addition, the nonlinear transformation term is used to enhance the model's ability to express complex signals, thereby extracting more representative low-dimensional features. Based on this, it can ensure that accurate and reliable fault detection and identification can be achieved under different equipment and different operating conditions.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and device for processing equipment sensor data for generalized fault diagnosis. Background Art

[0002] As the scale and complexity of industrial equipment continue to increase, the difficulty of equipment maintenance and management also increases. In real-world industrial scenarios, equipment fault diagnosis often relies on data collected from sensors, such as vibration signals, current, voltage, temperature, and sound. However, due to different equipment, different operating conditions, and different sensor configurations, the distribution of data often varies significantly. This difference leads to the so-called "domain shift" problem, whereby models trained in one domain (the source domain) are difficult to directly apply to fault diagnosis tasks in another domain (the target domain).

[0003] Current technologies typically use standardized methods based on single-device or single-domain data to preprocess raw data. However, this approach often fails to fully account for differences between multiple source domains and cannot effectively unify data distribution across different devices or operating conditions. Furthermore, existing feature extraction methods also face challenges, particularly when processing complex signals (such as mixed signals containing high-frequency transient components and low-frequency periodic vibrations). Effective feature decoupling and dimensionality reduction are difficult to achieve, resulting in poor separability in the feature space and, in turn, affecting the accuracy of the final fault diagnosis.

[0004] Therefore, there is an urgent need for a new type of data processing method that can unify data distribution through effective standardization means based on considering the differences in multiple source domains, and use advanced feature extraction technology to accurately capture complex signal characteristics, so as to improve the robustness and generalization ability of the fault diagnosis system and ensure accurate and reliable fault detection and identification under different equipment and different working conditions. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a device sensor data processing method and device for generalized fault diagnosis, which can ensure accurate and reliable fault detection and identification under different devices and different working conditions.

[0006] In a first aspect, an embodiment of the present invention provides a device sensor data processing method for generalized fault diagnosis, wherein the method includes: obtaining sensor data of a preset target domain device; the preset target domain device is used to characterize a device to be diagnosed of a preset device type, or a device to be diagnosed under a preset working condition; based on a preset multi-source domain standardization calculation parameter, the sensor data is standardized to obtain standardized data corresponding to the sensor data; wherein the multi-source domain standardization calculation parameter is calculated based on data characteristics corresponding to sensor data of multiple source domain devices; using a preset autoencoder, the standardized data is encoded based on a dynamic feature decoupling matrix of the standardized data to determine the low-dimensional features corresponding to the standardized data; wherein the autoencoder determines the low-dimensional features corresponding to the standardized data based on a preset nonlinear transformation term; classifying and identifying the low-dimensional features to determine the device fault type diagnosis result corresponding to the low-dimensional features; and determining the fault state corresponding to the target domain device based on the device fault type diagnosis result.

[0007] In combination with the first aspect, an embodiment of the present invention also provides a first implementation method of the first aspect, wherein the steps of classifying and identifying low-dimensional features and determining the equipment fault type diagnosis results corresponding to the low-dimensional features include: inputting the low-dimensional features into a pre-trained classifier, and outputting the fault type probability distribution corresponding to the low-dimensional features; based on the fault type probability distribution, determining the equipment fault type diagnosis results corresponding to the low-dimensional features.

[0008] In combination with the first aspect, an embodiment of the present invention also provides a second implementation of the first aspect, wherein the step of normalizing the sensor data based on the preset multi-source domain normalization calculation parameters to obtain the normalized data corresponding to the sensor data includes: normalizing the sensor data based on the numerical mean and numerical standard deviation of the preset multi-source domain normalization calculation parameters to obtain the normalized data corresponding to the sensor data. The calculation method of the preset multi-source domain normalization calculation parameters includes: obtaining sensor data from multiple source domain devices; comprehensively arranging the sensor data from multiple source domain devices according to the preset data characteristic classification rules to determine the data distribution sequence of each data characteristic in each source domain; the data characteristics include one or more of time domain statistical characteristics, frequency domain energy distribution characteristics, and signal amplitude distribution characteristics; based on the data distribution sequence, respectively calculating the numerical mean and numerical standard deviation corresponding to each data characteristic; and using the numerical mean and numerical standard deviation as the preset multi-source domain normalization calculation parameters.

[0009] In combination with the first aspect, an embodiment of the present invention also provides a third implementation of the first aspect, wherein a preset autoencoder is used to encode the standardized data based on the dynamic feature decoupling matrix of the standardized data, and the step of determining the low-dimensional features corresponding to the standardized data includes: obtaining preset nonlinear transformation terms corresponding to the standardized data; and pre-decoupling the dynamic features of the standardized data to obtain a dynamic feature decoupling matrix; based on the adaptive weight matrix and nonlinear transformation terms of the autoencoder, nonlinear activation processing is performed on the dynamic feature decoupling matrix to determine the low-dimensional features corresponding to the standardized data; wherein the adaptive weight matrix is ​​determined based on a preset weight update amount.

[0010] In combination with the first aspect, an embodiment of the present invention also provides a fourth implementation of the first aspect, wherein the calculation method of the preset nonlinear transformation term includes: using a preset nonlinear activation function to perform nonlinear activation processing on the encoding process of the autoencoder to construct a nonlinear transformation term corresponding to the standardized data.

[0011] In combination with the first aspect, an embodiment of the present invention also provides a fifth implementation of the first aspect, wherein the steps of decoupling the dynamic features of the standardized data and determining the dynamic feature decoupling matrix include: obtaining sensor data corresponding to multiple source domain devices collected in advance to obtain multi-source domain feature parameters; the multiple source domain devices include target domain devices corresponding to the standardized data; constructing graph data corresponding to the multi-source domain feature parameters; based on the graph data, calculating the cosine similarity between each feature parameter in the multi-source domain feature parameters, and determining the similarity weight parameters between the feature parameters; based on the similarity weight parameters, determining the non-Euclidean space distance corresponding to the multi-source domain feature parameters; based on the Pearson correlation coefficient between each feature parameter in the multi-source domain feature parameters, determining the inter-feature relationship corresponding to the multi-source domain feature parameters; and constructing the dynamic feature decoupling matrix corresponding to the standardized data based on the non-Euclidean space distance, the inter-feature relationship and the preset decoupling matrix adjustment factor.

[0012] In combination with the first aspect, an embodiment of the present invention also provides a sixth implementation of the first aspect, wherein the method further includes: based on the dynamic feature decoupling matrix, adaptively adjusting the gradient of the loss function of the autoencoder with respect to the weight to update the weight parameters of the autoencoder.

[0013] In combination with the first aspect, an embodiment of the present invention also provides a seventh implementation method of the first aspect, wherein the method for determining the preset weight update amount includes: obtaining the loss function of the autoencoder and the partial derivative of the loss function with respect to the weight corresponding to each iteration of the autoencoder; determining the dynamic gradient scaling factor corresponding to the loss function based on the partial derivative and the square term of the historical partial derivative corresponding to the partial derivative; determining the implicit gradient scaling update amount corresponding to the partial derivative based on the dynamic gradient scaling factor; and determining the weight update amount corresponding to the autoencoder based on the implicit gradient scaling update amount.

[0014] In combination with the first aspect, an embodiment of the present invention also provides an eighth implementation of the first aspect, wherein a method for calculating a loss function includes: obtaining a preset training sample set, and training an autoencoder based on the training sample set; the training sample set includes sensor parameters of multiple source domain devices; calculating the reconstruction error corresponding to the training sample set, and adaptively determining the sample weighting item parameter corresponding to the training sample set based on the reconstruction error; calculating the information flow parameter corresponding to the training sample set based on the interaction strength and mutual influence strength between the features of the training sample set; projecting the training sample set into a high-dimensional space, calculating the distribution difference corresponding to each source domain device of the training sample set, and determining the maximum mean difference parameter corresponding to the training sample set; distinguishing the features of each source domain device in the training sample set through a preset domain classifier, and adjusting the feature representation of the training sample set using the autoencoder to determine the domain classifier identification adversarial loss parameter corresponding to the training sample set; calculating the loss function corresponding to the autoencoder based on the sample weighting item parameter, information flow parameter, maximum mean difference parameter and domain classifier identification adversarial loss parameter.

[0015] In a second aspect, an embodiment of the present invention further provides an equipment sensor data processing device for generalized fault diagnosis, wherein the device includes: a data acquisition module for acquiring sensor data of a preset target domain device; the preset target domain device is used to characterize a device to be diagnosed of a preset device type, or a device to be diagnosed under a preset working condition; a data processing module for standardizing the sensor data based on preset multi-source domain standardization calculation parameters to obtain standardized data corresponding to the sensor data; wherein the multi-source domain standardization calculation parameters are calculated based on data characteristics corresponding to sensor data of multiple source domain devices; an execution module for encoding the standardized data based on a dynamic feature decoupling matrix of the standardized data using a preset autoencoder to determine the low-dimensional features corresponding to the standardized data; wherein the autoencoder determines the low-dimensional features corresponding to the standardized data based on a preset nonlinear transformation term; a classification module for classifying and identifying the low-dimensional features to determine the equipment fault type diagnosis result corresponding to the low-dimensional features; and an output module for determining the fault state corresponding to the target domain device based on the equipment fault type diagnosis result.

[0016] The embodiments of the present invention bring about the following beneficial effects: The present invention provides a method and apparatus for processing device sensor data for generalized fault diagnosis, which determines standardized calculation parameters for multiple source domains based on multiple device types or operating conditions, and accordingly standardizes the sensor data of the target domain device, eliminating the scale differences caused by the inherent characteristics of the device and the sensor configuration. Furthermore, a preset autoencoder is combined with a dynamic feature decoupling matrix to decouple the time-frequency features in the mixed signal, thereby improving the interpretability and separability of the feature space. In addition, the nonlinear transformation term is used to enhance the model's ability to express complex signals, thereby extracting more representative low-dimensional features. Based on this, it can ensure that accurate and reliable fault detection and identification can be achieved under different devices and different operating conditions.

[0017] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart of a method for processing device sensor data for generalized fault diagnosis provided by an embodiment of the present invention;

[0021] Figure 2 A flowchart of another method for processing device sensor data for generalized fault diagnosis provided by an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of a feature distribution effect provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of noise robustness comparison results provided by an embodiment of the present invention;

[0024] Figure 5 A schematic diagram of the training effect corresponding to a loss function provided in an embodiment of the present invention;

[0025] Figure 6A schematic diagram of a cross-domain diagnostic performance comparison result provided by an embodiment of the present invention;

[0026] Figure 7 A schematic structural diagram of a device sensor data processing apparatus for generalized fault diagnosis provided by an embodiment of the present invention;

[0027] Figure 8 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0029] Regarding the aforementioned technical issues, different equipment types can lead to significant variations in the basic characteristics of sensor data. For example, different types of wind turbines may produce different vibration patterns and noise levels due to varying design parameters and operating environments. Furthermore, changes in operating conditions can also affect sensor data. Even on the same equipment, factors such as load changes and speed adjustments can cause shifts in data distribution. Furthermore, sensor configuration is also a significant factor. Sensors installed in different locations or types may capture completely different signal characteristics, affecting data analysis results. These factors combine to make accurate and reliable fault diagnosis across equipment, operating conditions, and sensor configurations extremely challenging. On the one hand, this inconsistency in data distribution complicates subsequent processing steps, such as feature extraction and normalization. On the other hand, ensuring that the extracted features effectively reflect the actual health status of the equipment and have good generalization capabilities remains a pressing issue.

[0030] To solve the above technical problems, the embodiments of the present invention provide a device sensor data processing method and apparatus for generalized fault diagnosis, which can ensure accurate and reliable fault detection and identification under different devices and different working conditions.

[0031] To facilitate understanding, the present invention first describes a device sensor data processing method for generalized fault diagnosis. Figure 1 The flowchart corresponding to the embodiment of the present invention is shown. Figure 1 , the method comprises the following steps:

[0032] Step S102: Acquire sensor data of a preset target domain device.

[0033] In this embodiment of the present invention, sensor data collected from the device to be diagnosed (i.e., the preset target domain device) is first acquired. The preset target domain device represents the device to be diagnosed of a specific device type or under a specific operating condition. The collected sensor data includes, but is not limited to, physical quantities such as device vibration signals, current, voltage, temperature, and sound, comprehensively reflecting the device's operating status.

[0034] Step S104 : Based on preset multi-source domain normalization calculation parameters, the sensor data is normalized to obtain normalized data corresponding to the sensor data.

[0035] To alleviate the "domain offset" problem caused by different device types, operating conditions, and sensor configurations, an embodiment of the present invention adopts a multi-source domain-based standardization method to standardize the sensor data of target domain devices to unify the data distribution characteristics between different domains.

[0036] Specifically, an embodiment of the present invention constructs a multi-source domain dataset based on historical data from multiple device types or operating conditions, and statistically generates the calculation parameters required for standardization (such as mean, standard deviation, maximum and minimum values, etc.). These parameters serve as a preset standardization benchmark and are used to normalize or Z-score the sensor data of the target domain device, thereby eliminating the scale differences caused by the inherent characteristics of the device and the sensor configuration, narrowing the data distribution differences between the source domain and the target domain, and reducing the impact of domain offset. For example, the vibration amplitude range of the target domain device may be different from that of the source domain, but through the unified standardization parameters of the embodiment of the present invention, it can be mapped to the same dimensional space to reduce the impact of domain offset.

[0037] Step S106 , using a preset autoencoder, encoding the standardized data based on the dynamic feature decoupling matrix of the standardized data to determine the low-dimensional features corresponding to the standardized data.

[0038] After completing the standardization processing of the sensor data of the target domain device, the embodiment of the present invention further utilizes a preset autoencoder model in combination with a dynamic feature decoupling matrix to encode the standardized data to extract its corresponding low-dimensional feature representation.

[0039] Specifically, the autoencoder employed in the embodiments of the present invention is a pre-trained neural network structure consisting of an encoder and a decoder. The encoder compresses high-dimensional, standardized data into a low-dimensional embedding space, forming a discriminative low-dimensional feature vector; the decoder reconstructs the original input data to assist the model in learning effective feature representations. To improve interpretability and separability during feature extraction, the present invention introduces a dynamic feature decoupling matrix, which can separate key components of the signal during the encoding process. For example, these components include high-frequency transient components (such as impact vibrations); low-frequency periodic components (such as rotational imbalances); and modulated mixed signal components (such as frequency modulation caused by gear faults). This mechanism effectively decouples the time-frequency mixed features in complex signals, thereby enhancing the structural clarity of the feature space and improving the differentiation between different fault types. For example, after encoding the vibration signal of the gearbox target domain data, its low-dimensional features will retain the kurtosis characteristics of the impact component and the spectral characteristics of the periodic vibration.

[0040] Furthermore, the autoencoder used in this embodiment of the present invention incorporates nonlinear transformations to enhance the model's ability to model complex signals. Compared to traditional linear feature extraction methods, this approach can more accurately capture nonlinear variations in signals and extract more representative low-dimensional features.

[0041] Step S108: classify and identify the low-dimensional features to determine the equipment fault type diagnosis result corresponding to the low-dimensional features.

[0042] Step S110: determining the fault status corresponding to the target domain device based on the device fault type diagnosis result.

[0043] After standardizing the sensor data of the target domain equipment and extracting low-dimensional features, embodiments of the present invention further classify and identify the extracted low-dimensional features to determine the corresponding equipment fault type. Specifically, the low-dimensional feature vector encoded by the autoencoder is input into a pre-trained classifier model, which outputs a probability distribution of fault types. Based on the probability distribution output by the classifier, the fault type corresponding to the highest probability is selected as the diagnosis result (e.g., normal, bearing inner race fault, gear tooth breakage, etc.). The classifier can be a softmax classifier, support vector machine (SVM), random forest, or other supervised learning model that has been trained on historical datasets encompassing multiple source domains (different equipment and operating conditions) and exhibits good cross-domain generalization capabilities. In one embodiment, training samples include normal conditions and multiple fault types. Domain labels (e.g., source domain 1, source domain 2, source domain 3, etc.) can be assigned to each sample for model training.

[0044] In summary, the embodiment of the present invention provides a device sensor data processing method for generalized fault diagnosis, which determines multi-source domain standardized calculation parameters based on multiple device types or operating conditions, and accordingly standardizes the sensor data of the target domain device to eliminate the scale differences caused by the inherent characteristics of the device and the sensor configuration. Furthermore, a preset autoencoder is combined with a dynamic feature decoupling matrix to decouple the time-frequency features in the mixed signal, thereby improving the interpretability and separability of the feature space. In addition, the nonlinear transformation term is used to enhance the model's ability to express complex signals, thereby extracting more representative low-dimensional features. Based on this, it can ensure that accurate and reliable fault detection and identification can be achieved under different devices and different operating conditions.

[0045] Furthermore, based on the above embodiment, the embodiment of the present invention also provides another device sensor data processing method for generalized fault diagnosis. Figure 2 The flowchart corresponding to the embodiment of the present invention is shown. Figure 2 , the method comprises the following steps:

[0046] Step S202: Acquire sensor data of a preset target domain device.

[0047] Step S204 : Based on the preset multi-source domain normalization calculation parameters, the sensor data is normalized to obtain normalized data corresponding to the sensor data.

[0048] In specific implementations, embodiments of the present invention standardize sensor data based on the numerical mean and numerical standard deviation of preset multi-source domain normalization calculation parameters to obtain standardized data corresponding to the sensor data. Prior art standardization typically only calculates the mean and standard deviation for data from a single device or single domain, making it difficult to address the misalignment of feature distributions across devices. This results in inconsistent initial distributions of source and target domain data, increasing the difficulty of subsequent domain adaptation optimization. The multi-source domain normalization calculation parameters of embodiments of the present invention are calculated based on data characteristics corresponding to sample data from all source domains, encompassing sensor signals collected under different device types, different operating conditions, and various fault conditions. Specifically, embodiments of the present invention obtain sensor data from multiple source domain devices; comprehensively arrange the sensor data from the multiple source domain devices according to preset data characteristic classification rules to determine the data distribution sequence of each data characteristic in each source domain; the data characteristics include one or more of time domain statistical characteristics, frequency domain energy distribution characteristics, and signal amplitude distribution characteristics; based on the data distribution sequence, the numerical mean and numerical standard deviation corresponding to each data characteristic are calculated; and the numerical mean and numerical standard deviation are used as the preset multi-source domain normalization calculation parameters.

[0049] In summary, the embodiments of the present invention address data distribution differences caused by different device structures by performing joint statistical modeling of data across devices and operating conditions, achieving unified scale processing. This effectively eliminates dimensional differences in sensor data from target domain devices (for example, inconsistent units of vibration signal amplitudes collected by different accelerometers), ensures that different features have the same scale expression, and improves the stability and generalization capabilities of subsequent model processing. Even if the vibration amplitude ranges of device A and device B may differ by several times, the standardized processing of the embodiments of the present invention can bring the features of different devices into the same distribution range, providing a unified range of feature values ​​for cross-domain training.

[0050] In specific implementation, the present invention uses a hierarchical alignment normalization method to preprocess the input device data. Taking into account the overall data distribution, local statistical characteristics within the domain, and cross-domain category structure differences, the global mean and variance are first calculated based on all source domain data. Then, the intra-category alignment factor and the inter-domain drift compensation term are used to achieve a more generalizable data scale normalization, which can be expressed as:

[0051]

[0052] Where, To hierarchically align the standardized input device data, a multi-level distributed alignment approach is used. This not only unifies the global scale but also fine-tunes the offset within the domain and the scale differences between categories, ensuring consistent scale representation across devices and working conditions.

[0053] is the first The global mean of the dimensional features; is the first The global standard deviation of the dimension feature; is the source domain offset compensation term, which performs adaptive offset adjustment according to the current data domain to solve the problem of "alignment error accumulation" in cross-domain applications of traditional standardization methods. It is defined as ; Represents the mean of the r-th dimension feature in the d-th domain where the current input sample is located.

[0054] It is the intra-category scale alignment factor, which is used to improve the perception of scale differences between multiple fault categories, maintain the scale information of feature differences between fault categories, and avoid masking inter-class diagnostic features due to unified standardization.

[0055] The calculation method is expressed as ; is the adjustment factor, preferably, set to 0.3; is the total number of categories; is the mean of the r-th dimension feature in the c-th category (such as different fault types).

[0056] This is different from the existing technology. The standardization of the existing technology usually only calculates the mean and standard deviation of single-device or single-domain data. For example, the accelerometer of device A (the vibration signal of the device is collected by the accelerometer) has a range of ±10g, and that of device B is ±50g. The existing technology standardization will lead to misalignment of feature distribution across devices. The standardization method of the present invention unifies the two into the same dimensional space, eliminating the impact of inherent range differences between devices on the model, aligning the initial distribution of source domain and target domain data, and reducing the difficulty of subsequent domain adaptation optimization.

[0057] Furthermore, a preset autoencoder is used to perform nonlinear activation processing on the standardized data based on its dynamic feature decoupling matrix to determine the low-dimensional features corresponding to the standardized data. An autoencoder is a neural network structure consisting of an encoder and a decoder. The encoder is responsible for mapping the input device data to a latent low-dimensional space, while the decoder is responsible for restoring the device data from the latent low-dimensional space to the original input dimension. The autoencoder trains the model by minimizing the difference between input and output, enabling the model to learn a compact representation of the input device data. Existing feature extraction methods have difficulty effectively decoupling the time-frequency mixed features in complex signals, resulting in poor separability of the feature space and significant overlap of feature points of different fault types, affecting the accuracy of fault diagnosis.

[0058] During the autoencoder processing stage of this embodiment of the present invention, input device data is encoded and mapped into a latent low-dimensional space. Each encoder layer extracts features through a linear transformation. Furthermore, this embodiment of the present invention utilizes an implicit gradient scaling mechanism, so that the output of each layer is influenced not only by the features of the previous layer but also by the dynamic adjustment of the gradient magnitude, thereby reducing the problems of vanishing gradients and information loss when processing high-dimensional device data. Gradient scaling adaptively adjusts weight updates to address the complex time-frequency features of vibration signals (such as the superposition of impact components and periodic vibrations), preventing gradient explosion when capturing high-frequency components and vanishing gradients when learning low-frequency trend features.

[0059] During specific implementation, refer to the following steps S206-S210.

[0060] Step S206: Obtain a preset nonlinear transformation item corresponding to the standardized data.

[0061] The role of the nonlinear transformation term is to consider additional nonlinear transformations during the encoding process. The encoding process in traditional autoencoders is often a linear mapping (such as Wx + b), which has the defect of causing the gradient of the Sigmoid activation function to disappear in the saturation region, which easily leads to the loss of signal details. Taking the broken tooth fault of a gear as an example, the traditional linear encoding may lose the time-frequency details of the transient pulse, while the nonlinear transformation term retains the kurtosis characteristics of the signal through nonlinear transformation, so that the encoded low-dimensional features can distinguish between normal and fault states. In order to better capture the complex relationship in the equipment data. The embodiment of the present invention uses the hyperbolic tangent function to perform nonlinear activation processing on the encoding process of the autoencoder, determine the corresponding nonlinear transformation term, and enhance the expression ability of the transient characteristics of the signal. For example, the non-stationary shock waveform generated in the early stage of a gearbox failure avoids the potential information loss caused by simple linear mapping. In the specific implementation, the calculation method of the nonlinear transformation term is expressed as:

[0062]

[0063] Where, is a nonlinear transformation term. is the hyperbolic tangent function; is the nonlinear transformation term weight, preferably, set to 0.3. is the weight of the nonlinear transformation term, which is a training parameter and is updated by the gradient descent method; It is the weight bias term of the nonlinear transformation term, which is a training parameter and is updated by the gradient descent method. To standardize data.

[0064] Step S208: Obtain a dynamic feature decoupling matrix obtained by pre-decoupling the dynamic features of the standardized data.

[0065] Among these, redundant information exists between the features of device data. When vibration signals collected by multiple accelerometers exhibit strong correlation (e.g., repeated perception of the same vibration source by adjacent measurement points), the independence of feature learning is affected. The present invention employs a dynamic feature decoupling method to independently optimize the learning process of each feature during encoder training. A dynamic feature decoupling matrix is ​​then used to control the learning rate of each feature, minimizing the mutual influence between features and addressing the redundancy and spatial heterogeneity of multi-sensor vibration signals. Furthermore, considering the nonlinear relationships between features, embodiments of the present invention also adjust the learning rate of each feature based on a weighted term for local feature correlation, thereby improving the quality of the device data representation after dimensionality reduction. This differs from existing dimensionality reduction methods (e.g., principal component analysis), which assume linear independence between features and struggle to address the nonlinear redundancy of multi-sensor vibration signals (e.g., correlation between vibration signals at adjacent measurement points).

[0066] Specifically, the calculation method of the dynamic feature decoupling matrix includes the following steps:

[0067] a- Obtain sensor data corresponding to multiple source domain devices collected in advance to obtain multiple source domain feature parameters. The multiple source domain devices include the target domain device corresponding to the standardized data.

[0068] b- Construct graph data corresponding to the multi-source domain feature parameters; based on the graph data, calculate the cosine similarity between each feature parameter in the multi-source domain feature parameters, and determine the similarity weight parameter between the feature parameters.

[0069] c- Based on the similarity weight parameter, determine the non-Euclidean space distance corresponding to the multi-source domain feature parameters.

[0070] The calculation of non-Euclidean space distance measures the similarity between device data points by constructing a graph, and each point in the graph is a sample point. Each data point can be regarded as a node in the graph to determine which data points need to be taken into consideration. Furthermore, according to the characteristics of the data and application requirements, a suitable similarity measurement method can be selected to calculate the similarity and establish edges. For example, if the similarity exceeds the set threshold, an edge is established between the two nodes, thereby constructing a graph model that can reflect the similarity between data points. The graph distance calculation can be obtained based on the distance and connection weight of each sample point in the graph, expressed as:

[0071]

[0072] Where, It is the non-Euclidean space distance calculated through the graph. is the weight between the i-th node (corresponding to the i-th sample) and the j-th node (corresponding to the j-th sample) in the graph; is the eigenvector corresponding to the i-th sample in the encoded low-dimensional feature matrix; is the first in the encoded low-dimensional feature matrix The feature vector corresponding to each sample.

[0073] The weights between nodes are calculated by adapting the scale factor to handle the heterogeneity between data points of different devices. The calculation method is expressed as:

[0074]

[0075] Where, is the weight between the i-th node (corresponding to the i-th sample) and the j-th node (corresponding to the j-th sample) in the graph. is the adaptively adjusted scale factor, which characterizes the local scale of the similarity between the two device data points. Preferably, the adaptively adjusted scale factor is calculated by Node (corresponding to samples) and Node (corresponding to The cosine similarity of samples) is obtained; is a positive integer; Is a positive integer.

[0076] d-Determine the inter-feature relationships corresponding to the multi-source domain feature parameters based on the Pearson correlation coefficient between each feature parameter in the multi-source domain feature parameters.

[0077] The embodiment of the present invention measures the linear correlation between two variables by the Pearson correlation coefficient, whose value range is [-1, 1]. For example, 1 is completely positive correlation, 0 is not linear correlation, and -1 is completely negative correlation. The encoded low-dimensional features and The Pearson correlation coefficient between the encoded low-dimensional features , For the encoded low-dimensional features; For the The two features can be features corresponding to the same domain or features corresponding to different domains.

[0078] e-Based on the non-Euclidean spatial distance, the relationship between features and the preset decoupling matrix adjustment factor, a dynamic feature decoupling matrix corresponding to the standardized data is constructed.

[0079] In specific implementation, the calculation method of the dynamic feature decoupling matrix refers to the following formula:

[0080]

[0081] Where, is the dynamic feature decoupling matrix. is the non-Euclidean distance calculated by the graph, For the The encoded low-dimensional features and The Pearson correlation coefficient between the encoded low-dimensional features. It is a decoupling matrix adjustment factor used to suppress high-amplitude strongly activated features, preventing certain feature dimensions from being too significant and suppressing the learning of other dimensions. It is a regulatory factor that controls the correlation between features. Its function is to control the intensity of information interaction between features and retain key diagnostic information during the dimensionality reduction process. Set to 0.2, Set to 0.3. It is the overall L2 norm of the encoded low-dimensional feature matrix, which indicates the overall activity of the feature.

[0082] This is different from existing technologies. Existing dimensionality reduction methods (such as principal component analysis) assume that features are linearly independent and are difficult to handle the nonlinear redundancy of multi-sensor vibration signals (such as the correlation between vibration signals at adjacent measuring points).

[0083] In step S210 , based on the adaptive weight matrix and nonlinear transformation terms of the autoencoder, nonlinear activation processing is performed on the dynamic feature decoupling matrix to determine the low-dimensional features corresponding to the standardized data.

[0084] The adaptive weight matrix of the autoencoder is determined based on a preset weight update amount. Specifically, the method for determining the preset weight update amount includes the following steps:

[0085] a-Get the loss function of the autoencoder, as well as the partial derivative of the loss function with respect to the weights corresponding to each iteration of the autoencoder.

[0086] b-Determine the dynamic gradient scaling factor corresponding to the loss function based on the partial derivative and the square of the historical partial derivative corresponding to the partial derivative.

[0087] The dynamic gradient scaling factor controls the strength of gradient scaling and adaptively adjusts the gradient based on the influence of historical gradients. The gradient scaling factor combines the current gradient norm with the squared historical gradient term to suppress high-frequency features (large gradients) and enhance low-frequency features (small gradients). For example, when capturing gearbox impact signals, the gradient scaling mechanism can automatically reduce the gradient amplitude corresponding to high-frequency noise to avoid weight oscillation. When learning the slowly changing trend of bearing wear, the gradient is amplified to accelerate convergence. The calculation method is expressed as:

[0088]

[0089] Where, is the gradient scaling adjustment factor; represents the square of the partial derivative of the autoencoder loss function with respect to the weights for the first t iterations; is the L2 norm; is the adjustment factor for controlling the gradient variance; is a positive integer. Preferably, Set to 0.2, Set to 0.1. The gradient adjustment factor focuses on adjusting the instantaneous gradient amplitude to prevent instability caused by excessively large or small current gradients. The adjustment factor controlling the gradient variance combines the historical sliding window effect of the squared gradient to suppress oscillations caused by rapid weight changes. Together, these two factors balance response speed and numerical stability.

[0090] c-Determine the implicit gradient scaling update amount corresponding to the partial derivative based on the dynamic gradient scaling factor.

[0091] d- Determine the corresponding weight update amount of the autoencoder based on the implicit gradient scaling update amount.

[0092] The embodiment of the present invention uses an implicit gradient scaling mechanism to update the model. The implicit gradient scaling mechanism uses the scaled weight update amount to adaptively adjust the gradient size according to the gradient value of each layer. If the gradient is small, the implicit gradient scaling mechanism will increase the gradient value. If the gradient is too large, the implicit gradient scaling mechanism will reduce the gradient to improve the convergence speed and stability of the network at different stages. Specifically, the calculation method of the weight update amount is expressed as:

[0093]

[0094] Where, represents the amount of weight update after scaling; is the partial derivative of the autoencoder loss function with respect to the weights, is the loss function of the autoencoder; is the symbol of partial derivative; is the dynamic gradient scaling factor, calculated by the above steps.

[0095] In combination with the above steps, the data encoding process of the embodiment of the present invention is expressed as follows:

[0096]

[0097] Where, is the low-dimensional feature after encoding; is the Sigmoid activation function; is the weight matrix of the encoder; is the bias vector of the encoder; is a nonlinear transformation term. Represents the scaled weight update amount. Represents the normalized input device data. is a nonlinear transformation term. This differs from existing technologies, where the encoding layer of autoencoders typically uses a fixed learning rate or a simple linear activation function, making it difficult to process the mixed characteristics of high-frequency transient components (such as bearing fault impact) and low-frequency periodic vibrations (such as the gear meshing fundamental frequency) in vibration signals.

[0098] Furthermore, in the decoder processing stage of the autoencoder, the low-dimensional representation after dimensionality reduction is reconstructed back to the original high-dimensional features. The goal is to minimize the reconstruction error, thereby restoring the original device data while maintaining the main features of the device data. When calculating the error, the decoder also uses the implicit gradient scaling mechanism to dynamically adjust the gradient value to ensure the stability of the training process, which is expressed as:

[0099]

[0100] Where, It is the reconstructed device data after decoding; is the weight matrix of the decoder; is the bias vector of the decoder.

[0101] Furthermore, in order to verify the data processing effect corresponding to the embodiment of the present invention, the embodiment of the present invention also provides a characteristic distribution effect schematic diagram, referring to Figure 3 In this experiment, a double scatter plot was used to compare the separability of the feature space. Among them, the feature distribution of the traditional method showed a fuzzy linear clustering (left figure), and the feature points of different fault types (distinguished by color) overlapped significantly. The feature points of the method of the present invention (right figure) formed three clear concentric rings. The Euclidean distance of similar samples is reduced. This is because the nonlinear transformation term in the encoder enhances the expression ability of the kurtosis feature of the impact signal through the hyperbolic tangent function. At the same time, the dynamic feature decoupling matrix reduces the correlation interference of multi-sensor signals. In addition, the annular distribution pattern of the feature points of the present invention in the experimental data is highly consistent with the spectral characteristics of the gearbox fault, which confirms the decoupling ability of the method for time-frequency mixed features.

[0102] Step S212: classify and identify the low-dimensional features to determine the equipment fault type diagnosis result corresponding to the low-dimensional features.

[0103] Step S214: Determine the fault status corresponding to the target domain device based on the device fault type diagnosis result.

[0104] In summary, the embodiments of the present invention provide another method for processing device sensor data for generalized fault diagnosis by proposing a unified cross-domain standardization method. Compared to existing technologies that standardize data only for a single device or a single domain, the embodiments of the present invention unify the features of multi-source domain data (different devices and operating conditions) into the same dimensional space, eliminating the impact of inherent range differences between devices on the model. This aligns the initial distributions of source and target domain data, reducing the difficulty of subsequent domain adaptation optimization. Furthermore, the embodiments of the present invention pre-decouple the dynamic features of the data to generate a dynamic feature decoupling matrix. This can control the learning rate of each feature, minimize the mutual influence between features, avoid falling into local optimality, and address the redundancy and spatial heterogeneity of multi-sensor vibration signals.

[0105] In addition, the embodiment of the present invention uses the hyperbolic tangent function to perform nonlinear activation processing on the encoding process of the autoencoder, determines the corresponding nonlinear transformation terms, and then determines the low-dimensional features corresponding to the data. This method can enhance the ability to express the transient characteristics of the signal and avoid the potential information loss caused by simple linear mapping. Furthermore, the embodiment of the present invention also dynamically adjusts the gradient size based on the implicit gradient scaling mechanism to determine the weight update amount corresponding to the autoencoder, and then determines its adaptive weight matrix. It can effectively identify the mixed features of high-frequency transient components and low-frequency periodic vibrations, and solve the problems of gradient disappearance and gradient explosion in high-dimensional device data processing.

[0106] Furthermore, based on the above embodiment, the embodiment of the present invention also adaptively adjusts the gradient of the loss function of the autoencoder with respect to the weight based on the above dynamic feature decoupling matrix to update the weight parameters of the autoencoder.

[0107] In a specific implementation, during the backpropagation process, the reconstruction error of the autoencoder is propagated layer by layer through the network to calculate the gradient. The gradient is updated based on the adaptive adjustment of the dynamic feature decoupling matrix to improve the stability and convergence of the training process, which is expressed as:

[0108]

[0109]

[0110] Where, It is the parameter update operation; is the learning rate of the autoencoder; It is a dynamic feature decoupling matrix. Different from the prior art, the back propagation of the autoencoder training process in the prior art uses a fixed learning rate to update the weights, which easily leads to training oscillations. The present invention integrates the dynamic feature decoupling matrix into the parameter update process to achieve feature-level learning rate regulation. For example, in the early stages of training, the diagonal values ​​of the dynamic feature decoupling matrix corresponding to highly correlated sensor features (such as the axial and radial vibrations of the same bearing seat) are low, which suppresses the amplitude of their parameter updates and avoids falling into the local optimum too early. Independent features (such as temperature sensors and vibration signals) maintain a normal learning rate, accelerating the model's convergence to the global optimum.

[0111] Furthermore, in combination with the above embodiment, the weight update amount is determined based on the loss function of the autoencoder. Among them, the autoencoder of the prior art usually uses the mean square error as the loss function, which is sensitive to noise and lacks cross-domain constraints. It is difficult to improve the robustness of the model to noise and outliers, and it is also impossible to effectively optimize the cross-domain generalization ability. In order to enhance the robustness of the autoencoder model to noise and outliers, the embodiment of the present invention uses sample weighting terms to reduce the impact of outliers on error calculation. In specific implementation, the loss function calculation method of the autoencoder of the embodiment of the present invention includes the following steps:

[0112] 1) Obtain a preset training sample set and train the autoencoder based on the training sample set.

[0113] The training sample set includes sensor parameters of multiple source domain devices, wherein the specific data can refer to the sensor data of the target domain device.

[0114] 2) Calculate the reconstruction error corresponding to the training sample set, and based on the reconstruction error, adaptively determine the sample weighting item parameters corresponding to the training sample set.

[0115] The sample weighting term can better suppress the impact of outliers and prevent the reconstruction error from excessively amplifying the error caused by noisy device data, thereby improving the robustness of the autoencoder model. For example, when the sensor collects an outlier (signal distortion caused by transient electromagnetic interference), the sample weighting term adaptively calculates the error before reconstruction of the sample, resulting in a smaller sample weighting term value, automatically reducing the contribution of the sample to the loss function. The calculation method is expressed as:

[0116]

[0117] Where, is the weighted term of the i-th sample, is the L2 norm; is the weighted item adjustment factor. Preferably, Set to 0.3. Indicates the Normalized input device data; For the The decoded reconstructed device data.

[0118] 3) Based on the interaction strength and mutual influence strength between the features of the training sample set, the information flow parameters corresponding to the training sample set are calculated.

[0119] The role of the information flow term is to constrain the high-order dependencies between features. It is calculated based on the strength of the mutual influence between features during the dimensionality reduction process. For example, in the processing of gearbox vibration data, the information flow term can automatically establish the nonlinear dependency between the meshing frequency component and the sideband features. The calculation method is expressed as:

[0120]

[0121] Where, is the eigenvector corresponding to the i-th sample in the encoded low-dimensional feature matrix; is the eigenvector corresponding to the jth sample in the encoded low-dimensional feature matrix. To adjust the hyperparameters of high-order information flow, used to control the impact of gradient changes on regularization. Preferably, Set to 0.2. yes and The Pearson correlation coefficient between the features is used to capture the nonlinear dependencies between the features by calculating the correlation and gradient changes between the features, thereby adaptively adjusting the information flow path during the training process. is the gradient of the eigenvector corresponding to the i-th sample in the encoded low-dimensional feature matrix; is the gradient of the eigenvector corresponding to the jth sample in the encoded low-dimensional feature matrix.

[0122] 4) Project the training sample set into a high-dimensional space, calculate the distribution difference corresponding to each source domain device in the training sample set, and determine the maximum mean difference parameter corresponding to the training sample set.

[0123] The embodiment of the present invention also constrains the generalization ability of the autoencoder based on the calculated loss function. Specifically, the embodiment of the present invention implements the generalization ability constraint of the autoencoder through the maximum mean difference term and the adversarial loss term to minimize the distribution difference between the source domain and the target domain. The calculation method of the maximum mean difference term is expressed as:

[0124]

[0125] Where, is the maximum mean difference term, It is a kernel function (such as a Gaussian kernel) that projects the original device data into a high-dimensional space through mapping to calculate more accurate distribution differences. is the feature corresponding to the i-th sample in the source domain; is the feature corresponding to the i-th sample in the target domain. is the number of samples in the source domain; is the number of samples in the target domain.

[0126] 5) The features of each source domain device in the training sample set are distinguished by a preset domain classifier, and the feature representation of the training sample set is adjusted using an autoencoder to determine the domain classifier recognition adversarial loss parameters corresponding to the training sample set.

[0127] In order to enhance the model's ability to learn domain-invariant features, embodiments of the present invention also improve the model's generalization ability on unknown device data through an adversarial training strategy. Adversarial training can be implemented using a preset domain classifier (such as a Softmax classifier function), which attempts to distinguish between the features of the source domain and the target domain. During training, the encoder adjusts its feature representation to confuse the domain classifier, making it unable to effectively distinguish between the source domain and the target domain. The domain classifier attempts to output labels for the source domain and the target domain (1 for the source domain and 0 for the target domain). The adversarial loss term is calculated as follows:

[0128]

[0129] Where, is the counter-loss term. express expectations; represents the recognition probability of the domain classifier whose input is the source domain sample, Represents the recognition probability of the domain classifier when the input is a sample from the target domain.

[0130] 6) Calculate the loss function corresponding to the autoencoder based on the sample weighting parameter, information flow parameter, maximum mean difference parameter, and domain classifier recognition adversarial loss parameter.

[0131] In summary, in order to enhance the robustness of the autoencoder model to noise and outliers, the embodiment of the present invention uses sample weighting terms to reduce the impact of outliers on error calculation. The loss function calculation method of the autoencoder is expressed as:

[0132]

[0133] Where, is the total loss function; is the information flow adjustment factor, is an information flow item; preferably, Set to 0.2. It is a hyperparameter that controls the weight of the domain-invariant loss term and the learning intensity of the control domain-invariant features. is the maximum mean difference term. is the weight of the adversarial loss term; is the counter-loss term.

[0134] The number of samples input to the autoencoder for the current batch; is a positive integer; is the weighted term of the i-th sample. Represents the input device data after standardization of i; is the reconstructed device data after the i-th decoding. is the source domain feature, and ; is the target domain feature, and .

[0135] The autoencoder of the embodiment of the present invention is pre-trained and can be iterated based on the above steps (e.g., weight update amount, loss function, dynamic feature decoupling matrix, etc.) until a preset stopping condition is met, indicating that model training is complete. In one embodiment, the preset stopping condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000.

[0136] In summary, the embodiment of the present invention constrains the generalization ability of the autoencoder during its training process, so that the autoencoder can be trained on the source domain and used on the target domain. Based on this, the embodiment of the present invention provides a domain-invariant autoencoder to perform feature dimensionality reduction. In addition, the constraints of sample weighting terms, information flow terms, maximum mean difference terms and adversarial loss terms are adopted in the loss function of the autoencoder to enhance the robustness of the model to noise and outliers, while improving the generalization ability of the model through cross-domain constraints. Compared with the prior art autoencoder training process that uses a fixed learning rate to update weights, which easily leads to training oscillations, the present invention uses a dynamic gradient scaling factor and a dynamic feature decoupling matrix to adaptively adjust the gradient update amount during the autoencoder training process, thereby improving the stability and convergence speed of the training process.

[0137] Furthermore, in order to verify the robustness of the embodiment of the present invention to sensor noise, the embodiment of the present invention also provides a noise robustness comparison result diagram, refer to Figure 4 ,exist Figure 4 In the figure, the horizontal axis is set to the input noise intensity (normalized parameter ranging from 0 to 0.5), and the vertical axis shows the mean squared reconstruction error. As the noise intensity increases, the reconstruction error of the traditional autoencoder shows a steep upward trend. However, the upward trend of the reconstruction error of the proposed method is significantly smaller than that of the traditional autoencoder, and the slope of the curve is significantly flatter, indicating that the sample weighting mechanism in the loss function can automatically reduce the weight of abnormal samples. At the same time, combined with the weight update value based on the implicit gradient scaling mechanism mentioned above, the implicit gradient scaling mechanism suppresses the gradient mutation caused by noise, and the dynamic gradient adjustment improves training stability.

[0138] Furthermore, the embodiment of the present invention also provides a diagram of the training effect corresponding to the loss function, referring to Figure 5This experiment shows the optimization effect of the training process through a double Y-axis graph. The primary vertical axis shows the loss value curve, the secondary vertical axis is the verification accuracy, and the horizontal axis is the training round. The loss curve of the traditional method (blue dotted line) still fluctuates greatly after 50 rounds, and the accuracy only slowly increases to about 8%; while the loss value of the method of the present invention (red solid line) enters a stable convergence state after 30 rounds, and the accuracy quickly climbs to more than 90%, indicating that the adjustment mechanism of the dynamic gradient scaling factor automatically reduces the update amount when the gradient norm is large in the early stage of training to prevent the weight matrix from mutating; in the later gradient decay stage, the update amount is amplified to accelerate convergence, and the steep rising section of the accuracy curve of the present invention (10-30 rounds) corresponds to the smooth transition of the loss curve, indicating the dual improvement effect of implicit gradient scaling on convergence speed and stability.

[0139] In summary, the embodiment of the present invention also provides a schematic diagram of cross-domain diagnostic performance comparison results, referring to Figure 6 , shows the performance comparison of different methods in cross-domain fault diagnosis scenarios. Figure 6 In the experimental data, the horizontal axis represents the number of target domains (3 to 10 different equipment working conditions), the vertical axis represents the fault diagnosis accuracy, and the error bar reflects the data volatility. The experiment compares the traditional autoencoder, principal component analysis (PCA) dimensionality reduction, single source domain training and the method of the present invention. When the number of target domains increases from 3 to 10, the accuracy of the traditional method drops from about 70% to about 60% due to the domain offset problem. The accuracy of the method of the present invention is always stable in the range of 80%-85% through domain-invariant feature learning and dynamic feature decoupling mechanism, indicating that the cross-domain fault diagnosis accuracy of the method of the present invention is significantly better than other methods, and the error range is smaller. It verifies that the cross-domain normalization and maximum mean difference term constraints effectively eliminate the distribution differences between devices, so that the model can still maintain a stable feature extraction capability when facing unknown device data.

[0140] Furthermore, based on the above embodiment, the embodiment of the present invention also provides a device sensor data processing device for generalized fault diagnosis. Figure 7 The schematic diagram of the structure corresponding to the embodiment of the present invention is shown. Figure 7The device includes: a data acquisition module 100, which is used to acquire sensor data of a preset target domain device; the preset target domain device is used to characterize a device to be diagnosed of a preset device type, or a device to be diagnosed under a preset working condition; a data processing module 200, which is used to standardize the sensor data based on a preset multi-source domain standardization calculation parameter to obtain standardized data corresponding to the sensor data; wherein the multi-source domain standardization calculation parameter is calculated based on data characteristics corresponding to sensor data of multiple source domain devices; an execution module 300, which is used to use a preset autoencoder to encode the standardized data based on a dynamic feature decoupling matrix of the standardized data to determine the low-dimensional features corresponding to the standardized data; wherein the autoencoder determines the low-dimensional features corresponding to the standardized data based on a preset nonlinear transformation term; a classification module 400, which is used to classify and identify the low-dimensional features to determine a diagnosis result of a device fault type corresponding to the low-dimensional features; and an output module 500, which is used to determine a fault state corresponding to the target domain device based on the device fault type diagnosis result.

[0141] An embodiment of the present invention provides an apparatus for processing device sensor data for generalized fault diagnosis, which has the same technical features as the above-mentioned method embodiment, and can therefore solve the same technical problems and achieve the same technical effects.

[0142] Furthermore, the classification module 400 is also used to: input low-dimensional features into a pre-trained classifier, output the fault type probability distribution corresponding to the low-dimensional features; and determine the equipment fault type diagnosis result corresponding to the low-dimensional features based on the fault type probability distribution.

[0143] The data processing module 200 is further configured to perform normalization processing on the sensor data based on the numerical mean and numerical standard deviation of the preset multi-source domain normalization calculation parameters to obtain normalized data corresponding to the sensor data. The method for calculating the preset multi-source domain normalization calculation parameters includes: obtaining sensor data from multiple source domain devices; comprehensively arranging the sensor data from the multiple source domain devices according to preset data characteristic classification rules to determine the data distribution sequence of each data characteristic in each source domain; the data characteristics include one or more of time domain statistical characteristics, frequency domain energy distribution characteristics, and signal amplitude distribution characteristics; based on the data distribution sequence, respectively calculating the numerical mean and numerical standard deviation corresponding to each data characteristic; and using the numerical mean and numerical standard deviation as the preset multi-source domain normalization calculation parameters.

[0144] The above-mentioned execution module 300 is also used to: obtain preset nonlinear transformation terms corresponding to the standardized data; and, pre-decouple the dynamic features of the standardized data to obtain a dynamic feature decoupling matrix; based on the adaptive weight matrix and nonlinear transformation terms of the autoencoder, perform nonlinear activation processing on the dynamic feature decoupling matrix to determine the low-dimensional features corresponding to the standardized data; the adaptive weight matrix is ​​determined based on a preset weight update amount.

[0145] The execution module 300 is further configured to perform nonlinear activation processing on the encoding process of the autoencoder using a preset nonlinear activation function, and construct a nonlinear transformation term corresponding to the standardized data.

[0146] The above-mentioned execution module 300 is also used to: obtain sensor data corresponding to multiple source domain devices collected in advance to obtain multi-source domain feature parameters; the multiple source domain devices include target domain devices corresponding to the standardized data; construct graph data corresponding to the multi-source domain feature parameters; based on the graph data, calculate the cosine similarity between each feature parameter in the multi-source domain feature parameters, and determine the similarity weight parameter between the feature parameters; based on the similarity weight parameter, determine the non-Euclidean space distance corresponding to the multi-source domain feature parameters; based on the Pearson correlation coefficient between each feature parameter in the multi-source domain feature parameters, determine the inter-feature relationship corresponding to the multi-source domain feature parameters; based on the non-Euclidean space distance, the inter-feature relationship and the preset decoupling matrix adjustment factor, construct a dynamic feature decoupling matrix corresponding to the standardized data.

[0147] The device also includes a training module for adaptively adjusting the gradient of the loss function of the autoencoder with respect to the weight based on the dynamic feature decoupling matrix to update the weight parameters of the autoencoder.

[0148] The above-mentioned execution module 300 is also used to: obtain the loss function of the autoencoder, and the partial derivative of the loss function with respect to the weight corresponding to each iteration of the autoencoder; determine the dynamic gradient scaling factor corresponding to the loss function based on the partial derivative and the square term of the historical partial derivative corresponding to the partial derivative; determine the implicit gradient scaling update amount corresponding to the partial derivative based on the dynamic gradient scaling factor; determine the weight update amount corresponding to the autoencoder based on the implicit gradient scaling update amount.

[0149] The above-mentioned execution module 300 is also used to: obtain a preset training sample set and train the autoencoder based on the training sample set; the training sample set includes sensor parameters of multiple source domain devices; calculate the reconstruction error corresponding to the training sample set, and based on the reconstruction error, adaptively determine the sample weighting item parameter corresponding to the training sample set; calculate the information flow parameter corresponding to the training sample set based on the interaction strength and mutual influence strength between the features of the training sample set; project the training sample set into a high-dimensional space, calculate the distribution difference corresponding to each source domain device of the training sample set, and determine the maximum mean difference parameter corresponding to the training sample set; distinguish the features of each source domain device in the training sample set through a preset domain classifier, and use the autoencoder to adjust the feature representation of the training sample set to determine the domain classifier identification adversarial loss parameter corresponding to the training sample set; calculate the loss function corresponding to the autoencoder based on the sample weighting item parameter, information flow parameter, maximum mean difference parameter and domain classifier identification adversarial loss parameter.

[0150] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the above Figures 1 to 2 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 8 FIG. 8 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 81 and a memory 80, the memory 80 stores computer executable instructions that can be executed by the processor 81, and the processor 81 executes the computer executable instructions to implement the above Figures 1 to 2 Either of the methods shown. Figure 8 In the illustrated embodiment, the electronic device further includes a bus 82 and a communication interface 83 , wherein the processor 81 , the communication interface 83 and the memory 80 are connected via the bus 82 .

[0151] Among them, the memory 80 may include high-speed random access memory (RAM), and may also include non-volatile memory (non-volatile memory), such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 83 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 82 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture, on-chip bus standard) bus, wherein AMBA defines three types of buses, including APB (Advanced Peripheral Bus) bus, AHB (Advanced High-performance Bus) bus and AXI (Advanced eXtensible Interface) bus. The bus 82 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0152] The processor 81 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the processor 81 or by software instructions. The above processor 81 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor 81 reads the information in the memory and combines its hardware to complete the above Figures 1 to 2 Any of the methods shown.

[0153] The embodiments of the present invention provide a computer program product for a device sensor data processing method and apparatus for generalized fault diagnosis, including a computer-readable storage medium storing program code. The program code includes instructions that can be used to execute the methods described in the aforementioned method embodiments. For specific implementations, please refer to the method embodiments and will not be described in detail here. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating processes of the system described above can refer to the corresponding processes in the aforementioned method embodiments and will not be described in detail here. Furthermore, in the description of the embodiments of the present invention, unless otherwise specified or limited, the terms "installed," "connected," and "connected" should be interpreted broadly. For example, they can refer to fixed, removable, or integral connections; mechanical or electrical connections; direct connections, indirect connections through an intermediary, or internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances. If the functions described are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0154] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0155] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A device sensor data processing method for generalized fault diagnosis, characterized in that: The method comprises: Acquiring sensor data of a preset target domain device; the preset target domain device is used to represent a device to be diagnosed of a preset device type, or a device to be diagnosed under a preset working condition; Based on preset multi-source domain normalization calculation parameters, the sensor data is normalized to obtain normalized data corresponding to the sensor data; wherein the multi-source domain normalization calculation parameters are calculated based on data characteristics corresponding to the sensor data of multiple source domain devices; Using a preset autoencoder, encoding the standardized data based on a dynamic feature decoupling matrix of the standardized data to determine low-dimensional features corresponding to the standardized data; wherein the autoencoder determines the low-dimensional features corresponding to the standardized data based on a preset nonlinear transformation term; Classify and identify the low-dimensional features to determine a diagnosis result of the equipment fault type corresponding to the low-dimensional features; Determining a fault state corresponding to the target domain device based on the device fault type diagnosis result; The dynamic feature decoupling matrix is ​​obtained by decoupling the dynamic features of the standardized data, and includes: Acquire sensor data corresponding to a plurality of pre-collected source domain devices, respectively, to obtain multiple source domain feature parameters; the plurality of source domain devices include a target domain device corresponding to the standardized data; Constructing graph data corresponding to the multi-source domain feature parameters; calculating the cosine similarity between each feature parameter in the multi-source domain feature parameters based on the graph data, and determining a similarity weight parameter between the feature parameters; Determining the non-Euclidean space distance corresponding to the multi-source domain feature parameters based on the similarity weight parameter; Determining, based on the Pearson correlation coefficient between each feature parameter in the multi-source domain feature parameters, the inter-feature relationship corresponding to the multi-source domain feature parameters; Based on the non-Euclidean spatial distance, the interrelationships between the features and a preset decoupling matrix adjustment factor, a dynamic feature decoupling matrix corresponding to the standardized data is constructed.

2. The method according to claim 1, characterized in that The step of classifying and identifying the low-dimensional features and determining the equipment fault type diagnosis result corresponding to the low-dimensional features includes: Inputting the low-dimensional features into a pre-trained classifier, and outputting a probability distribution of the fault type corresponding to the low-dimensional features; Based on the fault type probability distribution, a diagnosis result of the equipment fault type corresponding to the low-dimensional feature is determined.

3. The method according to claim 1, characterized in that The step of normalizing the sensor data based on preset multi-source domain normalization calculation parameters to obtain normalized data corresponding to the sensor data includes: Based on the numerical mean and numerical standard deviation of the preset multi-source domain normalization calculation parameters, the sensor data is normalized to obtain normalized data corresponding to the sensor data; The preset method for calculating the multi-source domain normalization calculation parameters includes: Obtain sensor data from multiple source domain devices; According to preset data characteristic classification rules, the sensor data of multiple source domain devices are comprehensively arranged to determine the data distribution sequence of each data characteristic in each source domain; the data characteristics include one or more of time domain statistical characteristics, frequency domain energy distribution characteristics, and signal amplitude distribution characteristics; Based on the data distribution sequence, respectively calculate the numerical mean and numerical standard deviation corresponding to each data characteristic; The numerical mean and numerical standard deviation are used as preset multi-source domain normalization calculation parameters.

4. The method according to claim 1, wherein The step of encoding the standardized data using a preset autoencoder based on a dynamic feature decoupling matrix of the standardized data and determining low-dimensional features corresponding to the standardized data includes: Obtaining a preset nonlinear transformation term corresponding to the standardized data; and a dynamic feature decoupling matrix obtained by pre-decoupling the dynamic features of the standardized data; Based on the adaptive weight matrix of the autoencoder and the nonlinear transformation term, nonlinear activation processing is performed on the dynamic feature decoupling matrix to determine the low-dimensional features corresponding to the standardized data; wherein the adaptive weight matrix is ​​determined based on a preset weight update amount.

5. The method according to claim 4, characterized in that The calculation method of the preset nonlinear transformation term includes: A nonlinear activation function is used to perform nonlinear activation processing on the encoding process of the autoencoder to construct a nonlinear transformation term corresponding to the standardized data.

6. The method according to claim 1, characterized in that The method further comprises: Based on the dynamic feature decoupling matrix, the gradient of the loss function of the autoencoder with respect to the weight is adaptively adjusted to update the weight parameters of the autoencoder.

7. The method according to claim 4, characterized in that The method for determining the preset weight update amount includes: Obtaining the loss function of the autoencoder and the partial derivative of the loss function with respect to the weight corresponding to each iteration of the autoencoder; Determining a dynamic gradient scaling factor corresponding to the loss function based on the partial derivative and a square term of a historical partial derivative corresponding to the partial derivative; Determining an implicit gradient scaling update corresponding to the partial derivative based on the dynamic gradient scaling factor; Based on the implicit gradient scaling update amount, a weight update amount corresponding to the autoencoder is determined.

8. The method according to claim 7, characterized in that The calculation method of the loss function includes: Obtaining a preset training sample set, and training the autoencoder based on the training sample set; the training sample set includes sensor parameters of multiple source domain devices; Calculating a reconstruction error corresponding to the training sample set, and adaptively determining a sample weighting item parameter corresponding to the training sample set based on the reconstruction error; Calculating information flow parameters corresponding to the training sample set based on interaction strength and mutual influence strength between features of the training sample set; Projecting the training sample set into a high-dimensional space, calculating the distribution difference corresponding to each source domain device of the training sample set, and determining the maximum mean difference parameter corresponding to the training sample set; Distinguishing features of each source domain device in the training sample set using a preset domain classifier, adjusting feature representations of the training sample set using the autoencoder, and determining domain classifier recognition adversarial loss parameters corresponding to the training sample set; A loss function corresponding to the autoencoder is calculated based on the sample weighting term parameter, the information flow parameter, the maximum mean difference parameter, and the domain classifier identification adversarial loss parameter.

9. A device sensor data processing device for generalized fault diagnosis, characterized in that: The device comprises: A data acquisition module, configured to acquire sensor data of a preset target domain device; the preset target domain device is used to represent a device to be diagnosed of a preset device type, or a device to be diagnosed under a preset working condition; a data processing module, configured to perform normalization processing on the sensor data based on preset multi-source domain normalization calculation parameters to obtain normalized data corresponding to the sensor data; wherein the multi-source domain normalization calculation parameters are calculated based on data characteristics corresponding to sensor data of multiple source domain devices; an execution module, configured to use a preset autoencoder to encode the standardized data based on a dynamic feature decoupling matrix of the standardized data, and determine low-dimensional features corresponding to the standardized data; wherein the autoencoder determines the low-dimensional features corresponding to the standardized data based on a preset nonlinear transformation term; A classification module, configured to classify and identify the low-dimensional features and determine a diagnosis result of the equipment fault type indicated by the low-dimensional features; An output module, configured to determine a fault state corresponding to the target domain device based on the device fault type diagnosis result; In which, the dynamic feature decoupling matrix is ​​obtained by decoupling the dynamic features of the standardized data, and the execution module is also used to: obtain sensor data corresponding to multiple source domain devices collected in advance to obtain multi-source domain feature parameters; the multiple source domain devices include target domain devices corresponding to the standardized data; construct graph data corresponding to the multi-source domain feature parameters; based on the graph data, calculate the cosine similarity between each feature parameter in the multi-source domain feature parameters, and determine the similarity weight parameter between the feature parameters; based on the similarity weight parameter, determine the non-Euclidean space distance corresponding to the multi-source domain feature parameters; based on the Pearson correlation coefficient between each feature parameter in the multi-source domain feature parameters, determine the inter-feature relationship corresponding to the multi-source domain feature parameters; based on the non-Euclidean space distance, the inter-feature relationship and the preset decoupling matrix adjustment factor, construct the dynamic feature decoupling matrix corresponding to the standardized data.

Images