Early fault diagnosis method and device for device health management and electronic device

Abstract

This application provides a method, apparatus, and electronic device for early fault diagnosis in equipment health management. The method includes: acquiring real-time collected equipment operation monitoring data; inputting the equipment operation monitoring data into a pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operating parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer; determining a fused feature vector through the multi-source feature fusion layer; outputting an attention-weighted feature vector through the multi-head adaptive gating attention layer; extracting multi-scale deep features from the attention-weighted feature vector and weighting and aggregating them through the multi-scale residual classification layer to output a fault category probability vector; and determining the early fault diagnosis result of the equipment health status based on the fault category probability vector. This application can improve the accuracy of early fault prediction for equipment.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an early fault diagnosis method, apparatus, and electronic device for equipment health management. Background Technology

[0002] As industrial equipment develops towards larger scale, higher speed, and continuous operation, critical rotating equipment operates under complex conditions for extended periods. Its health status directly impacts the safety, reliability, and operational economy of the production system. Before experiencing a serious failure, equipment often undergoes a relatively long early degradation phase. The vibration response characteristics generated during this phase are weak and susceptible to noise and fluctuations in operating conditions. However, accurate identification during this phase provides a sufficient time window for predictive maintenance and condition-based repair, effectively preventing sudden downtime and major accidents. Therefore, early fault diagnosis methods for equipment health management have become a key research focus and an urgent engineering need in the field of industrial intelligent operation and maintenance. Summary of the Invention

[0003] The purpose of this application is to provide an early fault diagnosis method, device, and electronic equipment for equipment health management, which can improve the accuracy of early fault prediction.

[0004] Firstly, this application provides an early fault diagnosis method for equipment health management. The method includes: acquiring real-time collected equipment operation monitoring data; the equipment operation monitoring data includes: the original time-domain vibration signal corresponding to the equipment vibration data and the equipment operation parameter vector; inputting the equipment operation monitoring data into a pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operation parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer; through the multi-source feature fusion layer, performing multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing on the equipment operation monitoring data to determine the fused feature vector; through the multi-head adaptive gating attention layer, dynamically adjusting the activation intensity of each feature dimension in the fused feature vector according to the equipment operation parameter vector to focus on the feature subspace most relevant to the fault, and outputting an attention-weighted feature vector; through the multi-scale residual classification layer, extracting multi-scale deep features from the attention-weighted feature vector and weighting and aggregating them to output a fault category probability vector, and determining the early fault diagnosis result of the equipment health status based on the fault category probability vector.

[0005] Furthermore, the steps described above, which involve multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing of equipment operation monitoring data through a multi-source feature fusion layer to determine the fused feature vector, include: performing the following steps through the multi-source feature fusion layer: performing multi-band adaptive spectrum reconstruction on the original time-domain vibration signal corresponding to the equipment vibration data to determine the reconstructed spectrum vector; using a joint recursive matrix construction method, combining the reconstructed spectrum vector with the envelope information of the original time-domain vibration signal to generate a time-frequency joint feature vector that can simultaneously characterize the time-domain dynamics and frequency-domain energy distribution pattern; and fusing the spectrum reconstruction vector, the time-frequency joint feature vector, and the equipment operation parameter vector to determine the fused feature vector.

[0006] Furthermore, the steps described above for performing multi-band adaptive spectrum reconstruction on the original time-domain vibration signal corresponding to the equipment vibration data and determining the reconstructed spectrum vector include: performing a fast Fourier transform on the original time-domain vibration signal corresponding to the equipment vibration data and taking the amplitude of a single-sided spectrum to obtain the corresponding spectrum amplitude vector; calculating the cumulative energy and gradient of the spectrum amplitude based on the spectrum amplitude vector, and adaptively determining the boundary points of the frequency band division by locating the local maxima of the cumulative energy gradient to separate the scattered and weak fault features into independent sub-bands; for each sub-band, adaptively calculating the enhancement coefficient based on the relative relationship between the sub-band average energy and the global average energy, and using the hyperbolic tangent function for nonlinear enhancement and normalization processing to obtain the amplitude of the reconstructed spectrum at each frequency index corresponding to each sub-band; and splicing multiple reconstructed spectra according to the original frequency order based on the amplitude of the reconstructed spectrum at each frequency index corresponding to all sub-bands to form a complete, feature-enhanced reconstructed spectrum vector.

[0007] Furthermore, the steps described above, which employ a joint recursive matrix construction method to combine the reconstructed spectral vector with the envelope information of the original time-domain vibration signal to generate a time-frequency joint feature vector capable of simultaneously characterizing time-domain dynamics and frequency-domain energy distribution patterns, include: extracting the envelope signal from the original time-domain vibration signal; dividing the envelope signal and the reconstructed spectral vector into several non-overlapping segments, and calculating the root mean square value and spectral centroid of each segment to obtain the envelope root mean square sequence and the spectral centroid sequence; constructing a joint recursive matrix based on the root mean square sequence and the spectral centroid sequence; and reshaping the joint recursive matrix into a one-dimensional vector as the time-frequency joint feature vector.

[0008] Furthermore, the training process of the aforementioned early equipment fault diagnosis model is as follows: Obtain the equipment condition monitoring data training set; train the multi-source feature fusion layer, multi-head adaptive gating attention layer, and multi-scale residual classification layer of the operating parameter modulation using the equipment condition monitoring data training set to obtain the early equipment fault diagnosis model; during the training process, a composite loss function combining adaptive focus cross-entropy loss and operating parameter perception feature consistency loss is used to calculate the loss value.

[0009] Furthermore, the steps for obtaining the equipment condition monitoring data training set mentioned above include: acquiring historical equipment operation monitoring data; the historical equipment operation monitoring data includes: continuously collecting vibration signal samples of the equipment at different operating stages and different health states at a fixed sampling frequency, and simultaneously collecting the equipment operation parameter vectors corresponding to the vibration signal samples; labeling each vibration signal sample and the corresponding equipment operation parameter vector in the equipment operation monitoring data with a health status label; and integrating the labeled vibration signal sequence, the corresponding equipment operation parameter vector and its fault category label to construct a structured equipment condition monitoring data training set.

[0010] Furthermore, the composite loss function is determined as follows: Based on the equipment operating parameter vector, the sample adaptive focus parameter is determined; based on the sample adaptive focus parameter and the predicted probability of the sample to the true category, the adaptive focus cross-entropy loss is determined; based on the equipment operating parameter vector, the bandwidth parameter of operating parameter similarity, and the indicator function based on the true fault category label, the operating parameter perceived feature consistency loss is determined; the adaptive focus cross-entropy loss and the operating parameter perceived feature consistency loss are added together according to their weights to obtain the composite loss function for model training.

[0011] Secondly, this application also provides an early fault diagnosis device for equipment health management. The device includes: a data acquisition module for acquiring real-time collected equipment operation monitoring data; the equipment operation monitoring data includes: the original time-domain vibration signal corresponding to the equipment vibration data and the equipment operation parameter vector; a model input module for inputting the equipment operation monitoring data into a pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operation parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer; a model prediction module for performing multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing on the equipment operation monitoring data through the multi-source feature fusion layer to determine the fused feature vector; through the multi-head adaptive gating attention layer, dynamically adjusting the activation intensity of each feature dimension in the fused feature vector according to the equipment operation parameter vector to focus on the feature subspace most relevant to the fault and outputting an attention-weighted feature vector; through the multi-scale residual classification layer, extracting multi-scale deep features from the attention-weighted feature vector and weighting and aggregating them to output a fault category probability vector, and determining the early fault diagnosis result of the equipment health status based on the fault category probability vector.

[0012] Thirdly, this application also provides an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the method described in the first aspect above.

[0013] Fourthly, this application also provides a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method described in the first aspect above.

[0014] The early fault diagnosis method, apparatus, and electronic equipment for equipment health management provided in this application have the following advantages compared to the prior art: 1. A multi-band adaptive spectrum reconstruction method based on energy distribution gradient is proposed. By accumulating energy gradient, frequency bands are automatically divided and sub-bands are independently enhanced and normalized. This effectively suppresses strong background noise and main rotating frequency components, and significantly highlights the weak and dispersed sideband characteristics in early faults, thereby improving the separability of fault characteristics from the source.

[0015] 2. Construct a joint recursive matrix feature representation method that integrates time-domain envelope and frequency-domain reconstructed spectrum, and map the time-domain nonlinear dynamic characteristics and frequency-domain energy distribution pattern of vibration signal into a joint phase space. This method can characterize the subtle but discriminative time-frequency coupling structural features in the early fault evolution process.

[0016] 3. By adopting a multi-source feature fusion and multi-head gated attention deep network structure with operating parameter modulation, operating condition parameters such as speed and load are explicitly introduced into the feature modulation and attention weight generation process, which realizes adaptive modeling of the differences in fault feature contribution under different operating conditions, and enhances the robustness and generalization ability of the model under varying operating conditions.

[0017] 4. A composite optimization objective combining adaptive focus cross-entropy loss and operational parameter-aware feature consistency loss is proposed. This objective alleviates the problems of scarce early fault samples and class imbalance, while guiding the model to learn more stable and consistent feature representations under similar operating conditions, thereby improving the accuracy and reliability of early fault identification. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an early fault diagnosis method for equipment health management, provided as an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an early failure model of a device provided in an embodiment of this application; Figure 3 A structural block diagram of an early fault diagnosis device for equipment health management provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Existing fault diagnosis methods for equipment health management often have the following problems: 1. Most existing methods directly extract features from the original vibration signal or global spectrum, relying on fixed frequency band division or overall normalization processing, which makes it difficult to effectively separate and enhance the weak characteristic frequency bands corresponding to early faults in the context of strong noise and strong dominant frequency.

[0022] 2. Existing time-domain analysis methods and frequency-domain analysis methods are often independent of each other, lacking joint modeling of the intrinsic relationship between time-domain dynamic changes and frequency-domain energy distribution, resulting in the failure to fully explore the key time-frequency coupling information in the early fault evolution process.

[0023] 3. Existing deep learning-based fault diagnosis models typically treat operating parameters as additional inputs or completely ignore their influence, failing to characterize the modulation effect of operating condition changes on vibration feature distribution and fault representation. The models are prone to performance degradation under varying operating conditions.

[0024] 4. Existing loss functions are mostly based on standard cross-entropy, which is difficult to cope with the problem of a small number of early fault samples and small differences between categories. At the same time, they lack constraints on the consistency of features of samples of the same type under similar working conditions, resulting in insufficient stability of the model in the identification of minority classes and complex working conditions.

[0025] Based on this, embodiments of this application provide an early fault diagnosis method, apparatus, and electronic device for equipment health management. To facilitate understanding of this embodiment, the early fault diagnosis method for equipment health management disclosed in this application will be described in detail first. Figure 1 A flowchart of an early fault diagnosis method for equipment health management provided in this application embodiment is included, the method specifically including the following steps: Step S102: Acquire real-time collected equipment operation monitoring data; the equipment operation monitoring data includes: the original time-domain vibration signal corresponding to the equipment vibration data and the equipment operation parameter vector; In practice, high-sensitivity vibration acceleration sensors installed on the bearing housings or casings of critical rotating equipment (such as motors, gearboxes, and fans) continuously collect raw time-domain vibration signals of the equipment at different operating stages and health states at a fixed sampling frequency (e.g., 12.8kHz or 25.6kHz). Simultaneously, operating parameters closely related to the equipment's operating status, such as rotational speed and load, are collected synchronously through the equipment control system or additional sensors to construct an operating parameter vector.

[0026] Step S104: Input the equipment operation monitoring data into the pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operating parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer, such as... Figure 2 As shown; Early-stage faults exhibit weak vibration characteristics that are closely coupled with equipment operating conditions. Conventional deep learning models directly classify feature vectors, ignoring the modulating effect of operating parameters on fault characteristics and the differential contributions of multi-scale fault characteristics.

[0027] This embodiment constructs a deep network that integrates operational parameter modulation and multi-scale residual learning. The network takes spectral reconstruction features and time-frequency joint features as inputs, combines operational parameters, and extracts multi-scale deep features sensitive to early faults through adaptive fusion and attention mechanisms, and then classifies them.

[0028] Step S106: Through the multi-source feature fusion layer, multi-band adaptive spectrum reconstruction, joint recursive matrix construction and feature fusion processing are performed on the equipment operation monitoring data to determine the fused feature vector; Step S108: Through a multi-head adaptive gating attention layer, the activation intensity of each feature dimension in the fused feature vector is dynamically adjusted according to the device operating parameter vector to focus on the feature subspace most relevant to the fault and output the attention-weighted feature vector. Step S110: Through the multi-scale residual classification layer, the attention-weighted feature vector is processed to extract multi-scale deep features and aggregate them with weights to output a fault category probability vector. Based on the fault category probability vector, the early fault diagnosis result of the equipment health status is determined.

[0029] The process of feature extraction and prediction in steps S106-110 will be described in detail later.

[0030] Step S106 above, which involves performing multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing on the equipment operation monitoring data through a multi-source feature fusion layer to determine the fused feature vector, includes: The following steps are performed through the multi-source feature fusion layer: (1) Perform multi-band adaptive spectrum reconstruction on the original time-domain vibration signal corresponding to the equipment vibration data to determine the reconstructed spectrum vector; Equipment vibration data is often affected by strong high-frequency noise interference, and the fault characteristic frequency band is dispersed and the energy is weak. Conventional methods directly perform fast Fourier transform on the original time-domain vibration signal to obtain the global spectrum. However, the strong background noise and main rotation frequency components in this spectrum often overwhelm the weak sideband characteristics that reflect early faults, such as slight bearing wear or shaft misalignment.

[0031] This embodiment employs an adaptive frequency band division and reconstruction strategy based on energy distribution. First, the original vibration signal is converted to the frequency domain. Then, based on the abrupt change points of the spectral energy gradient, the entire frequency band is divided into multiple sub-bands. The spectral amplitude of each sub-band is independently normalized and enhanced, reconstructing a new spectral map that can highlight subtle fault characteristics. The specific steps are as follows: (1.1) Perform a fast Fourier transform on the original time-domain vibration signal corresponding to the equipment vibration data, and take the amplitude of the single-sided spectrum to obtain the corresponding spectrum amplitude vector; Specifically, for the i-th sample, the original time-domain vibration signal of the i-th sample is defined as follows: Data collected by a vibration sensor, dimension: Given a vibration sequence of a fixed time length, perform a Fast Fourier Transform on the signal and take the amplitude of one-sided spectrum to obtain the one-sided spectrum amplitude vector of the i-th sample. , dimension Characterizes the energy distribution of a signal in the frequency domain, where ;in, This represents the sample index, with a value range of [value range missing]. , The total number of samples; This represents the length of the original time-domain signal, with an example value of 8192. Represents the length of a single-sided spectrum, when hour, The example value is 4097.

[0032] This step involves real-time data prediction using a model. Therefore, the i-th sample refers to the current real-time data sample. For the model training process, which involves multiple training samples, the above processing needs to be performed on each sample. Subsequent steps follow the same logic and will not be elaborated further.

[0033] (1.2) Based on the spectral amplitude vector, calculate the cumulative energy and gradient of the spectral amplitude. By locating the local maximum point of the cumulative energy gradient, adaptively determine the boundary point of the frequency band division so as to separate the scattered and weak fault features into independent sub-bands. Specifically, the calculation starts from the first frequency index and goes to the second. Cumulative energy of each frequency index And calculate its normalized gradient. ; Then, search All local maxima points, which constitute the set of frequency band division boundary points, are represented as:

[0034] in, This represents the set of indices for all frequency band boundary points; Indicates from the first frequency index to the second. The cumulative energy of each frequency index represents the total spectral energy up to a certain frequency. Represents the one-sided spectral magnitude vector of the i-th sample. In the The amplitude at each frequency index represents the vibration intensity of that frequency component; This represents the frequency index, with a value range of [value range missing]. ; Indicates difference from Frequency index; Indicates from the first frequency index to the second. The cumulative energy of each frequency index; Indicates the first The normalized cumulative energy gradient at each frequency index characterizes the degree of energy change with frequency, and its peak point usually corresponds to the boundary of different energy distribution intervals. Indicates from the first frequency index to the second. The cumulative energy of each frequency index, that is, the total energy of the entire spectrum, and This is used to normalize the gradient; Indicates the first A partition boundary point index, satisfying that its corresponding cumulative energy gradient is a local maximum; Indicates the first Normalized cumulative energy gradient at each boundary point; Indicates the first Normalized cumulative energy gradient at each frequency index; Indicates the first Normalized cumulative energy gradient at each frequency index; This represents the logical AND operator; This represents the total number of frequency bands, and its value is equal to the number of boundary points plus 1, automatically determined by the algorithm. In specific implementations, the following conventions apply. As the starting index, To terminate the index, The maximum value is constrained by the preset minimum bandwidth.

[0035] (1.3) For each sub-band, the enhancement coefficient is adaptively calculated based on the relative relationship between the sub-band average energy and the global average energy, and the hyperbolic tangent function is used for nonlinear enhancement and normalization to obtain the amplitude of the reconstructed spectrum at each frequency index for each sub-band. Each sub-band is processed independently. Based on the relative relationship between the sub-band average energy and the global average energy, the enhancement coefficient is adaptively calculated, and a hyperbolic tangent function is used for nonlinear enhancement and normalization to obtain the amplitude of the reconstructed spectrum at each frequency index. This relatively enhances the amplitude of the low-energy frequency band and suppresses high-energy background components, as expressed below:

[0036] In the formula, Indicates the reconstructed spectrum at the th The amplitude at each frequency index represents the relative enhanced energy intensity of each frequency component in the reconstructed spectrum, which is used to more effectively detect early faults. This represents the hyperbolic tangent activation function, used to compress the dynamic range of data and introduce a nonlinear mapping; Indicates the first The adaptive enhancement coefficient for each frequency band is inversely proportional to the average energy of that sub-band; the lower the energy of the sub-band, the stronger the enhancement effect. The calculation method is expressed as follows: ; Indicates the first The minimum value of the original spectral amplitude within each frequency band, i.e. This is used to perform zero-bias processing on the amplitude of the sub-band, so as to make more effective use of the sensitive interval of the hyperbolic tangent function; Indicates finding the first... The maximum value of the expression within the parentheses in each frequency band is used to achieve amplitude normalization within the sub-band. Indicates the first Index of boundary points. The average energy of the entire spectrum is expressed as follows: This characterizes the overall energy level of the signal; Indicates the first The average energy of each frequency band is calculated as follows: This characterizes the energy level of the subband; This represents a very small positive integer, used to prevent the denominator from being zero and to ensure numerical stability. Examples of its values ​​are shown below. .

[0037] (1.4) Based on the amplitude of the reconstructed spectrum at each frequency index corresponding to each sub-band, multiple reconstructed spectra are spliced ​​together in the order of the original frequencies to form a complete reconstructed spectrum vector with enhanced features.

[0038] All the reconstructed spectral segments of the sub-bands are concatenated in their original frequency order to form a complete, feature-enhanced reconstructed spectral vector, represented as:

[0039] In the formula, Indicates the first Reconstruct the spectrum vector from each sample, with dimension [ ]. This vector effectively suppresses strong background components and relatively amplifies weak fault features; This indicates a vector concatenation operation.

[0040] (2) By using the joint recursive matrix construction method, the reconstructed spectrum vector is combined with the envelope information of the original time-domain vibration signal to generate a time-frequency joint feature vector that can simultaneously characterize the time-domain dynamics and frequency-domain energy distribution pattern; The nonlinear dynamic characteristics of fault evolution in the time domain are analyzed by conventional techniques, which analyze the time domain or frequency domain characteristics separately, ignoring the joint correlation mode between the two. However, the subtle vibration characteristics of early faults are often reflected in the mode of energy change over time in a specific frequency band.

[0041] This embodiment employs a joint recursive matrix construction method, combining the enhanced reconstructed spectrum with the envelope information of the original vibration signal to generate a time-frequency joint feature vector that can simultaneously characterize the time-domain dynamics and frequency-domain energy distribution patterns. The specific steps are as follows: (2.1) Extract the envelope signal from the original time-domain vibration signal; The envelope signal of the original vibration time-domain signal is extracted to capture the macroscopic profile of the signal amplitude change over time, reflecting the time-domain fluctuation characteristics of the vibration energy. Specifically, the Hilbert transform is used to obtain the modulus of the analytic signal, thus obtaining the envelope signal of the i-th sample. , dimension Characterizes the time-varying trend of the instantaneous amplitude of the signal; among which, Represents the Hilbert transform operator; The modulus (amplitude) operator for complex numbers.

[0042] (2.2) Divide the envelope signal and the reconstructed spectrum vector into several non-overlapping segments, and calculate the root mean square value and the spectral centroid of each segment to obtain the envelope root mean square sequence and the spectral centroid sequence; The envelope signal and the reconstructed spectral vector are divided into several non-overlapping segments, and representative statistics (root mean square value and spectral centroid) of each segment are calculated, thereby converting the high-dimensional signal into a low-dimensional time-domain and frequency-domain feature sequence, as follows:

[0043]

[0044] In the formula, Represents the envelope signal of the i-th sample. The root mean square value of each segment represents the average level of vibrational energy within that time period; Indicates the first The reconstructed spectrum vector of the nth sample The spectral centroid of a segment represents the frequency at which energy is concentrated within that frequency band; Indicates the first The envelope signal of each sample in time index The amplitude at that point represents the instantaneous amplitude of the vibration signal at that moment; Indicates the first The reconstructed spectrum vector of the nth sample is at the th... The amplitude at each frequency index represents the energy intensity of that frequency component in the enhanced spectrum; This represents the length of each segment. For a time-domain envelope signal, the length of each segment is... (Assuming divisibility), for the frequency domain reconstructed spectrum, the length of each segment is... (Assuming divisibility); This represents the total number of segments, which is also the length of the final feature sequence; an example value is 64. Represents the time index of a time-domain signal; Indicates the first The index range contained in each fragment.

[0045] Furthermore, by combining the representative statistics of all segments, we obtain two segments of length [missing information]. Feature sequence: Root mean square envelope sequence and spectral centroid sequence .

[0046] (2.3) Construct a joint recursive matrix based on the root mean square sequence and the spectral centroid sequence; A joint recursive matrix is ​​constructed based on the root mean square sequence and the spectral barycenter sequence. This matrix simultaneously considers the recursive characteristics of the time-domain envelope and the frequency-domain barycenter in the two-dimensional phase space, and introduces a time decay weight to emphasize short-term dynamic correlation, expressed as:

[0047] In the formula, Represents the joint recursive matrix The Line number The elements of the column, with values ​​between 0 and 1, are used to quantify the similarity of the system states at two different times in the time-frequency joint phase space; denoted as the Herveside step function, the function value is 1 when the independent variable is greater than or equal to 0, and 0 otherwise; This represents the recursion threshold, used to determine whether two state points are close enough to cause "recursion". It can be a proportion of the standard deviation of the Euclidean distance between all point pairs in the joint phase space, with an example value of 0.2 times the standard deviation. Represents the L2 norm; Represents the natural exponential function; Represents the root mean square sequence of the envelope The Middle The value at each time point; Represents the spectral centroid sequence The Middle The value at each time point; Indicates by and The two-dimensional feature points represent the system at time t. The state; This represents the row index of the joint recursive matrix, i.e., the index of the first time point, with a value range of [value missing]. arrive ; This represents the column index of the joint recursive matrix, which is also the index of the second time point, and its value range is... arrive ; The scale parameter that controls the rate of time decay. The larger the value, the slower the weight decays over time; the example value is 5.0.

[0048] (2.4) Reshape the joint recursive matrix into a one-dimensional vector as the time-frequency joint eigenvector.

[0049] The joint recursive matrix is ​​reshaped into a one-dimensional vector, which serves as the joint time-frequency eigenvector. The first... Time-frequency joint feature vector of each sample Its dimensions are This vector simultaneously encodes the dynamic evolution pattern and recursive characteristics of the vibration signal in both the time-domain envelope and the frequency-domain centroid; among which, This represents the operation of expanding a matrix row by row into a one-dimensional vector.

[0050] (3) The spectrum reconstruction vector, the time-frequency joint feature vector and the equipment operation parameter vector are fused together to determine the fused feature vector.

[0051] The spectral reconstruction vector, the time-frequency joint feature vector, and the externally acquired equipment operating parameter vector are fused. The operating parameters are used to generate adaptive weights to modulate the contributions of different feature sources, thereby enhancing the model's robustness to changes in operating conditions. The fused feature vector is represented as follows:

[0052] In the formula, Indicates the first The fused feature vector of each sample, with dimension [ ]. This characterizes the joint features after modulation by the operating parameters; Indicates the first The vector of running parameters for each sample, with dimension 1. It characterizes the working conditions of the equipment and provides working condition information to the network to assist feature modulation. It is collected by external sensors or working condition records. In one implementation, it specifically includes two parameters, namely rotational speed (unit: revolutions per minute) and load (unit: Newton-meter), and normalization processing is required before input. This represents the weight matrix used for projected spectral features, with dimensions of [missing information]. , used to map spectral features from the original dimension to the hidden dimension, are trainable parameters; This represents the bias vector used for projected spectral features, with dimension . , used to adjust the bias of the projected spectral features, is a trainable parameter; This represents the weight matrix used for projecting the joint time-frequency features, with dimension 1. , is used to map joint time-frequency features from the original dimension to the hidden dimension, and is a trainable parameter; This represents the bias vector used for projecting the joint time-frequency features, with dimension . , used to adjust the bias of the projected time-frequency joint features, is a trainable parameter; This represents the weight matrix used to generate the spectral feature modulation weights, with dimensions of . , used to calculate the modulation coefficients for spectral features based on operating parameters, are trainable parameters; This represents the bias vector used to generate the modulation weights for spectral features, with dimension 1. , used to adjust the bias of the modulation weight calculation, is a trainable parameter; This represents the weight matrix used to generate the time-frequency feature modulation weights, with dimensions of . , used to calculate the modulation coefficients for the joint time-frequency features based on the operating parameters, are trainable parameters; This represents the bias vector used to generate the time-frequency feature modulation weights, with dimension 1. , used to adjust the bias of the modulation weight calculation, is a trainable parameter; This represents the Sigmoid activation function, used to compress the output of a linear transform to... Intervals are used to generate modulation weights; This represents the element-wise multiplication operator; This represents the activation function of the rectified linear unit, used to introduce nonlinearity and filter negative values; The representation layer normalization operation is used to standardize the fused features to stabilize the training process; The hidden dimension represents the fused features, defines the spatial size of the intermediate features in the network, and is a hyperparameter with an example value of 256.

[0053] The above step S108, through a multi-head adaptive gating attention layer, dynamically adjusts the activation intensity of each feature dimension in the fused feature vector according to the device operating parameter vector, in order to focus on the feature subspace most relevant to the fault and output an attention-weighted feature vector. The specific implementation process is as follows: The fused feature vector may contain redundant information. A multi-head gating attention mechanism is introduced to dynamically adjust the activation intensity of each feature dimension based on operating parameters, focusing on the feature subspace most relevant to the fault, and outputting an attention-weighted feature vector, represented as:

[0054] In the formula, Indicates the first The attention-weighted feature vector of each sample, with dimension [ ]. This represents the deep features after attention filtering, which are used to provide fault-sensitive feature representations for subsequent classification layers. This represents the total number of attention heads, used to segment features into multiple subspaces for parallel processing. It is a hyperparameter with an example value of 8. Indicates the first The fusion feature vector of each sample The There are segments, and the dimension of each segment is . , characterization assigned to the first A subset of features for each attention head; Indicates the first One attention head is used to generate the weight matrix of the gating weights, with dimensions of . , is used to map the intermediate representation after modulation of the running parameters to gated weights, and is a trainable parameter; Indicates the first One attention head is used to process the weight matrix of the running parameters, with dimensions of . , used to perform linear transformations on the running parameters, are trainable parameters; Indicates the first An intermediate bias vector is used when the attention head processes runtime parameters, with dimension [missing information]. , used to adjust the bias of the transformation results of the running parameters, are trainable parameters; Indicates the first The output bias vector of the attention head when generating the gating weights has a dimension of . , used to adjust the bias of the final gating weights, are trainable parameters; This represents the Swish activation function, used to enhance non-linear expressive power; This indicates a vector concatenation operation, used to merge the outputs of multiple attention heads back into a complete feature vector; This represents the intermediate dimension in the attention mechanism, used to define the size of the intermediate features in the transformation of the runtime parameters. It is a hyperparameter, with an example value of 64.

[0055] It should be noted that the multi-head adaptive gating attention layer captures the correlation of features in different subspaces in parallel through multi-head design, and uses the running parameters to generate adaptive gating weights to suppress irrelevant features.

[0056] The specific implementation process of step S110 above, which involves extracting multi-scale deep features from the attention-weighted feature vector through a multi-scale residual classification layer, weighting and aggregating the features, outputting a fault category probability vector, and determining the early fault diagnosis result of the equipment health status based on the fault category probability vector, is as follows: Early fault features exhibit differentiated characteristics at different scales. By extracting multi-scale deep features through parallel residual blocks and weighted aggregation, the model's ability to identify weak multi-scale fault patterns is enhanced, outputting a fault category probability vector, represented as:

[0057] In the formula, Indicates the first The fault category probability vector of each sample has dimension . The confidence level of the model's prediction for each fault category is used to output the final classification result, and is obtained through calculation. This represents the total number of multi-scale residual blocks, used to define the number of scales for feature extraction. It is a hyperparameter with an example value of 3. Indicates the first The weighted scalar of each residual block, used to balance the contributions of features at different scales, is a trainable parameter. Indicates the first The weight matrix of the first layer in each residual block has dimensions of . , used to map input features to the hidden space of residual blocks, are trainable parameters; Indicates the first The bias vector of the first layer in each residual block has a dimension of . , used to adjust the bias of the first layer transformation result, is a trainable parameter; This represents the Gaussian error linear unit activation function, used to enhance the nonlinear fitting ability of the model; Indicates the first The weight matrix of the second layer in each residual block has dimensions of . , used to map hidden features back to the original dimension, are trainable parameters; Indicates the first The bias vector of the second layer in each residual block has a dimension of , used to adjust the bias of the second-layer transformation result, is a trainable parameter; This represents the weight matrix of the output classification layer, with dimension 1. , used to map multi-scale aggregated features to the class space, are trainable parameters; This represents the bias vector of the output classification layer, with dimension . , used to adjust the bias of the classification results, is a trainable parameter; This represents the Softmax function, used to convert a linear output into a probability distribution; This indicates the total number of fault categories, including health status. It is used to define the number of categories for the classification task and is determined by the specific diagnostic task. Indicates the first The hidden dimension of each residual block defines the spatial size of the features within that residual block. It is a hyperparameter, and different values ​​can be set for different blocks to capture multi-scale features. Examples of possible values ​​are shown below. .

[0058] In practice, each residual block learns the residual of the input features through two layers of transformation, and different blocks are configured with different hidden dimensions. To focus on different scales, and ultimately through learnable weighted scalars Integration.

[0059] The training process of the early fault diagnosis model for the above-mentioned equipment is as follows: (1) Obtain the training set of equipment status monitoring data; (1.1) Acquire historical equipment operation monitoring data; historical equipment operation monitoring data includes: continuously collecting vibration signal samples of the equipment at different operating stages and different health states at a fixed sampling frequency, and synchronously collecting the equipment operation parameter vectors corresponding to the vibration signal samples; Comprehensive and representative equipment condition monitoring data are acquired to construct a training dataset for training the deep network. Specifically, high-sensitivity vibration acceleration sensors installed on the bearing housings or casings of key rotating equipment (such as motors, gearboxes, and fans) continuously collect raw time-domain vibration signals of the equipment at different operating stages and health states at a fixed sampling frequency (e.g., 12.8kHz or 25.6kHz). Simultaneously, operating parameters closely related to the equipment's operating status, such as rotational speed and load, are synchronously collected through the equipment control system or additional sensors to construct an operating parameter vector.

[0060] (1.2) Label the health status of each vibration signal sample and the corresponding equipment operation parameter vector in the equipment operation monitoring data; To construct the training set required for supervised learning, each collected vibration signal sample and its corresponding operating parameters need to be precisely labeled. The labeling categories are determined based on the actual health status of the equipment. For example, this may include a "healthy" category and several "early failure" categories. Early failure categories can be defined according to the specific diagnostic object. For example, for rolling bearings, early failure categories may include "early minor wear of the inner ring," "early minor wear of the outer ring," and "early minor damage to the rolling elements"; for rotating shaft systems, they may include "slight imbalance" and "initial misalignment." The label for each sample data segment is determined by the corresponding equipment disassembly and inspection results, simulated fault implantation experiment records, or expert evaluation based on long-term monitoring trends, ensuring the authenticity and accuracy of the labels.

[0061] (1.3) Integrate the labeled vibration signal sequence, the corresponding equipment operating parameter vector and its fault category label to construct a structured equipment condition monitoring data training set.

[0062] (2) The multi-source feature fusion layer, multi-head adaptive gating attention layer and multi-scale residual classification layer of the operating parameters are trained by the equipment condition monitoring data training set to obtain the early fault diagnosis model of the equipment. During the training process, the loss value is calculated by a composite loss function that combines adaptive focus cross-entropy loss and operating parameter perception feature consistency loss.

[0063] After completing the construction of the training dataset, feature extraction, and network structure definition (see the feature extraction and prediction structure section in the aforementioned real-time prediction process), the iterative training and parameter update phase of the deep network begins.

[0064] First, all trainable parameters in the network are initialized, such as the weight matrix and bias vector, which are typically initialized using a random initialization strategy. In each training iteration, a batch of samples is randomly selected from the training dataset and sequentially processed through the forward propagation process described in steps S106-S110: first, multi-band adaptive spectrum reconstruction is performed on the vibration signal of each sample in the batch; then, its time-frequency joint features are extracted; and then the spectrum reconstruction vector, the time-frequency joint feature vector, and the running parameter vector are input together into a deep network that fuses running parameter modulation and multi-scale residual learning. The network outputs a fault category probability prediction vector for each sample.

[0065] Then, based on the total loss function defined above, the total loss value for this batch of samples is calculated. The total loss value measures the difference between the predicted result and the true label under the current network parameters, and is combined with the feature consistency constraint perceived by the running parameters. The core of training lies in calculating the gradient of the loss with respect to each trainable parameter in the network using the backpropagation algorithm based on this loss value.

[0066] Then, an adaptive moment estimation optimization algorithm is used to update all trainable parameters based on the calculated gradient information, aiming to optimize the total loss value in the direction of reduction, thereby continuously improving the network's classification accuracy for early faults and its robustness to changes in operating conditions.

[0067] The aforementioned forward propagation, loss calculation, backpropagation, and parameter update processes are repeated cyclically, iterating through the samples in the training dataset multiple times. Iterative training continues until a preset stopping condition is met. The stopping condition typically includes reaching the maximum number of training iterations, or the model's performance (e.g., classification accuracy) on a separately reserved validation dataset no longer improving or even starting to decline, indicating that the model may have been sufficiently trained or is at risk of overfitting. At this point, training is terminated, and the network parameters with the best validation performance are saved, resulting in a well-trained fault diagnosis model.

[0068] Early failure samples are scarce and the differences between classes are small. Conventional cross-entropy loss can easily lead to the model overfitting the majority class and underfitting the minority class, while ignoring the influence of running parameters on feature distribution.

[0069] This embodiment employs a composite loss function combining adaptive focus cross-entropy loss and runtime parameter-aware feature consistency loss to enhance learning on difficult samples and improve the model's generalization ability under varying conditions. The specific steps are as follows: (1) Determine the adaptive focus parameters of the samples based on the device operation parameter vector; determine the adaptive focus cross-entropy loss based on the adaptive focus parameters of the samples and the predicted probability of the samples to the true class; To address the difficulty in classifying early fault samples, a focus loss concept is introduced, and the focus parameters are adaptively adjusted based on the operating parameters. This makes the model pay more attention to difficult samples under the current operating conditions, as expressed below:

[0070] In the formula, This represents the adaptive focus cross-entropy loss. The smaller the value, the higher the model's classification accuracy, providing a supervisory signal for the model's classification accuracy. This represents a set of sample indices in a batch, used to specify the range of samples currently participating in the loss calculation; This indicates the batch size, which is the number of samples contained in a batch. An example value is 32. Indicates the first Each sample corresponds to its true category The predicted probability is the first Fault category probability vector of each sample The Each element represents the model's confidence in the true class of the sample; Indicates the first The true fault category label for each sample, with a value of An integer in the form of a classification supervision standard answer; Indicates the first The adaptive focus parameter for each sample dynamically adjusts its weight in the loss based on the running parameters. The calculation method is expressed as follows: ; This indicates the maximum value of the focus parameter, used to limit the upper limit of the focus parameter. An example value is 5.0. This represents the weight vector used to generate the focus parameters, with dimensions of... , used to calculate focus parameters based on running parameters, are trainable parameters; This represents a bias scalar used to generate the focus parameters, which is used to adjust the bias of the focus parameter calculation and is a trainable parameter.

[0071] It should be noted that the focus loss function By running parameter vector Dynamically calculate the adaptive focus parameters of the sample When the sample classification is difficult, This will increase the weight of the loss, forcing the model to learn more effectively.

[0072] (2) Determine the consistency loss of operating parameter perception features based on the device operating parameter vector, the bandwidth parameter of operating parameter similarity, and the indicator function based on the real fault category label; To enhance the stability of the model's feature representation under varying operating conditions, a consistency loss is introduced to encourage similar samples to have similar feature representations under similar operating parameters, expressed as:

[0073] In the formula, This represents the consistency loss of perceived features in operating parameters. The smaller the value, the more consistent the features of similar samples under similar operating conditions. The bandwidth parameter represents the similarity of operating parameters and is used to control the decay rate of the similarity weight as the parameter differences increase. An example value is 1.0. Indicates an indicator function, when The function value is 1 when the function is active and 0 otherwise, and is used to filter pairs of similar samples. Indicates the first The true fault category label for each sample; Indicates the first Attention-weighted feature vectors for each sample.

[0074] It should be noted that, The item assigns greater weight to sample pairs with similar operating parameters, and combines this with the category label. It specifically constrains the feature distance of similar samples under similar operating conditions.

[0075] (3) The adaptive focus cross-entropy loss and the running parameter perception feature consistency loss are added together according to their weights to obtain the composite loss function for model training.

[0076] The adaptive focus cross-entropy loss and the runtime parameter-aware feature consistency loss are added together with weights to obtain the total loss function for model training, which is expressed as:

[0077] In the formula, This represents the total loss function, which is the final objective that needs to be minimized during model training and is used to provide overall supervision signals for model optimization. This represents the weighting coefficient that balances the two types of losses, controls the strength of feature consistency regularization, and is a hyperparameter with an example value of 0.1.

[0078] Once the deep network has completed training and obtained optimized parameters, it can be deployed in real-world industrial scenarios to achieve online or offline early fault diagnosis for equipment health management, as described in steps S102-110 of this application.

[0079] In the application phase, for target equipment requiring health status assessment, the system collects vibration monitoring signals and corresponding operating parameters under its current operating conditions in real time or periodically. The collected raw vibration signals first undergo a preprocessing and feature extraction process identical to that in the training phase: namely, performing multi-band adaptive spectrum reconstruction to convert the noise-affected raw signals into a reconstructed spectrum that enhances weak fault features; joint recursive matrix construction and time-frequency joint feature extraction to obtain feature vectors encoding time-frequency dynamic patterns. Then, the reconstructed spectrum vector, time-frequency joint feature vector, and real-time operating parameter vector of the current sample are input together into a pre-trained deep network that integrates operating parameter modulation and multi-scale residual learning. The network's forward propagation process automatically fuses the operating parameter modulation features and, after multi-head attention filtering and multi-scale residual learning, outputs a fault category probability vector. This probability vector indicates the likelihood that the current equipment state belongs to various preset health categories (including healthy states and various early fault types).

[0080] The diagnostic system selects the category with the highest probability value as the final diagnostic result output. Furthermore, the system can set a confidence threshold; when the highest probability falls below the threshold, a "status uncertain" warning is issued, indicating the need for further investigation. By continuously monitoring and diagnosing the equipment in this way, the system can identify fault types in their early stages, before serious performance degradation or secondary damage occurs. This provides crucial decision support for predictive maintenance, enabling a shift from periodic or reactive maintenance to precise and economical condition-based maintenance, effectively improving equipment reliability, safety, and operational economics.

[0081] Based on the above method embodiments, this application also provides an early fault diagnosis device for equipment health management, see [link to relevant documentation]. Figure 3 As shown, the device includes: a data acquisition module 32, used to acquire real-time collected equipment operation monitoring data; the equipment operation monitoring data includes: the original time-domain vibration signal corresponding to the equipment vibration data and the equipment operation parameter vector; a model input module 34, used to input the equipment operation monitoring data into a pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operation parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer; a model prediction module 36, used to determine the fused feature vector by performing multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing on the equipment operation monitoring data through the multi-source feature fusion layer; by dynamically adjusting the activation intensity of each feature dimension in the fused feature vector according to the equipment operation parameter vector through the multi-head adaptive gating attention layer, focusing on the feature subspace most relevant to the fault, and outputting an attention-weighted feature vector; by extracting multi-scale deep features from the attention-weighted feature vector through the multi-scale residual classification layer and weighting and aggregating them, outputting a fault category probability vector, and determining the early fault diagnosis result of the equipment health status according to the fault category probability vector.

[0082] Furthermore, the aforementioned model prediction module 36 is used to perform the following steps through the multi-source feature fusion layer: perform multi-band adaptive spectrum reconstruction on the original time-domain vibration signal corresponding to the equipment vibration data to determine the reconstructed spectrum vector; use a joint recursive matrix construction method to combine the reconstructed spectrum vector with the envelope information of the original time-domain vibration signal to generate a time-frequency joint feature vector that can simultaneously characterize the time-domain dynamics and frequency-domain energy distribution pattern; and fuse the spectrum reconstruction vector, the time-frequency joint feature vector, and the equipment operating parameter vector to determine the fused feature vector.

[0083] Furthermore, the aforementioned model prediction module 36 is used to perform a fast Fourier transform on the original time-domain vibration signal corresponding to the equipment vibration data, and take the amplitude of the single-sided spectrum to obtain the corresponding spectrum amplitude vector; based on the spectrum amplitude vector, the cumulative energy and gradient of the spectrum amplitude are calculated, and the boundary points of the frequency band division are adaptively determined by locating the local maxima of the cumulative energy gradient, so as to separate the scattered and weak fault features into independent sub-bands; for each sub-band, the enhancement coefficient is adaptively calculated according to the relative relationship between the sub-band average energy and the global average energy, and the hyperbolic tangent function is used for nonlinear enhancement and normalization processing to obtain the amplitude of the reconstructed spectrum at each frequency index corresponding to each sub-band; based on the amplitude of the reconstructed spectrum at each frequency index corresponding to all sub-bands, the multiple reconstructed spectra are spliced ​​together in the original frequency order to form a complete, feature-enhanced reconstructed spectrum vector.

[0084] Furthermore, the aforementioned model prediction module 36 is used to extract the envelope signal from the original time-domain vibration time-domain signal; divide the envelope signal and the reconstructed spectrum vector into several non-overlapping segments, and calculate the root mean square value and spectral centroid of each segment to obtain the envelope root mean square sequence and the spectral centroid sequence; construct a joint recursive matrix based on the root mean square sequence and the spectral centroid sequence; and reshape the joint recursive matrix into a one-dimensional vector as the time-frequency joint feature vector.

[0085] Furthermore, the aforementioned device also includes: a model training module, used to perform the following training process for an early equipment fault diagnosis model: acquiring a training set of equipment condition monitoring data; training a multi-source feature fusion layer, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer modulated by operating parameters using the equipment condition monitoring data training set to obtain an early equipment fault diagnosis model; during the training process, using a composite loss function combining adaptive focus cross-entropy loss and operating parameter perception feature consistency loss to calculate the loss value.

[0086] Furthermore, the aforementioned model training module is used to acquire historical equipment operation monitoring data. This historical equipment operation monitoring data includes: continuously collecting vibration signal samples of the equipment at different operating stages and health states at a fixed sampling frequency, and simultaneously collecting equipment operation parameter vectors corresponding to the vibration signal samples; labeling each vibration signal sample and its corresponding equipment operation parameter vector in the equipment operation monitoring data with a health status label; and integrating the labeled vibration signal sequence, the corresponding equipment operation parameter vector, and its fault category label to construct a structured equipment condition monitoring data training set.

[0087] Furthermore, the composite loss function is determined as follows: Based on the equipment operating parameter vector, the sample adaptive focus parameter is determined; based on the sample adaptive focus parameter and the predicted probability of the sample to the true category, the adaptive focus cross-entropy loss is determined; based on the equipment operating parameter vector, the bandwidth parameter of operating parameter similarity, and the indicator function based on the true fault category label, the operating parameter perceived feature consistency loss is determined; the adaptive focus cross-entropy loss and the operating parameter perceived feature consistency loss are added together according to their weights to obtain the composite loss function for model training.

[0088] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts of the device embodiment not mentioned can be referred to the corresponding content in the aforementioned method embodiment.

[0089] This application also provides an electronic device, such as... Figure 4 The diagram shows the structure of the electronic device, which includes a processor 41 and a memory 40. The memory 40 stores computer-executable instructions that can be executed by the processor 41, and the processor 41 executes the computer-executable instructions to implement the above-described method.

[0090] exist Figure 4 In the illustrated embodiment, the electronic device further includes a bus 42 and a communication interface 43, wherein the processor 41, the communication interface 43, and the memory 40 are connected via the bus 42.

[0091] The memory 40 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 43 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 42 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 42 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0092] Processor 41 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 41 or by instructions in software form. Processor 41 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can 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 can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory, and the processor 41 reads the information in the memory and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiment.

[0093] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to implement the above-described method. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0094] The computer program products of the methods, apparatus, and electronic devices provided in the embodiments of this application include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementations, please refer to the method embodiments, which will not be repeated here.

[0095] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0096] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a 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 several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0098] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An early fault diagnosis method for equipment health management, characterized in that, The method includes: Acquire real-time equipment operation monitoring data; the equipment operation monitoring data includes: the original time-domain vibration signal corresponding to the equipment vibration data and the equipment operation parameter vector; The equipment operation monitoring data is input into a pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operating parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer. The multi-source feature fusion layer performs multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing on the equipment operation monitoring data to determine the fused feature vector, including: performing the following steps through the multi-source feature fusion layer: Multi-band adaptive spectrum reconstruction is performed on the original time-domain vibration signal corresponding to the equipment vibration data to determine the reconstructed spectrum vector; A joint recursive matrix construction method is employed to combine the reconstructed spectral vector with the envelope information of the original time-domain vibration signal to generate a time-frequency joint feature vector capable of simultaneously characterizing time-domain dynamics and frequency-domain energy distribution patterns. This includes: extracting the envelope signal from the original time-domain vibration signal; dividing the envelope signal and the reconstructed spectral vector into several non-overlapping segments, and calculating the root mean square value and spectral centroid of each segment to obtain an envelope root mean square sequence and a spectral centroid sequence; constructing a joint recursive matrix based on the root mean square sequence and the spectral centroid sequence; and reshaping the joint recursive matrix into a one-dimensional vector as the time-frequency joint feature vector. The reconstructed spectrum vector, the time-frequency joint feature vector, and the device operating parameter vector are fused to determine the fused feature vector; The multi-head adaptive gating attention layer dynamically adjusts the activation intensity of each feature dimension in the fused feature vector according to the device operating parameter vector, so as to focus on the feature subspace most relevant to the fault and output the attention-weighted feature vector. Through the multi-scale residual classification layer, multi-scale deep features are extracted from the attention-weighted feature vector and aggregated with weights to output a fault category probability vector. Based on the fault category probability vector, the early fault diagnosis results of the equipment health status are determined.

2. The method according to claim 1, characterized in that, The steps of performing multi-band adaptive spectrum reconstruction on the original time-domain vibration signal corresponding to the equipment vibration data and determining the reconstructed spectrum vector include: Perform a fast Fourier transform on the original time-domain vibration signal corresponding to the vibration data of the equipment, and take the amplitude of the single-sided spectrum to obtain the corresponding spectrum amplitude vector. Based on the spectral amplitude vector, the cumulative energy and gradient of the spectral amplitude are calculated. By locating the local maxima of the cumulative energy gradient, the boundary points for dividing the frequency bands are adaptively determined so as to separate the scattered and weak fault features into independent sub-bands. For each sub-band, the enhancement coefficient is adaptively calculated based on the relative relationship between the sub-band average energy and the global average energy. The hyperbolic tangent function is then used for nonlinear enhancement and normalization to obtain the amplitude of the reconstructed spectrum at each frequency index for each sub-band. Based on the amplitude of the reconstructed spectrum at each frequency index corresponding to each sub-band, multiple reconstructed spectra are spliced ​​together in the original frequency order to form a complete, feature-enhanced reconstructed spectrum vector.

3. The method according to claim 1, characterized in that, The training process of the early fault diagnosis model for the equipment is as follows: Obtain a training set of equipment status monitoring data; The equipment condition monitoring data training set is used to train the multi-source feature fusion layer, multi-head adaptive gating attention layer, and multi-scale residual classification layer of the operating parameters to obtain the early fault diagnosis model of the equipment. During training, a composite loss function combining adaptive focus cross-entropy loss and runtime parameter-aware feature consistency loss is used to calculate the loss value.

4. The method according to claim 3, characterized in that, The steps for obtaining the equipment status monitoring data training set include: Acquire historical equipment operation monitoring data; the historical equipment operation monitoring data includes: continuously collecting vibration signal samples of the equipment at different operating stages and different health states at a fixed sampling frequency, and synchronously collecting equipment operation parameter vectors corresponding to the vibration signal samples; Each vibration signal sample and its corresponding equipment operation parameter vector in the equipment operation monitoring data are labeled with a health status tag. The labeled vibration signal sequences, corresponding equipment operating parameter vectors, and their fault category labels are integrated to construct a structured equipment condition monitoring data training set.

5. The method according to claim 3, characterized in that, The composite loss function is determined as follows: Based on the device operating parameter vector, the sample adaptive focus parameter is determined; based on the sample adaptive focus parameter and the predicted probability of the sample to the true class, the adaptive focus cross-entropy loss is determined. Based on the device operating parameter vector, the bandwidth parameter of operating parameter similarity, and the indicator function based on the real fault category label, the consistency loss of operating parameter perception features is determined. The adaptive focus cross-entropy loss and the running parameter-aware feature consistency loss are added together by weights to obtain the composite loss function for model training.

6. An early fault diagnosis device for equipment health management, characterized in that, The device includes: The data acquisition module is used to acquire real-time collected equipment operation monitoring data; the equipment operation monitoring data includes: the original time-domain vibration signal corresponding to the equipment vibration data and the equipment operation parameter vector; The model input module is used to input the equipment operation monitoring data into a pre-trained equipment early fault diagnosis model; the equipment early fault diagnosis model includes: a multi-source feature fusion layer modulated by operating parameters, a multi-head adaptive gating attention layer, and a multi-scale residual classification layer. The model prediction module is used to perform multi-band adaptive spectrum reconstruction, joint recursive matrix construction, and feature fusion processing on the equipment operation monitoring data through the multi-source feature fusion layer to determine the fused feature vector. This includes: performing the following steps through the multi-source feature fusion layer: performing multi-band adaptive spectrum reconstruction on the original time-domain vibration signal corresponding to the equipment vibration data to determine the reconstructed spectrum vector; using a joint recursive matrix construction method, combining the reconstructed spectrum vector with the envelope information of the original time-domain vibration signal to generate a time-frequency joint feature vector that can simultaneously characterize the time-domain dynamics and frequency-domain energy distribution pattern, including: extracting the envelope signal from the original time-domain vibration signal; dividing the envelope signal and the reconstructed spectrum vector into several non-overlapping segments, and calculating the root mean square value and spectral centroid of each segment to obtain... The system generates an envelope root mean square sequence and a spectral centroid sequence. Based on these sequences, a joint recursive matrix is ​​constructed. This joint recursive matrix is ​​then reshaped into a one-dimensional vector, serving as a time-frequency joint feature vector. The reconstructed spectral vector, the time-frequency joint feature vector, and the device operating parameter vector are fused to determine a fused feature vector. Through a multi-head adaptive gating attention layer, the activation intensity of each feature dimension in the fused feature vector is dynamically adjusted according to the device operating parameter vector to focus on the feature subspace most relevant to the fault, outputting an attention-weighted feature vector. Through a multi-scale residual classification layer, multi-scale deep features are extracted from the attention-weighted feature vector and weighted and aggregated to output a fault category probability vector. Based on this fault category probability vector, an early fault diagnosis result for the device's health status is determined.

7. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described in any one of claims 1 to 5.

Images