A power equipment adaptive diagnosis method and system for an industrial edge end

Abstract

The application discloses a kind of power equipment adaptive diagnosis method and system for industrial edge end, comprising: three indexes of gradient importance, memory penalty and frequency energy importance are calculated on network activation map, and channel level binary sparse mask is generated;Forward propagation retains complete activation value, and when reverse caching, sparse storage is carried out according to mask;Only the first and last layers are normalized parameter maintenance multi-particle copy, and particle number is adjusted with working condition change rate, and sparse mask is used to project dimension reduction on kernel function;With particle swarm diversity as constraint, sparse rate as target, Lagrange dual problem is constructed, and diversity threshold is driven by working condition change rate, and double-time-scale dual descent closed loop update is adopted.The application realizes the precision and memory efficiency balance of variable working condition fault diagnosis under the limited resources of edge end, and has good industrialization prospect.

Description

Technical Field

[0001] This invention belongs to the field of industrial internet and intelligent equipment operation and maintenance technology, specifically involving an adaptive diagnostic method and system for power equipment at the industrial edge, applicable to fault diagnosis of transmission components such as rolling bearings and gearboxes, and can be deployed on industrial internet platforms, intelligent edge computing devices and high-end equipment predictive maintenance systems. Background Technology

[0002] Power equipment (such as fans, compressors, pump sets, gearboxes, etc.) is widely used in industries such as power, metallurgy, petrochemicals, and mining. Failures in their key transmission components (bearings, gears, etc.) can lead to unplanned downtime, causing significant economic losses or even safety accidents. Deep learning-based fault diagnosis methods, due to their automatic feature extraction and high-precision discrimination capabilities, have become the mainstream technology in this field.

[0003] However, the actual operating conditions of power equipment are usually dynamic. Factors such as load fluctuations, speed changes, start-stop switching, and environmental interference cause the statistical distribution of monitoring signals to continuously drift. Traditional deep learning models are based on the assumption that training and test data are independently and identically distributed. Once the operating conditions change, the diagnostic accuracy of the pre-trained model decreases significantly, and the false diagnosis rate increases. Therefore, the model needs to have online adaptive capabilities under changing operating conditions.

[0004] Test-Time Adaptation (TTA) allows models to adaptively update online using unlabeled target data during the inference phase without modifying the original training process, making it an effective paradigm for solving the aforementioned adaptation problem under varying operating conditions. However, in industrial settings, models are typically deployed on edge devices with extremely limited available memory and computing power. Existing TTA methods face the following shortcomings when deployed at the edge:

[0005] (1) The activation cache memory usage during backpropagation is too high. TTA needs to perform backpropagation to update model parameters during inference, which requires caching the activation values ​​of intermediate layers during forward propagation for gradient calculation. The memory usage of activation values ​​is much higher than that of the model parameters themselves, which constitutes a deployment bottleneck under the limited memory resources at the edge, and seriously restricts the promotion and application of deep learning-based fault diagnosis methods in industrial fields.

[0006] (2) Lack of physical priors for industrial signals. The adaptation process of existing TTA methods relies solely on gradient information and does not utilize the frequency domain physical characteristics of industrial signals (such as bearing fault characteristic frequencies, gear meshing frequencies, etc.). Therefore, it is unable to distinguish between key feature channels carrying weak fault impact information and noise interference channels. In noisy industrial environments, without physical prior guidance, these key channels are easily misjudged as redundant information and suppressed, leading to a decrease in diagnostic reliability.

[0007] (3) Single-model adaptation is prone to getting trapped in local optima, while multi-model adaptation has excessive computational overhead. Single models lack constraints in the unsupervised adaptation process and are prone to irreversible collapse caused by erroneous pseudo-labels under complex and variable working conditions. Introducing multi-particle Bayesian inference can maintain diversity, but traditional methods need to calculate the interaction between particles in the global parameter space. The computational complexity increases with the parameter scale, exceeding the computing power capacity of edge devices.

[0008] To address the above issues, it is necessary to develop an adaptive diagnostic method and system for power equipment at the industrial edge. Summary of the Invention

[0009] To address the aforementioned issues, this invention discloses an adaptive diagnostic method and system for power equipment at the industrial edge. It employs a progressive architecture of "time-frequency sensing sparse mask generation, forward-stabilized reverse sparse caching, computationally adaptive hierarchical particle inference, and condition-driven dual closed-loop adaptive processing." Using vibration acceleration signals as input and fault category probability vectors as output, it achieves a dynamic balance between pruning rate and model diversity. This enables online adaptive diagnostics under varying operating conditions without increasing hardware costs, demonstrating promising industrialization prospects.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] An adaptive diagnostic method for power equipment at the industrial edge includes the following steps:

[0012] Step 1, Time-Frequency Aware Sparse Mask Generation: On the output activation map of each layer of the network or selected key layers, calculate the three indices of gradient importance, memory penalty, and frequency domain energy importance respectively, fuse them to generate hierarchical importance, and construct a channel-level binary sparse mask accordingly; at the same time, extract the transient statistics of the vibration signal as a proxy index of the rate of change of the working condition.

[0013] Step 2, Forward Stabilized Reverse Sparse Cache: During forward propagation, the complete activation values ​​are retained and the running statistics of the normalized layer are updated; the mask generated in Step 1 takes effect when the activation values ​​are written to the reverse propagation cache, only the data of the unpruned channels are cached, stored in sparse tensor format, and the original complete tensor is released immediately.

[0014] Step 3, computational power adaptive hierarchical particle inference: maintain multiple particle copies only on the normalized parameters of the first feature extraction layer and the last classification layer of the network, and dynamically adjust the number of particles according to the rate of change of the working condition; use the sparse mask to project and reduce the dimensionality of the parameter space so that the distance metric of the kernel function is only performed in the active channel subspace.

[0015] Step 4, Working Condition-Driven Dual Closed-Loop Adaptation: Using the geometric diversity of the particle swarm as a hard constraint and the sparsity rate as the optimization objective, a Lagrange dual problem is constructed; the diversity threshold is driven by the working condition change rate, and a dual time-scale descent algorithm is used to achieve closed-loop update; the prediction results of each particle are averaged to output the fault category probability vector.

[0016] The core innovation of this invention can be summarized as follows:

[0017] ① Existing activation sparsity methods (such as forward pruning) prune activation values ​​during forward propagation, leading to drastic drift in batch normalization layer statistics and decreased adaptive stability. This invention postpones the sparsity operation until the activation values ​​are written to the backpropagation cache, preserving complete features during forward propagation. This ensures memory release while maintaining the stability of normalization layer statistics.

[0018] ② Existing Bayesian multi-particle methods require calculating inter-particle kernel functions across the entire parameter space, resulting in complexity that increases quadratically with the parameter size, limiting their use at the edge. This invention utilizes the isomorphism between sparse masks and normalized layer parameters in the channel dimension for projection dimensionality reduction, restricting kernel function calculation to the pruned active subspace and allocating multi-particle replicas only to the first and last layers. This design significantly reduces computational complexity while maintaining multi-path exploration capabilities.

[0019] ③ Existing methods mostly employ open-loop static control, which cannot dynamically adjust the adaptation strategy according to changes in operating conditions and model state. This invention constructs a Lagrange dual optimization framework: using the rate of change of operating conditions as a feedforward signal, it actively relaxes pruning constraints when operating conditions change abruptly; using particle swarm diversity as a feedback signal, it automatically adjusts the sparsity intensity through dual descent on two time scales, achieving a dynamic optimal balance between "saving memory" and "maintaining accuracy".

[0020] The three innovations mentioned above are not independent additions, but rather synergistic: Innovation ① provides the basis for memory release, Innovation ② achieves low-cost diversity maintenance within the released subspace, and Innovation ③ dynamically adjusts ① and ② through feedforward-feedback coupling, enabling the system to maintain a "memory-precision" balance under different operating conditions.

[0021] Furthermore, in step one, the frequency domain energy importance is calculated as follows: Fourier transform is performed on the activation map along the time-series dimension to calculate its energy proportion in the fault characteristic frequency range; the fault characteristic frequency range is calculated in real time based on the equipment's operating parameters and transmission structure geometric parameters, or determined through spectral envelope analysis of the equipment's historical steady-state operating data.

[0022] Furthermore, in step one, the proxy index for the rate of change of operating conditions is any single item or weighted combination of the root mean square of energy, peak factor, absolute mean, skewness, and kurtosis.

[0023] Furthermore, in step one, the pruning rate is determined by both the hierarchical basic importance and the active relaxation factor; the active relaxation factor is driven by the rate of change of operating conditions, the pruning rate decreases when the operating conditions change drastically, and increases when the operating conditions are stable.

[0024] Furthermore, in step two, the runtime statistics of each normalization layer are updated normally during forward propagation, and the sparsification operation does not change the forward computation graph.

[0025] Furthermore, in step three, the method for dynamically adjusting the number of particles according to the rate of change of working conditions is as follows: when it is necessary to increase the number of particles, orthogonal Gaussian noise is applied to initialize the split based on the particle with the highest current diversity; when it is necessary to reduce the number of particles, the particle with the lowest historical contribution is discarded.

[0026] Furthermore, in step three, the kernel function adopts a radial basis function, and its distance metric is calculated only in the active channel subspace where the binary sparse mask has a value of 1.

[0027] Furthermore, in step three, the updating of multi-particle parameters employs Stan variational gradient descent, Monte Carlo random deactivation, or Markov chain Monte Carlo methods.

[0028] Furthermore, in step four, the diversity safety threshold is dynamically determined by the rate of change of operating conditions. When the rate of change of operating conditions increases, the threshold increases, and when the rate of change of operating conditions decreases, the threshold decreases.

[0029] Furthermore, in step four, in the dual-timescale descent algorithm, the main variable is updated with a first learning rate, and the dual variable is updated with a second learning rate, wherein the second learning rate is less than the first learning rate.

[0030] The present invention also provides an adaptive diagnostic system for power equipment at the industrial edge, for performing the above method, comprising:

[0031] The module includes a data acquisition module, a time-frequency sensing sparse mask generation module, a forward-stabilized reverse sparse cache module, a computing power adaptive hierarchical particle inference module, a working condition driven dual closed-loop adaptive module, and a diagnostic result output module.

[0032] Each of the above modules is used to execute the methods described in steps one to four. They can be software functional units running in edge computing devices or logical functional modules formed by a processor executing computer programs.

[0033] The beneficial effects of this invention are as follows:

[0034] (1) Resource efficiency: The reverse caching sparsity mechanism reduces the memory usage of activation values ​​without affecting the stability of the normalization layer statistics; through triple dimensionality reduction of only the first and last layers, mask modulation kernel function and dynamic adjustment of particle number, multi-particle inference can run at the edge.

[0035] (2) Diagnostic reliability: The frequency domain energy importance index ensures that the feature channels corresponding to the key fault frequency band are not mistakenly pruned, thereby improving the feature extraction capability of the model in a strong noise environment; the operating condition change rate simultaneously drives the adjustment of the pruning rate and the diversity threshold, so that the system maintains stable diagnostic performance under different operating conditions.

[0036] (3) This invention can be deployed on industrial edge computing devices to support the construction of predictive maintenance systems for equipment in pillar industries of the national economy such as power, metallurgy, and petrochemicals. It has direct industrial application value for ensuring the safe operation of major equipment and reducing unplanned downtime losses. Attached Figure Description

[0037] Figure 1 This is a flowchart of the overall method of the present invention.

[0038] Figure 2 The network structure diagram of the model is shown in the example. Detailed Implementation

[0039] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0040] This embodiment uses rolling bearings as the diagnostic object. The dataset was collected using a vibration acceleration sensor on a bearing fault diagnosis test bench. Wire cutting technology was used to machine the outer ring, inner ring, and rolling elements of the bearing to simulate three typical faults. The dataset includes four operating states: normal (N), inner ring fault (IB), outer ring fault (OB), and rolling element fault (TB). The sampling frequency is 13kHz. Three speed conditions are included: 500 r / min, 700 r / min, and 900 r / min.

[0041] The bearing used in this embodiment is model UCP210, and its internal bearing is a UC210 deep groove ball bearing.

[0042] This embodiment sets up a variable operating condition migration task: the training set uses 500r / min operating condition data (labeled), and the test set uses 700r / min and 900r / min operating condition data (unlabeled), which arrive in a streaming manner in time order to simulate the operating condition shift caused by changes in rotation speed in a real industrial scenario.

[0043] Data partitioning: Under the 500 r / min operating condition, 600 samples of each type of fault are provided, divided in a 3:1 ratio into a training set (450 samples per class, totaling 1800 samples) and a validation set (150 samples per class, totaling 600 samples) for pre-training. Under the two target operating conditions of 700 r / min and 900 r / min, 100 samples of each type of fault are used as test sequences (i.e., 400 test samples per operating condition, totaling 800 samples for both operating conditions). All test samples are unlabeled and used only for online adaptation and evaluation.

[0044] The network input layer receives the original vibration signal with dimensions B×1×T, where B is the batch size (16 in this embodiment) and T is the time length of each sample (2400 in this embodiment).

[0045] This embodiment uses a one-dimensional convolutional neural network, such as Figure 2 As shown, the network has a 9-layer structure, with 7 layers containing batch normalization layers. The first layer is the Stem layer, which is responsible for mapping the single-channel input to 64-channel features and is selected as the first feature extraction layer in multi-particle updates. The ninth layer is a fully connected classification layer, which maps the features to a K-dimensional output (K=4 in this embodiment, corresponding to four classes of running states) and is selected as the last classification layer in multi-particle updates. The second to eighth layers are intermediate layers, all of which maintain single-particle updates.

[0046] Example 1

[0047] like Figure 1 As shown in this embodiment, an adaptive diagnostic method for power equipment at the industrial edge includes:

[0048] Step 1: Detailed Implementation of Rolling Bearing Variable Operating Condition Diagnosis

[0049] For the output activation graph of layer l (B represents batch size, C) l L is the number of channels. l (Time length)

[0050] (1) Gradient importance (G) l )

[0051] ;

[0052] in For unsupervised adaptive loss, this embodiment uses entropy minimization loss. Let be the weight gradient of the c-th channel in the l-th layer.

[0053] (2) Memory penalty (S) l ):

[0054] To penalize layers that consume a lot of memory (incentivizing them to generate a higher pruning rate):

[0055] ;

[0056] Where size (⋅) is the number of bytes of physical memory occupied by the activation graph, ε = 10 −8 To prevent logarithmic operations from overflowing with extremely small positive numbers. In this formula, the layer with the larger memory footprint of the activation graph has, the higher its S... l The smaller the value, the higher the pruning rate, thereby freeing up physical memory for large feature maps.

[0057] (3) Importance of frequency domain energy (F) l )

[0058] ;

[0059] Where Ω fault The fault characteristic frequency range This is a Fast Fourier Transform (FFT). In this embodiment, frequency domain energy importance is calculated only in the first layer, and deeper layers directly inherit the frequency attention weights from the first layer. The reason is that after the network passes through convolutional and pooling layers with a stride greater than 1, the temporal length of the feature map is shortened, and the equivalent sampling rate has changed. If the original physical frequency is used to index the FFT spectrum in deeper layers, it may locate the wrong frequency band.

[0060] The bearing used in this embodiment is a UCP210, which is a UC210 deep groove ball bearing. Based on the bearing's geometric parameters (pitch circle diameter, rolling element diameter, number of rolling elements) and the rotational frequency at the current speed, the fault characteristic frequencies of the outer ring, inner ring, and rolling elements are calculated, and ±5Hz of each frequency point is taken as the fault characteristic interval. The training set has a rotational speed of 500 r / min, and the test set includes speeds of 700 r / min and 900 r / min. The fault characteristic frequency shifts proportionally with the rotational speed; therefore, the fault characteristic interval is dynamically calculated based on the real-time rotational frequency of the current target operating condition. In the FFT spectrum of the first-layer activation graph A1, the energy proportion F1 is calculated according to the fault characteristic interval corresponding to the current operating condition.

[0061] (4) Proximity index for the rate of change of operating conditions:

[0062] ;

[0063] Where RMS is the root mean square energy, CF is the peak factor, and RMS ref and CF ref The reference steady-state value is the statistical mean of the initial batches of data (e.g., 5-10 batches).

[0064] (5) Mask generation:

[0065] Normalization and integration: Where Norm(·) is the maximum-minimum normalized to [0, 1]. This embodiment uses a product method for fusion.

[0066] Active relaxation factor: α is a preset sensitivity coefficient, the value of which is preset according to the equipment type and the fluctuation range of the operating conditions (e.g., between 1.5 and 3.0, and in this embodiment, it is exemplarily taken as 2.0).

[0067] Pruning rate: In this embodiment, if δ l If the calculation result exceeds the interval [0, 1], then take its nearest boundary value (0 if less than 0, 1 if greater than 1).

[0068] According to δ l Sort each channel in descending order of importance score, retaining the top (1-δ) channels. l The most important channels are set to 1 (mask value is set to 1), and the remaining channels are set to 0, generating a channel-level binary mask. .

[0069] Step 2: Implementation of Forward Stable Reverse Sparse Cache

[0070] Activation values ​​A of each layer during forward propagation l The entire process is preserved, and the mean and variance of the BN layer are updated normally. Before backpropagation, a sparsity operation is performed on the activation values ​​of each layer: Only the activation values ​​of the unpruned channels are compressed and stored as sparse tensors (non-zero activation values ​​plus position indices), and the original complete tensors are released immediately after the cache is written.

[0071] Step 3: Specific Implementation of Computational Power Adaptive Hierarchical Particle Inference

[0072] In this embodiment, only the BN parameters of layer L1 and layer L9 undergo multi-particle updates. The basic particle number P base =3, Dynamic Particle Number:

[0073] ;

[0074] Where P min and P max These represent the upper and lower limits of the particle count, set according to the edge computing power (3 and 5 in this example); γ is the sensitivity coefficient (1.5 in this example). When P t > P t-1 At that time, using the particle with the highest current diversity as the parent particle, orthogonal Gaussian noise is applied for splitting initialization. When P t < P t-1At that time, calculate the historical contribution of each particle (cumulative confidence of predictions in the most recent N batches), discard the particle with the lowest contribution, and retain the best principal particle.

[0075] Particle updates, facilitated by the kernel function, ensure that particles remain separated in the parameter space to maintain diversity. Simultaneously, to reduce computational overhead at the edges, the distance metric of the kernel function should only be performed within the unpruned active channel subspace. In this embodiment, the sparse kernel function is the RBF kernel, defined as follows:

[0076] ;

[0077] in Here, h represents the parameters of the i-th particle in the l-th layer, and h represents the kernel bandwidth (adaptive median distance bandwidth is used in this embodiment).

[0078] Based on the kernel function described above, the SVGD algorithm is used to update the particles in the first and last layers, as shown in the following formula:

[0079] ;

[0080] Where ϵ is the learning rate for updating the main variable (i.e., the fast-scale learning rate of the main variable mentioned in step four), and in this embodiment, ϵ is exemplarily taken as 5 × 10. −4 The parameters corresponding to the pruned channels are not included in the calculation of the kernel function's repulsion effect.

[0081] When the mask is projected onto the L9 fully connected layer, since the temporal dimension has been compressed by the preceding global average pooling layer, the one-dimensional channel-level sparse mask is directly broadcast-aligned with the input feature dimension (column vector) of the weight matrix of the fully connected layer, thereby shielding the pruned input neuron connections when the repulsion effect of the kernel function is calculated.

[0082] Step 4: Specific Implementation of Dual Closed-Loop Adaptive Operation Driven by Operating Conditions

[0083] First, we measure the diversity of the particle swarm. Let there be a total of P particles. t Let there be n particles, and denot θ as the total number of trainable parameters for the i-th particle. i The mean parameter and population variance of the particle swarm are as follows:

[0084] , ;

[0085] Use V t This represents a measure of the diversity of the current particle swarm. The larger the value, the greater the difference in parameters among the particles, and the more thoroughly the model explores different optimization paths.

[0086] Secondly, calculate the diversity safety threshold under the current operating conditions. The more drastic the changes in operating conditions, the higher the diversity required by the model to adapt to the new data distribution. The diversity safety threshold is dynamically determined by the rate of change of operating conditions.

[0087] ;

[0088] in λ is the base threshold (0.3 in this example), and λ is the coefficient (0.5 in this example). When the operating conditions are stable, the threshold falls back, allowing particles to converge; when the operating conditions change abruptly, the threshold rises, forcing the particle swarm to maintain sufficient dispersion.

[0089] Then, update the sparsity penalty coefficient (dual variable) β. This coefficient controls the balance between sparsity and diversity: the larger β is, the stronger the penalty for high sparsity, tending to reduce sparsity (i.e., retain more channels); the smaller β is, the weaker the penalty for high sparsity, tending to increase sparsity (i.e., prune more channels to save memory). The update of β is as follows:

[0090] ;

[0091] Where η is the learning rate of the dual variable, and its value should be much smaller than the learning rate ϵ of the main variable. In this example, it is taken as 1×10. −5 .when As β increases, the pruning rate decreases, and the system retains more feature channels to restore particle swarm diversity; when diversity is sufficient ( As β decreases, the pruning rate increases, and the system increases the pruning rate to save memory. The updated β result is fed back to the next batch of mask generation process, forming a closed-loop control. In this embodiment, the initial value of β is exemplarily set to β0 = 0.01.

[0092] Finally, the predicted probability distributions of each particle are averaged to obtain the final diagnostic output:

[0093] ;

[0094] in Let x be the predicted class probability vector for the i-th particle pair with input x. The final output is the class with the highest probability. .

[0095] Experimental results:

[0096] The following table compares the test accuracy and peak memory usage of this invention with existing methods:

[0097] method Test accuracy (%) Peak memory usage (MB) Source (Unadapted original model) 73.2 356 Traditional TTA method (updates BN layer parameters without using any sparsity strategy) 82.3 356 This invention 92.7 58

[0098] This embodiment achieved a test accuracy of 92.7% in continuous variable operating condition migration tasks, an improvement of 19.5 percentage points compared to the unadapted model and 10.4 percentage points compared to the traditional TTA method. Peak memory usage was 58MB, a reduction of 83.7% compared to the traditional solution (356MB). These experimental results demonstrate that the present invention achieves both a significant improvement in diagnostic accuracy and an order-of-magnitude reduction in memory usage in typical industrial scenarios with limited memory at the edge, verifying its feasibility and effectiveness in practical deployment in the field of intelligent manufacturing.

[0099] To further verify the effectiveness of the core innovations of this invention, the following ablation experiment was conducted:

[0100] method Test accuracy (%) Peak memory usage (MB) illustrate <![CDATA[Variant A: Step 1 Remove the frequency energy importance indicator F l , and the pruning basis is only G l × S l > 88.7 58 Verify the contribution of frequency domain physical priors Variant B: Full-layer particleization (all normalized layers are particleized) 94.2 >1GB Validating the importance of hierarchical particleization in edge deployment Variant C: Remove dual loop closure (fixed divergence threshold) 89.5 58 Verify the contribution of closed-loop control to stability. This invention 92.7 58

[0101] The above results indicate that frequency domain physical priors and closed-loop control each contribute to diagnostic accuracy; although variant B (fully particle-based) has higher accuracy, its memory requirements far exceed those of the method of this invention, which would severely limit its deployment at the edge.

[0102] The parameter values ​​and network structure described above are merely a specific example of the present invention and do not constitute a limitation on the scope of protection of the present invention. Those skilled in the art can adjust the specific values ​​according to the actual device and edge resource conditions, which still falls within the scope of the present invention.

[0103] Example 2: Gearbox Variable Operating Condition Diagnosis

[0104] This embodiment uses a gearbox as the diagnostic object. The publicly available dataset was collected on a gearbox fault simulation test bench. The test gear pair consists of a driving gear (55 teeth) and a driven gear (75 teeth). An accelerometer was installed on the gearbox housing to collect vibration signals at a sampling frequency of 5120Hz. The dataset includes six health states: normal, driven gear pitting, driven gear tooth breakage, driving gear tooth wear, a combined fault of driven gear pitting and driving gear tooth wear, and a combined fault of driven gear tooth breakage and driving gear tooth wear. It also includes four load conditions: 0W, 10W, 25W, and 50W.

[0105] This embodiment sets up a variable load condition migration task. The training set (source domain) uses two load condition data (labeled) of 0W and 10W, and the test set (target domain) uses two load condition data (unlabeled) of 25W and 50W. They arrive sequentially in a streaming manner to simulate the continuous adaptation process of gradually changing load in a real industrial scenario.

[0106] The data partitioning method and network structure are the same as in Example 1, and will not be repeated here.

[0107] [Explanation of Key Differences in Step One]

[0108] The difference between this embodiment and Embodiment 1 lies in the fault characteristic frequency range Ω in step one. fault The method for determining the fault characteristics is as follows: For gearboxes, the fault characteristics manifest as the meshing frequency and its sidebands. The meshing frequency is determined by the product of the number of gear teeth and the rotational frequency. When a local fault occurs in a gear, the fault energy is concentrated at the meshing frequency and its two sidebands. In this embodiment, the range of ±3 times the rotational frequency of the meshing frequency is used as the fault characteristic interval. With changes in load, when the rotational speed remains constant, the position of the meshing frequency remains unchanged, while the amplitude of the sideband changes but its position remains unchanged; if the rotational speed changes, the meshing frequency shifts proportionally with the rotational frequency, and the fault characteristic interval should be dynamically calculated based on the real-time rotational frequency under the current operating conditions. The specific implementation methods of the remaining steps are the same as in Embodiment 1.

[0109]

Experimental Results

[0110] On continuous variable load test sequences of 25W and 50W, the average test accuracy of this invention in gearbox fault diagnosis tasks was 89.1%, an improvement of 13.6% compared to the Source (unadapted model); the peak memory usage was 58MB. These results verify the applicability of this invention to gearbox-type equipment.

[0111] Each step of this invention can be deployed as an edge software functional module, running on devices such as embedded industrial control computers (e.g., Jetson Xavier NX), intelligent acquisition terminals, or edge gateways. The data acquisition module interfaces with the accelerometer, and the diagnostic results are sent to the host computer or cloud platform via industrial Ethernet or fieldbus for reference by maintenance personnel.

[0112] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for online fault diagnosis of power equipment under varying operating conditions, characterized in that, Includes the following steps: Step 1, Time-Frequency Aware Sparse Mask Generation: On the output activation map of each layer of the network or selected key layers, calculate the three indices of gradient importance, memory penalty, and frequency domain energy importance respectively, fuse them to generate hierarchical importance, and construct a channel-level binary sparse mask accordingly; at the same time, extract the transient statistics of the vibration signal as a proxy index of the rate of change of the working condition. Step 2, Forward Stabilization and Backward Sparse Cache: During forward propagation, the complete activation values ​​are preserved to keep the running statistics of the normalization layer stable; the mask generated in Step 1 takes effect when the activation values ​​are written to the backpropagation cache, only caching the data of the unpruned channels, storing it in sparse tensor format, and the original complete tensor is released immediately. Step 3, computational power adaptive hierarchical particle inference: Multiple particle copies are maintained only on the normalized parameters of the first feature extraction layer and the last classification layer of the network, and the number of particles is dynamically adjusted according to the rate of change of the working condition. The sparse mask is used to project and reduce the dimensionality of the parameter space, so that the distance metric of the kernel function is performed only in the active channel subspace. Step 4, Working condition driven dual closed-loop adaptation: Using the geometric diversity of the particle swarm as a hard constraint and the sparsity as the optimization objective, construct the Lagrange dual problem; The diversity threshold is driven by the rate of change of the operating conditions and is updated in a closed loop using a dual time-scale descent algorithm. The prediction results of each particle are averaged to output a fault category probability vector.

2. The method according to claim 1, characterized in that, In step one, the frequency domain energy importance is calculated as follows: Fourier transform is performed on the activation map along the time dimension to calculate its energy proportion in the fault characteristic frequency range; the fault characteristic frequency range is calculated in real time based on the equipment's operating parameters and transmission structure geometric parameters, or determined by spectral envelope analysis of the equipment's historical steady-state operating data.

3. The method according to claim 1, characterized in that, In step one, the pruning rate is determined by both the hierarchical basic importance and the active relaxation factor. The active relaxation factor is driven by the rate of change of operating conditions. When the operating conditions change drastically, the pruning rate decreases, and when the operating conditions are stable, the pruning rate increases. The proxy index for the rate of change of operating conditions is any single item or weighted combination of the root mean square energy, peak factor, absolute mean, skewness, and kurtosis.

4. The method according to claim 1, characterized in that, In step two, the runtime statistics of each normalization layer are updated normally during forward propagation, and the sparsification operation does not change the forward computation graph.

5. The method according to claim 1, characterized in that, In step three, the number of particles changes dynamically with the rate of change of the working conditions: when it is necessary to increase the number of particles, orthogonal Gaussian noise is applied to initialize the split based on the particle with the highest current diversity; when it is necessary to reduce the number of particles, the particle with the lowest historical contribution is discarded.

6. The method according to claim 1, characterized in that, In step three, the kernel function adopts a radial basis function, and its distance metric is calculated only in the active channel subspace where the binary sparse mask has a value of 1.

7. The method according to claim 1, characterized in that, In step three, the updating of multi-particle parameters employs Stan variational gradient descent, Monte Carlo random deactivation, or Markov chain Monte Carlo methods.

8. The method according to claim 1, characterized in that, In step four, the diversity safety threshold is dynamically determined by the rate of change of operating conditions. When the rate of change of operating conditions increases, the threshold increases, and when the rate of change of operating conditions decreases, the threshold decreases.

9. The method according to claim 1, characterized in that, In step four, in the dual-timescale descent algorithm, the main variable is updated with a first learning rate, and the dual variable is updated with a second learning rate, where the second learning rate is less than the first learning rate.

10. An adaptive diagnostic system for power equipment at the industrial edge, characterized in that, include: The data acquisition module is used to collect vibration acceleration signals of key transmission components of power equipment; A time-frequency aware sparse mask generation module is used to perform the operation described in step one of claim 1; A forward stable reverse sparse cache module is used to perform the operation described in step two as claimed in claim 1; The computing power adaptive hierarchical particle inference module is used to perform the operation described in step three as in claim 1; The working condition driven dual closed-loop adaptive module is used to perform the operation described in step four as in claim 1; The diagnostic results output module is used to output the probability vector of the fault category.

Images