A deep learning-based industrial equipment fault prediction method and system

Abstract

The application discloses an industrial equipment fault prediction method and system based on deep learning, comprising: obtaining historical operation data of an industrial equipment to be measured within a set period and preprocessing the historical operation data to generate a sample data set; inputting the sample data set into an industrial equipment fault prediction model constructed based on a deep learning algorithm for learning and training, iteratively updating model parameters by using an adaptive optimizer during the training process, and dynamically adjusting model weights; deploying the trained fault prediction model to a local server, receiving an operation data stream of the industrial equipment in real time in a model operation stage, performing fault probability prediction on the operation data stream, and outputting corresponding fault types and occurrence probabilities. The application combines efficient data preprocessing technology and an adaptive optimization mechanism of a deep learning algorithm, and focuses on solving the problems of low real-time prediction accuracy of equipment operation data and poor model generalization ability.

Description

Technical Field

[0001] This invention relates to the technical field of industrial equipment fault prediction, and in particular to a method and system for industrial equipment fault prediction based on deep learning. Background Technology

[0002] Traditional prediction methods rely on conventional statistical analysis, which depend on the physical parameters of the equipment or expert knowledge, and perform poorly on high-dimensional, highly nonlinear, or non-steady-state data. In recent years, artificial intelligence technology, especially fault prediction methods based on machine learning and deep learning, has gradually become a research hotspot. These methods have the advantages of being data-driven, having strong generalization ability, and being better able to capture complex nonlinear relationships.

[0003] Existing industrial equipment fault prediction methods do not pay enough attention to data preprocessing and model adaptability, making it difficult to dynamically adapt to changes in the operating environment of industrial equipment and maintain stable model performance over a long period of time. Therefore, existing technologies suffer from insufficient generalization of prediction models, poor real-time performance, and weak adaptability. In particular, for complex industrial operating data with multiple dimensions and parameters, the prediction accuracy is difficult to improve effectively and cannot meet the requirements of rapid response and high reliability in actual industrial production.

[0004] CN116089870A discloses a method for predicting industrial equipment faults under small sample conditions based on meta-learning. This technology utilizes meta-learning to achieve cross-domain knowledge transfer, thereby reducing the model's need for target domain data. However, this technical solution relies excessively on the similarity of data distribution between the source domain and the target domain. When the operating conditions of the equipment differ significantly, the model's generalization performance is greatly reduced, and its robustness is poor.

[0005] CN117349619A discloses an industrial equipment fault prediction method that uses Copula functions to construct input features. This method determines input features through Spearman rank correlation and analytic hierarchy process to improve fault prediction accuracy. However, this method still belongs to the traditional statistical analysis framework. The extraction and selection of features rely too much on empirical functions, making it difficult to adapt to dynamically changing equipment operating conditions and unable to efficiently process large-scale, real-time monitoring data. It has certain limitations in actual industrial environments.

[0006] In summary, existing industrial equipment fault prediction technologies generally suffer from poor data adaptability, poor generalization performance, and insufficient real-time prediction accuracy. The deep learning-based industrial equipment fault prediction method proposed in this invention addresses the problems of low real-time prediction accuracy and poor model generalization ability by combining efficient data preprocessing technology and the adaptive optimization mechanism of deep learning algorithms. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the specification abstract and the title of the invention, to avoid obscuring the purpose of this section, the specification abstract, and the title of the invention. Such simplifications or omissions shall not be used to limit the scope of the invention.

[0008] In view of the aforementioned existing problems, the present invention is proposed.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: acquiring historical operating data of the industrial equipment under test within a set time period and preprocessing it to generate a sample dataset;

[0010] The sample dataset is input into an industrial equipment fault prediction model built based on a deep learning algorithm for learning and training. During the training process, an adaptive optimizer is used to iteratively update the model parameters and dynamically adjust the model weights.

[0011] The trained fault prediction model is deployed to a local server. During the model operation phase, the operation data stream of industrial equipment is received in real time. Fault probability prediction is performed on the operation data stream, and the corresponding fault type and its probability of occurrence are output.

[0012] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method described in this invention, the historical operating data includes at least the historical vibration signal, current and voltage signal, bearing temperature signal, speed signal, lubricating oil pressure signal, and equipment operating status of the industrial equipment under test.

[0013] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method of the present invention, the historical operating data is preprocessed to generate the sample dataset, including:

[0014] Wavelet thresholding is used to remove random noise;

[0015] Missing data were filled using spline interpolation, and outliers were removed using the three sigma criterion.

[0016] Each physical quantity is normalized according to its dimensions;

[0017] The normalized data is segmented and labeled with corresponding fault category tags according to a fixed sliding window strategy, and then summarized into a set to obtain the sample dataset.

[0018] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method of the present invention, the sample dataset is input into the industrial equipment fault prediction model constructed based on the deep learning algorithm for learning and training, including:

[0019] The fault prediction model is a series structure of one-dimensional convolutional neural network, bidirectional long short-term memory network, and attention mechanism.

[0020] The cross-entropy loss function is used and iterative training is performed using the AdamW optimizer;

[0021] During training, a Dropout random inactivation rate of 0.3 was set to suppress overfitting, and an early stopping strategy was used to monitor the validation set loss.

[0022] After obtaining the optimal hyperparameter combination through five-fold cross-validation, save the model weights with the highest F1-score in the validation set.

[0023] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method described in this invention, dynamically adjusting the model weights after iteratively updating the model parameters using an optimizer includes:

[0024] During the training of the fault prediction model, the weighted accuracy index of the most recent N prediction results is continuously calculated and compared with the historical baseline;

[0025] When performance degradation is detected to exceed the degradation threshold, incremental online learning is triggered, and newly collected labeled data is inserted into the experience replay buffer.

[0026] After determining the data distribution drift based on the drift detection module, a finite number of iterations of fine-tuning updates are performed to correct the model weights;

[0027] After the update is complete, the performance metrics are recalculated. If they exceed the degradation threshold, the old model weights are replaced; otherwise, the system is rolled back to the historical best weights.

[0028] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method described in this invention, the trained fault prediction model is deployed to a local server, including:

[0029] The fault prediction model is subjected to 8-bit quantization and graph optimization to reduce inference latency;

[0030] The model service is encapsulated using a containerized approach and provides an inference interface to the outside world through a RESTful API.

[0031] Enable the tensor parallel inference engine on the server GPU accelerator card;

[0032] It supports a hot update mechanism, allowing new model weights to be loaded without interrupting service.

[0033] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method described in this invention, the method involves receiving the operational data stream of the industrial equipment in real time during the model running phase, performing fault probability prediction on the operational data stream, and outputting the corresponding fault type and its probability of occurrence, including:

[0034] A fixed sliding window strategy is used to extract data segments from the continuously acquired raw running data stream, and normalization processing is performed on them to obtain normalized tensors;

[0035] The normalized tensor is input into the inference interface deployed on the local server, and the fault prediction model is called to perform vectorized inference, outputting a multidimensional probability vector.

[0036] Candidate fault categories corresponding to the maximum probability values ​​are selected from the multidimensional probability vectors to form prediction result pairs;

[0037] If the maximum probability is greater than or equal to the dynamic threshold, the prediction result pair is returned and published in key-value format in the OPCUA event channel.

[0038] As a preferred embodiment of the deep learning-based industrial equipment fault prediction method described in this invention, the candidate fault types include at least bearing wear, misalignment, rotor imbalance, insufficient lubrication, coil short circuit, overload operation, abnormal temperature rise, and power fluctuation.

[0039] As a preferred embodiment of the deep learning-based industrial equipment fault prediction system of the present invention, it includes: one or more processors;

[0040] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the deep learning-based industrial equipment fault prediction method described above.

[0041] As a preferred embodiment of the computer-readable medium for storing software according to the present invention, the software includes instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, the operations including the process of the aforementioned deep learning-based industrial equipment fault prediction method.

[0042] The beneficial effects of this invention are:

[0043] 1. By denoising, outlier processing, missing value imputation, and data standardization of the collected data, the data errors introduced by the equipment operating environment and the accuracy deviation of the sensor itself are eliminated, the quality of the input data is improved, irrelevant interference factors are eliminated, and the reliability of subsequent model training is improved.

[0044] 2. By using deep learning algorithms (such as a combination of convolutional neural networks, long short-term memory networks and attention mechanisms), the nonlinear temporal characteristics of industrial equipment operation data can be deeply mined. At the same time, an adaptive optimizer (such as the AdamW optimizer) is used to automatically find the globally optimal parameters and dynamically adjust the model weights based on the loss function feedback during training. This automatically adapts to the data characteristics and patterns under different industrial equipment operating conditions, avoiding the experience bias and misjudgment problems that are easy to occur when manually setting model parameters in traditional methods. This improves the generalization performance, stability and adaptability of the fault prediction model, and achieves the effect of long-term stable operation of the model performance.

[0045] 3. By deploying the model locally on the server, the inference process is brought closer to the data source on the production site, reducing network transmission latency. At the same time, the model receives operational data in real time, uses a sliding window method to dynamically capture and process the data, and outputs a multi-dimensional probability vector. Finally, based on the probability threshold, the specific fault type and its corresponding probability of occurrence are determined, enabling timely and accurate identification of potential equipment fault states and precise quantification of the probability of fault occurrence. This facilitates maintenance personnel or control systems to plan maintenance strategies in advance, avoid unexpected downtime events, reduce the risk of downtime caused by unexpected equipment failures, improve equipment operational reliability and production continuity, thereby reducing economic losses and improving production efficiency. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0047] Figure 1 This is a flowchart illustrating the deep learning-based industrial equipment fault prediction method of the present invention.

[0048] Figure 2 This is a schematic diagram of the softmax operation framework of the deep learning network shown in this invention. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0050] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] According to an embodiment of the present invention, in combination Figure 1 and Figure 2 The flowchart shown illustrates a deep learning-based method for predicting industrial equipment faults, which includes the following steps:

[0053] S1. Obtain historical operating data of the industrial equipment under test within a set time period and preprocess it to generate a sample dataset. Note that the following points should be noted in this step:

[0054] A multimodal sensor array is deployed at the key operating parts of the industrial equipment under test. The sensor array includes at least a triaxial vibration sensor, a current transformer, a voltage sampling module, a bearing temperature thermocouple, a photoelectric tachometer, and a lubricating oil pressure transmitter.

[0055] Synchronous sampling is performed according to a unified time reference, with a sampling frequency f. s Satisfy f s ≥2f max , where f max This is the highest effective frequency of the equipment vibration signal;

[0056] The multidimensional original signal stream obtained from continuous sampling is denoted as:

[0057]

[0058] Where t is the sampling time, R(t) is the six-dimensional original signal column vector, and r v (t) represents the vibration signal, r i (t) represents the current signal, r u (t) is the voltage signal, r T (t) represents the bearing temperature, r ω (t) is the rotational speed signal, r p (t) represents the lubricating oil pressure;

[0059] Perform piecewise wavelet thresholding denoising on each component of R(t);

[0060] Selecting Daubechies-4 basis wavelets, the decomposition level is [number missing]. Among them, f b For the signal fundamental frequency, a soft threshold is used as the threshold function:

[0061]

[0062] in, The denoised signal, where λ is the wavelet threshold, σ is the standard deviation of noise estimation, and N is the number of sampling points;

[0063] Obtain the denoised vector:

[0064]

[0065] For R dn (t) Perform timestamp integrity check;

[0066] Missing sampling points were filled using cubic spline interpolation.

[0067] Three Sigma criteria are applied to each dimension to remove outliers. Outlier criteria are as follows:

[0068] |r k (t)-μ k |3σ k ,k∈{v,i,u,T,ω,p}

[0069] Where, μ k ,σ k The mean and standard deviation within the current sliding window are used; outlier samples are initially set as missing and then repaired using spline interpolation.

[0070] Furthermore, map to [0,1] according to the min-max rule:

[0071]

[0072] Obtain a normalized signal R with consistent dimensions and good comparability. norm (t);

[0073] Given a window length τ and a step size δ (e.g., τ = 3s, δ = 0.5s), perform sliding cut on the normalized sequence:

[0074] X j ={R norm (t)∣t∈[t0+(j-1)δ,t0+(j-1)δ+τ]}

[0075] For each window X j Statistical monitoring system operating logs, matching fault event tags y j ∈0,1,2,…,C, where 0 represents the normal state and 1…C represents the codes for each fault type;

[0076] Summarize the initial sample pairs (X) i ,y j This forms the original training set D0;

[0077] If D0 has class imbalance (the number of samples in any c class is n) c <αnmax If α = 0.1), then apply the improved SMOTE oversampling to the minority class to generate synthetic samples;

[0078] The balanced training sample set D bal ;

[0079] D in chronological order bal Renumber and define the final sample dataset:

[0080] D = {(X i ,y i |i=1,2,…,N}

[0081] in, M = 6 represents the number of sampling points within the window, and M = 6 represents the number of signal channels. i The corresponding fault category label is N, where N is the total number of samples.

[0082] In an optional implementation, the sample dataset is stored serialized in HDF5 format, with each record containing a data tensor X. i Tag y i and timestamp t i This is used for subsequent deep learning model training.

[0083] S2. Input the sample dataset into the industrial equipment fault prediction model constructed based on a deep learning algorithm for learning and training. During the training process, an adaptive optimizer is used to iteratively update the model parameters and dynamically adjust the model weights. It should be noted that in this step:

[0084] In the sample dataset D, the training set, validation set and test set are divided in a ratio of 8:1:1, and the batch size of the training set is set to 128. Random balanced sampling is used to ensure that the sample size of each fault category is consistent.

[0085] A series structure of "one-dimensional convolutional neural network → bidirectional long short-term memory network → multi-head attention layer" is constructed. The number of kernels in the convolutional layer is 64, the kernel width is 5, the stride is 1, the number of hidden units in the bidirectional LSTM is set to 128, the number of heads in the attention layer is 4, and a normalization layer is added at the end of each layer.

[0086] The training objective is defined as cross-entropy loss (LCE), and the AdamW adaptive optimizer is used with an initial learning rate of 1.5 × 10⁻⁶. -3 The weight decays to 1×10 -4 The gradient clipping threshold is 5;

[0087] Dropout (p=0.3) is inserted at the output of the convolutional layer and the bidirectional LSTM hidden unit to suppress overfitting. This is effective when the validation set loss does not significantly decrease within 12 consecutive iterations (<10). -4Early stop is triggered when )

[0088] D is divided into five parts, and the training is repeated for 5 rounds. The hyperparameters of kernel width, number of hidden units and learning rate are searched using a grid search. The weights Θ with the highest F1-score in the validation set are saved. * ;

[0089] During training, the weighted accuracy of the most recent N=256 predictions is calculated in real time. (α i (with time decay weights), and compared with the historical baseline A w0 Compare;

[0090] If A w w0 -Δ d (Δ d =0.05) and Jensen–Shannon divergence D JS (P new ||P old If the new labeled sample is greater than θ (θ = 0.2), then the new labeled sample is written to the experience replay buffer, and 5 rounds of iterative fine-tuning are performed. After the accuracy is restored, the old weights are replaced; otherwise, the system rolls back to the historical optimal θ. best .

[0091] As an example, weight Θ * The calculation formula is as follows:

[0092]

[0093] Where, Θ * Let B be the final optimized network parameter tensor, B be the batch size for each gradient update, and b be the sample index within the batch. This represents the truth label (0 or 1) of the b-th sample in class c. The corresponding model output probability is given, where C is the total number of fault types and θ is the drift regularization coefficient. For the unbounded convolution integral over the frequency domain variable ω, erf(·) is the Gaussian error function used to measure the difference between the old and new distributions, P new (ω) and P old (ω) represents the power spectral density of the current streaming data and the training baseline, respectively, and σ ω The standard deviation of the ω-domain is used for normalization, η is the coefficient of the sparse regularization term, and ||Θ||1 is the l1 norm of the parameter tensor to enhance sparsity. Let ε be the magnitude of the model constrained by the l2 norm, and let ε be a very small positive constant to prevent the denominator from being zero.

[0094] It should be noted that the closer the target value is to 0, the smaller the model classification error, the less the distribution drift, and the sparser the parameters, resulting in the best overall performance.​

[0095] For example, the mathematical expression of the trained and quantified fault prediction model is as follows:

[0096]

[0097] Where X is a data tensor of length L and number of channels M = 6; Θ c ,Θ l ,Θ h These are the parameters for the convolution, bidirectional LSTM, and attention layers, respectively. To output the mapping weights and biases, p = [p1, ..., p2]. C ] T Let C = 8 be the probability vector for each fault type, and F = {bearing wear, misalignment, rotor imbalance, insufficient lubrication, coil short circuit, overload operation, abnormal temperature rise,}.

[0098] This invention effectively captures local temporal features and long-term dependencies in industrial equipment operation data through a series structure of a one-dimensional convolutional neural network → a bidirectional LSTM network → a multi-head attention mechanism. For example, it captures sudden anomalies in bearing wear vibration signals and slow trend changes in temperature rise anomalies. This solves the problem that traditional single structures cannot handle temporal features at different scales simultaneously and improves the model's ability to recognize complex data patterns.

[0099] Preferably, by monitoring and calculating the difference between the prediction accuracy and the historical baseline in real time, and combining the Jensen-Shannon divergence to determine data drift, the model's adaptive online fine-tuning is achieved. This solves the problem that traditional deep learning models cannot effectively adapt to dynamic changes in equipment operating conditions after static deployment, and ensures the long-term stability and high robustness of the prediction accuracy.

[0100] S3. Deploy the trained fault prediction model to a local server. During the model execution phase, receive the operational data stream from the industrial equipment in real time, perform fault probability prediction on the operational data stream, and output the corresponding fault type and its probability of occurrence. Note that the following should be noted in this step:

[0101] The trained fault prediction model is subjected to symmetric 8-bit integer quantization and layer fusion graph optimization to generate a low-latency inference graph. During the quantization process, a calibration set is used to calculate the saturation threshold and an integer learnable scale factor is inserted to ensure that the accuracy loss is <1%.

[0102] Package M8 into an image using Docker. fault-pred:ver deploys the FastAPI framework in the image, exposes the RESTful inference endpoint / v1 / predict, receives HTTP POST data tensor payload and returns the prediction result, and also has a built-in gRPC interface in the image for OPC UA gateway to call;

[0103] Load Triton Inference Server on the local server GPU accelerator card, enable TensorRT tensor parallel backend, with parameters instanceGroupCount=4 and pipelineDepth=2, so that the single stream latency is <8ms;

[0104] Configure a Kubernetes rolling update strategy. When a new weighted packet is detected and passes the CI / CD pipeline signature verification, M8 will be automatically hot-replaced within 0.5 seconds. The old session will be recycled after inference is completed within the Grace Period, ensuring 0 packet loss.

[0105] The industrial gateway reports raw operational data at 1kHz using the MQTT protocol. The preprocessing service extracts fragments X(t) using a fixed sliding window (τ,δ) = (3s, 0.5s) and performs Min-Max normalization processing consistent with the training phase to obtain tensors.

[0106] Calling / v1 / predict inputs X′(t) into model M8 to obtain the logits vector z(t) = [z1,…,z C ] T And mapped to a probability vector using SoftMax:

[0107]

[0108] Call / v1 / predict and input X′ k Obtain the probability vector p k =M8(X′) k );

[0109] For example, maximum probability filtering:

[0110]

[0111] For example, dynamic threshold calculation:

[0112] T0 = ​​0.65, k = 0.3

[0113] Where, λ k ∈[0,1] represents the normalized value of the current load level, collected by the PLC, ρ k∈[0,1] represents the average false alarm rate over the past 300 seconds;

[0114] like The equipment is in the early stages of a malfunction and an alarm will be issued.

[0115] like Then output a warning message;

[0116] Otherwise, record it as a normal sample;

[0117] For example, alarm or warning information is output in JSON format:

[0118] {"faultType":"F★","probability":P★,

[0119] "timestamp":ts,"level":"ALARM|WARN"}

[0120] Published via OPC UA event nodes / Fault / Alert and saved to the InfluxDB fault database. pred ;

[0121] As an example, failure probability prediction is performed in real time using the following formula:

[0122]

[0123] Where, p k,c Predict the probability of window k for class c. To maximize probability, For the corresponding fault category, Status k For predicting status labels, ALARM indicates an alarm, WARN indicates a warning, and OK indicates normal.

[0124] Preferably, the present invention effectively solves the problem that when traditional deep learning models are deployed to industrial sites, floating-point operations consume too much GPU resources, resulting in latency that cannot meet real-time requirements. By using symmetric 8-bit quantization and graph fusion optimization, the present invention strictly controls the inference latency to within 8ms, and realizes real-time and refined monitoring of equipment operating status.

[0125] Furthermore, by encapsulating the FastAPI inference service in Docker containers and integrating it with OPC UA / gRPC interfaces using RESTful interfaces, the problem of complex interfaces and difficult deployment in traditional industrial equipment monitoring platforms is effectively solved, greatly improving the convenience of model deployment. In particular, a dynamic threshold calculation and real-time probability decision mechanism is proposed, which adaptively adjusts the threshold according to the real-time load intensity and short-term false alarm rate to achieve fine classification of fault prediction results (alarm, warning, normal), ensuring an optimal trade-off between false alarms and missed alarms, and significantly improving the accuracy and reliability of fault diagnosis.

[0126] The aforementioned preprocessing methods for the collected data can be carried out using existing technologies and methods, and will not be elaborated further in this example.

[0127] In addition to the above embodiments, other aspects of the present invention also propose a deep learning-based industrial equipment fault prediction system, comprising: one or more processors and a memory.

[0128] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the deep learning-based industrial equipment fault prediction method of the foregoing embodiments, in particular... Figure 1 The flowchart of the method is shown.

[0129] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the deep learning-based industrial equipment fault prediction method of the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0130] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0131] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0132] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0133] In any case, the language can be either compiled or interpreted.

[0134] Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit.

[0135] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0136] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0137] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0138] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0139] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for predicting industrial equipment faults, characterized in that, include: Acquire historical operating data of the industrial equipment under test within a set time period and preprocess it to generate a sample dataset; The sample dataset is input into an industrial equipment fault prediction model built based on a deep learning algorithm for learning and training. During the training process, an adaptive optimizer is used to iteratively update the model parameters and dynamically adjust the model weights. After iteratively updating the model parameters using the optimizer, the dynamic adjustment of the model weights includes: during the training of the fault prediction model, continuously calculating the weighted accuracy index of the most recent N prediction results and comparing it with the historical baseline; when performance degradation is detected to exceed the degradation threshold, triggering incremental online learning and inserting newly collected labeled data into the experience replay buffer; after determining the data distribution drift based on the drift detection module, performing a finite-iteration fine-tuning update to correct the model weights; after the update is completed, recalculating the performance index, and if it is higher than the degradation threshold, replacing the old model weights; otherwise, rolling back to the historical optimal weights. The trained fault prediction model is deployed to a local server. During the model running phase, the operation data stream of industrial equipment is received in real time. Fault probability prediction is performed on the operation data stream, and the corresponding fault type and its probability of occurrence are output. Deploying the trained fault prediction model to a local server includes: performing 8-bit quantization and graph optimization on the fault prediction model to reduce inference latency; encapsulating the model service in a containerized manner and providing an inference interface through a RESTful API; enabling a tensor parallel inference engine on the server's GPU accelerator card; and supporting a hot update mechanism to allow loading new model weights without interrupting the service. During the model execution phase, the system receives real-time operational data streams from industrial equipment, performs fault probability prediction on these data streams, and outputs the corresponding fault types and their probabilities of occurrence. This includes: extracting data segments from the continuously collected raw operational data streams using a fixed sliding window strategy, performing normalization processing on these segments to obtain normalized tensors; inputting the normalized tensors into the inference interface deployed on a local server, calling the fault prediction model for vectorized inference, and outputting multidimensional probability vectors; selecting candidate fault types corresponding to the maximum probability from the multidimensional probability vectors to form prediction result pairs; if the maximum probability is greater than or equal to a dynamic threshold, returning the prediction result pair and publishing it in key-value format in the OPCUA event channel. The candidate fault types include at least bearing wear, misalignment, rotor imbalance, insufficient lubrication, coil short circuit, overload operation, abnormal temperature rise, and power fluctuation.

2. The industrial equipment fault prediction method based on deep learning according to claim 1, characterized in that, The historical operating data includes at least the historical vibration signals, current and voltage signals, bearing temperature signals, speed signals, lubricating oil pressure signals, and equipment operating status of the industrial equipment under test.

3. The industrial equipment fault prediction method based on deep learning according to claim 2, characterized in that, The historical operational data is preprocessed to generate the sample dataset, including: Wavelet thresholding is used to remove random noise; Missing data were filled using spline interpolation, and outliers were removed using the 3σ criterion. Each physical quantity is normalized according to its dimensions; The normalized data is segmented and labeled with corresponding fault type labels according to the fixed sliding window strategy, and then summarized into a set to obtain the sample dataset.

4. The industrial equipment fault prediction method based on deep learning according to claim 3, characterized in that, The sample dataset is input into an industrial equipment fault prediction model built based on a deep learning algorithm for learning and training, including: The fault prediction model is a cascaded structure that uses a one-dimensional convolutional neural network bidirectional long short-term memory network attention mechanism. The cross-entropy loss function is used and iterative training is performed using the AdamW optimizer; During training, a Dropout random inactivation rate of 0.3 was set to suppress overfitting, and an early stopping strategy was used to monitor the validation set loss. After obtaining the optimal hyperparameter combination through five-fold cross-validation, the weights of the models with the highest F1 scores in the validation set are saved.

5. A deep learning-based industrial equipment fault prediction system, characterized in that, include: One or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the deep learning-based industrial equipment fault prediction method as described in any one of claims 1 to 4.

6. A computer-readable medium for storing software, characterized in that: The software includes instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the deep learning-based industrial equipment fault prediction method as described in any one of claims 1 to 4.

Images