Compressor mixed fault prediction method and system based on LSH and LSTM

Abstract

The invention relates to the technical field of equipment fault prediction, in particular to a compressor mixed fault prediction method and system based on LSH and LSTM, and the method comprises the steps: data collection, preprocessing, LSTM model construction and training, incremental learning and fault prediction. Data preprocessing is combined with median filtering and wavelet denoising, and similar data clusters are formed through LSH processing; constructing an attention layer-containing LSTM network, and highlighting key features through attention scores; when the state of the compressor changes, new data are matched with corresponding data clusters through LSH, and only correlation model parameters are finely adjusted; the system comprises a data acquisition module, a preprocessing module, a model training module, an incremental learning module and a fault prediction module. According to the invention, data processing efficiency and model adaptability are improved, real-time performance and accuracy of fault prediction are enhanced, and equipment maintenance cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment failure prediction, in particular to a compressor hybrid failure prediction method and system based on LSH and LSTM. BACKGROUND

[0002] In modern industrial production, compressor-type equipment, as a key dynamic device, is widely used in petroleum chemical industry, power and many other fields. Its stable operation is crucial to ensure production continuity, improve production efficiency and product quality.

[0003] Traditional compressor failure prediction methods are mainly based on physical models and empirical rules. However, the operation process of the compressor is complex and affected by many factors such as temperature, pressure, vibration, etc., making it difficult to establish an accurate physical model. Moreover, with the increasing complexity and variability of industrial environments, fixed rule-based methods are difficult to adapt to changes in equipment operating conditions under different working conditions.

[0004] In recent years, data-driven failure prediction methods have gradually emerged. In patent CN110503133A, a compressor fault diagnosis method based on deep belief network (DBN) is proposed, which improves model performance through a combination of unsupervised pre-training and supervised fine-tuning. However, this method has two significant drawbacks: first, it does not consider the dynamic characteristics of time series data, making it difficult to capture the gradual change process of equipment state; second, the fixed structure of the model cannot adapt to changes in working conditions, and the entire network needs to be retrained every time the working condition is adjusted, resulting in high computational cost.

[0005] Another patent CN119882650A uses convolutional neural network (CNN) to process vibration signal image features, which improves feature extraction capability, but has limitations: 1) only for single type sensor data (vibration signal), without fusion of multi-source heterogeneous data; 2) lacks incremental learning mechanism, model performance significantly decreases when equipment operating parameters drift. Experiments show that when the compressor load changes more than 15%, the prediction accuracy of this model will decrease by 8-12 percentage points.

[0006] The current mainstream long short-term memory network (LSTM) is widely used due to its strong processing capability for time series data. LSTM can learn long-term dependencies in equipment operation data and effectively capture the trend of changes in equipment operating conditions. However, LSTM models still face three challenges in practical industrial applications:

[0007] First, the working state of equipment in industrial environment changes frequently. For example, the operating conditions of the compressor will be different in different seasons or under different production tasks. When the working state changes, the traditional LSTM model needs to be retrained to adapt to the new situation. However, retraining the entire model not only consumes time and effort (such as the method of CN110503133A which takes 4-6 hours to complete a retraining), but also in the actual production process, the equipment operation cannot be interrupted for a long time to train the model, which limits its application in industrial scenarios with high real-time requirements.

[0008] Second, with the continuous growth of industrial data, processing massive data requires higher computing resources. The traditional LSTM model has a slow training speed when processing large-scale data, making it difficult to meet the real-time monitoring requirements. For example, in some large chemical companies, the compressor generates thousands of data per day, and the image processing method in CN119882650A will make the data dimension expand by 3-5 times, further increasing the computational burden.

[0009] Third, there are various noise interferences in industrial data. Sensor failure can cause data deviation (such as pulse noise in vibration signals), and environmental interference can introduce Gaussian noise (such as temperature signal fluctuations). Existing methods such as CN110503133A only use mean filtering to denoise, which has limited effect on suppressing pulse noise; CN119882650A uses frequency domain filtering, but it will lose the transient fault features in the signal. These noises can significantly reduce the prediction accuracy of the model and can reduce the fault recognition rate by 15%-20% in testing.

[0010] In summary, the existing technology has three core problems: 1) poor model adaptability, unable to quickly respond to changes in operating conditions (such as CN110503133A); 2) low efficiency in processing large data (such as the image processing method of CN119882650A); 3) weak anti-noise ability. This makes existing methods based on DBN, CNN or LSTM difficult to meet the real-time and accurate fault prediction requirements in industrial scenarios, and there is an urgent need to develop new hybrid prediction models. SUMMARY

[0011] The purpose of the present application is to provide a compressor hybrid fault prediction method and system based on LSH and LSTM, to solve the problems of poor model adaptability, low efficiency in processing large data, and sensitivity to noise data faced by existing LSTM-based fault prediction methods when applied in industrial environments.

[0012] To achieve the above purpose, the present application provides the following technical solutions:

[0013] A compressor hybrid fault prediction method based on LSH and LSTM, comprising the following steps:

[0014] S10, data collection: real-time collection of multi-dimensional time series data during the operation of the compressor through sensors installed on the compressor to form an original data set;

[0015] S20, data preprocessing: sequentially performing denoising processing and LSH processing on the original data set to obtain preprocessed data; wherein the denoising processing adopts a method combining median filtering and wavelet denoising, and the LSH processing maps the denoised data to corresponding hash buckets through a hash function to determine similar data clusters;

[0016] S30, LSTM model construction and training: constructing an LSTM network model containing an attention layer, and training the model using a training set in the preprocessed data set, wherein the attention layer highlights key features by calculating attention scores of time step hidden states;

[0017] S40, incremental learning and model updating: when the working state of the compressor changes, the new data is matched to the corresponding similar data cluster after LSH processing, and only the LSTM model parameters associated with the data cluster are fine-tuned;

[0018] S50, fault prediction: using the trained and updated model to predict real-time preprocessed data, and outputting the operating state and fault type of the compressor.

[0019] As a preferred embodiment, the specific steps of the LSH processing in the data preprocessing include:

[0020] S21: converting the denoised data into a fixed-dimensional high-dimensional feature vector through principal component analysis;

[0021] S22: selecting a hash function group based on a random projection to construct a hash function group;

[0022] S23: mapping the feature vector to the hash bucket through the hash function group to form multiple similar data clusters;

[0023] S24: after feature extraction and hash mapping, the new data is quickly located to the corresponding similar data cluster.

[0024] Converting the fixed-dimensional feature vector through PCA reduces redundancy, the hash function group based on random projection ensures the accuracy of similar data mapping, and hash mapping forms similar data clusters and realizes the quick positioning of new data, which not only improves the processing efficiency of massive industrial data and ensures the accurate clustering of similar working condition data, but also provides a clear data basis for the incremental learning of the subsequent LSTM model, so that the model can quickly adapt to the changes in the working condition of the compressor, and the real-time and adaptability of fault prediction is enhanced.

[0025] As a preferred embodiment, the LSTM network model containing an attention layer includes:

[0026] Input layer: receives the LSH processed feature vector, the dimension is consistent with that of the feature vector;

[0027] At least two LSTM hidden layers: learn the long-term dependence of time series through input gate, forget gate and output gate;

[0028] Attention layer: calculate the attention score of each time step hidden state, in order to measure the importance of each time step hidden state, a scoring function is used to calculate the attention score a t , assuming that the LSTM hidden layer output is H = [h1, h2, …, hT], where ht is the hidden state at time t, T is the time step, the hidden state h t is mapped to a new dimensional space through a learnable linear transformation, and et is obtained, the formula is:

[0029] ;

[0030] Where, h t is the hidden state at time t, W e is the weight matrix, b e is the bias vector, T is the time step, and a t represents the relative importance of the hidden state h t at time t, the greater the value, the more important the information at this moment to the current prediction task. The sum of a t of all time steps is 1;

[0031] Output layer: the context vector output by the attention layer is spliced with the last time hidden state, and the fault prediction result is output through the fully connected layer.

[0032] By specifying the specific composition and connection mode of the input layer, at least two LSTM hidden layers, attention layer and output layer, especially the quantitative calculation of the importance of the time step hidden state (at), the model can accurately focus on the key features and time nodes for fault prediction. For example, in the compressor vibration data, some abnormal fluctuations at certain moments may be a precursor to failure, and the attention mechanism can give them higher weights, effectively improving the relevance and accuracy of fault feature extraction, reducing irrelevant information interference, and thus improving the accuracy of fault prediction.

[0033] As a preferred, the incremental learning strategy specifically comprises:

[0034] S41: when the new data matches the similar data cluster, select a preset number of new samples from the cluster;

[0035] S42: fix most of the parameters in the trained model, and only adjust the weight parameters related to the features of the new data;

[0036] S43: Fine-tuning the selected parameters using the cross-entropy loss function and the Adam optimizer until the performance of the validation set meets the standard.

[0037] Through the step-by-step strategy of "selecting new samples, fixing most parameters, and fine-tuning related parameters", combined with the cross-entropy loss function and the Adam optimizer, the model can quickly adapt to changes in equipment working conditions. Without retraining the entire model, only the weight parameters related to new data features are fine-tuned, significantly reducing the consumption of computing resources and the time of model updating. For example, when the compressor's pressure feature changes due to production task adjustment, fine-tuning the parameters related to the pressure feature can quickly adapt the model to the new working condition, ensuring real-time performance while maintaining stable prediction performance.

[0038] As a preferred, the specific steps of the denoising processing are:

[0039] S51 Median filtering: setting the window size to 5, performing sliding window processing on one-dimensional time series data, and replacing the center value of the window with the middle value in the window;

[0040] S52 Wavelet denoising: using db4 wavelet basis function for 3-level decomposition, and reconstructing the signal after threshold processing of each sub-band coefficient.

[0041] Median filtering (window size 5) specifically removes impulse noise in one-dimensional time series such as vibration, and wavelet denoising (db4 wavelet basis + 3-level decomposition) reduces the influence of Gaussian noise while preserving the main features of the signal. The two steps form a complementary denoising mechanism. Both sudden data changes (impulse noise) caused by sensor failures during compressor operation and persistent noise (Gaussian noise) caused by environmental interference can be effectively suppressed, significantly improving data quality, providing reliable input for subsequent LSH processing and model training, and enhancing the model's anti-interference ability in complex industrial environments.

[0042] On the other hand, the present application also provides a compressor hybrid fault prediction system based on LSH and LSTM, comprising:

[0043] Data acquisition module: composed of vibration, temperature, and pressure sensors installed on the compressor, used for real-time acquisition of multi-dimensional time series data;

[0044] Data preprocessing module: containing a denoising unit and an LSH processing unit, the denoising unit uses median filtering and wavelet denoising methods to process raw data, and the LSH processing unit maps the denoised data into similar data clusters;

[0045] Model training module: building an LSTM network with attention mechanism, training the model with training set data and optimizing through validation set;

[0046] Incremental learning module: locate similar data clusters according to LSH processing results, and update model correlation parameters incrementally;

[0047] Fault prediction module: receive real-time preprocessed data, and output compressor operating state and fault warning information.

[0048] As preferred, the LSH processing unit comprises:

[0049] Feature extraction subunit: convert multi-dimensional data into fixed-dimensional feature vectors through principal component analysis;

[0050] Hash function subunit: construct multiple sets of hash functions using a random projection hash function family;

[0051] Mapping subunit: map feature vectors to hash buckets through hash functions to realize fast clustering of similar data.

[0052] The feature extraction subunit reduces the dimensionality of multi-dimensional data to fixed-dimensional feature vectors through PCA, reducing data redundancy; the hash function subunit uses a random projection hash function family to ensure the accuracy of similar data mapping; the mapping subunit realizes fast clustering of feature vectors to hash buckets. The three work together to enable new data to quickly locate similar data clusters in massive historical data, for example, in compressor data running in different seasons, similar working condition historical data groups can be quickly matched, greatly shortening the data retrieval time, providing an efficient data basis for incremental learning, and improving the efficiency of model processing large-scale industrial data.

[0053] As preferred, the LSTM network in the model training module comprises:

[0054] Input layer: receive feature vectors with a dimension of 50;

[0055] 3 LSTM hidden layers: each hidden layer contains 64 LSTM units;

[0056] Attention layer: calculate time step weights through learnable parameters;

[0057] Output layer: output normal, minor fault, and serious fault probability distributions through a softmax function.

[0058] The input layer receives a 50-dimensional feature vector, balancing feature integrity and computational complexity; the three LSTM hidden layers with 64 units each enhance the learning ability of long-term dependencies in time series and can capture subtle trends in compressor operating conditions; the attention layer dynamically adjusts the time step weight through learnable parameters; the output layer outputs the probability distribution of three fault types through the softmax function, clearly distinguishing the severity of the faults. This structure ensures the model's expressive power while accurately classifying fault types, making it easier for maintenance personnel to take targeted measures based on the prediction results, such as planned maintenance for "minor faults" and emergency shutdown for "severe faults," improving the effectiveness of maintenance decisions.

[0059] As a preferred embodiment, the working mechanism of the incremental learning module is as follows:

[0060] Receiving the similar data cluster identification output by the LSH processing unit;

[0061] S61: Extracting feature samples related to the cluster from newly collected data;

[0062] S62: Freezing the parameters associated with other data clusters in the model and only fine-tuning the input layer to hidden layer weights corresponding to the current cluster;

[0063] S63: Updating the parameters through backpropagation until the validation set accuracy improves to the preset threshold.

[0064] Through the mechanism of "extracting related samples → freezing irrelevant parameters → fine-tuning target weights → backpropagation update", combined with the validation set accuracy threshold control, the model can adapt to new working conditions without destroying existing training results and accurately optimize parameters related to the current data cluster. For example, when the compressor's temperature feature is abnormal due to changes in raw materials, only the input layer to hidden layer weights related to the temperature feature are fine-tuned, allowing the model to quickly adapt and avoid overfitting, while the threshold control ensures the stability of the updated model, further improving the model's adaptability and reliability in dynamic industrial environments.

[0065] As a preferred embodiment, the denoising unit of the data preprocessing module performs the following operations:

[0066] S71: Applying median filter with a window size of 5 to remove impulse noise from the vibration signal;

[0067] S72: Using db4 wavelet basis for 3-level decomposition to denoise temperature and pressure signals;

[0068] S73: Outputting the denoised multi-dimensional time series data to the LSH processing unit.

[0069] Different types of signals (vibration signals, temperature / pressure signals) adopt differentiated denoising strategies: vibration signals are susceptible to pulse interference, and median filtering with a window size of 5 can effectively remove them; temperature and pressure signals are greatly affected by Gaussian noise, and 3-layer decomposition with db4 wavelet basis can accurately denoise. This targeted processing not only ensures the denoising effect of various signals, but also maximizes the preservation of effective features, such as fault impact features in vibration signals and gradual trends in temperature signals, which can be completely preserved to provide high-quality data for subsequent LSH processing and model training, reduce prediction bias caused by noise, and improve the overall robustness of the model.

[0070] Compared with the prior art, the beneficial effects of the present application are:

[0071] 1. Quickly adapt to changes in working conditions: By introducing the LSH algorithm for data preprocessing, new data can be quickly mapped to similar data clusters, and only the LSTM model parameters corresponding to the similar data clusters are fine-tuned, realizing the rapid adaptation of the model to changes in the working conditions of the equipment. Compared with the traditional method of retraining the entire model, it greatly saves time and computing resources. For example, in the case where the traditional method may take several hours or even several days to retrain the model, the method of the present application can complete the adjustment of the model in only tens of minutes, meeting the real-time requirements of industrial production and ensuring that the equipment can still be accurately predicted in a timely manner when the working conditions change.

[0072] 2. Improve prediction accuracy: The traditional LSTM network structure is optimized to increase the attention mechanism, so that the model can pay more attention to key features related to fault prediction. Whether in normal working conditions or when the working conditions change, the model can more accurately extract fault features, thereby improving the accuracy of fault prediction. Experiments show that, under the same data set, the model of the present application improves the fault prediction accuracy by 10%-15% compared with the unoptimized LSTM model, effectively reducing the occurrence of false positives and false negatives, and providing more reliable protection for the stable operation of the equipment.

[0073] 3. Efficiently process massive data: The application of LSH algorithm enables quick positioning of similar data when processing massive industrial data, avoiding repeated processing of all data. In the face of growing compressor operation data, the data processing efficiency is significantly improved, and the computational burden of the model is reduced. For example, in the scenario of processing thousands of data per day, the processing time of the model is shortened by 30%-50% compared with the traditional method, enabling the model to complete the prediction task in a shorter time and meeting the demand for real-time data processing in industrial production.

[0074] 4. Enhanced Noise Resistance: Median filtering and wavelet denoising are employed in the data preprocessing stage to effectively remove noise and outliers from the data. The denoised data is then input into the model, reducing the interference of noise on the model's prediction results and improving the model's robustness to noisy data. Even when occasional sensor malfunctions cause data deviations, the model of this invention can still maintain high prediction accuracy, enhancing the model's reliability in complex industrial environments.

[0075] 5. Reduced Maintenance Costs: Accurate fault prediction and timely maintenance decision recommendations enable maintenance personnel to prepare for repairs in advance, rationally allocate maintenance resources, and avoid production interruptions and unnecessary maintenance costs caused by sudden equipment failures. Simultaneously, real-time monitoring of equipment operating status enables preventative maintenance, extending equipment lifespan and further reducing enterprise equipment maintenance costs. For example, through the application of this invention's system, unplanned downtime of compressor equipment can be reduced by 30%-40%, bringing economic benefits to the enterprise. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0077] Figure 1 This is a block diagram of the compressor hybrid fault prediction system based on LSH and LSTM of the present invention. Detailed Implementation

[0078] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0079] Example 1: Fault Prediction Method Applicable to Reciprocating Compressors

[0080] 1.1 Data Acquisition

[0081] Data acquisition is performed using multiple types of sensors installed on the reciprocating compressor, including:

[0082] Vibration sensor: installed in the compressor crankcase and cylinder head, with a sampling frequency of 10kHz, to collect vibration acceleration signals in the X, Y, and Z directions;

[0083] Temperature sensor: installed at the intake and exhaust valves and bearing housing, with a sampling frequency of 1kHz, to collect temperature signals;

[0084] Pressure sensor: installed in the intake and exhaust pipeline, sampling frequency is 500 Hz, collect the intake pressure and exhaust pressure signal.

[0085] Continuous collection for 30 days, forming a raw data set containing normal operation, gas valve leakage (minor fault), bearing wear (serious fault) three states, the total amount of data is 500,000.

[0086] 1.2 Data preprocessing

[0087] De-noising processing:

[0088] The median filter with window size of 5 is performed on the vibration signal to remove impulse noise;

[0089] The temperature and pressure signals are decomposed by db4 wavelet basis for 3 layers, and the soft threshold is set to 0.02. The de-noised signal is reconstructed.

[0090] LSH processing:

[0091] The de-noised data is converted into a 50-dimensional feature vector by principal component analysis;

[0092] Select 10 random projection hash functions to construct a hash function group, and each hash function has a dimension of 10;

[0093] Map the feature vector to 1024 hash buckets to form 128 similar data clusters.

[0094] 1.3 LSTM model construction and training

[0095] Model structure:

[0096] Input layer: receive 50-dimensional feature vector;

[0097] 3 LSTM hidden layers: each contains 64 LSTM units, and the activation function is tanh;

[0098] Attention layer: calculate the attention score of each time step hidden state, the formula is: ,

[0099] Where (W e ) is a 64x64 weight matrix, b e is a 64-dimensional bias vector, and the time step T=30;

[0100] Output layer: concatenate the context vector and the last time hidden state (dimension 128), and output the probability of 3 classes of faults (normal, minor fault, serious fault) through the fully connected layer.

[0101] Training parameters: training set, validation set, test set ratio 7:2:1; batch size 32, iteration number 50, learning rate 0.001; cross-entropy loss function and Adam optimizer are used.

[0102] 1.4 Incremental learning and model updating

[0103] When the compressor load is increased from 80% to 100%, the new data is matched to the similar data cluster numbered #36 after LSH processing:

[0104] 2000 new samples are selected from the cluster;

[0105] Freeze the parameters associated with other data clusters (accounting for 85%), and only fine-tune the weights of the input layer to the first LSTM hidden layer;

[0106] Fine-tune the iteration number 10, learning rate 0.0001, and stop when the validation set accuracy is increased from 92.3% to 94.1%.

[0107] 1.5 Fault prediction results

[0108] The test set prediction accuracy is 95.7%, of which the slight fault recognition rate is 94.2% and the serious fault recognition rate is 97.3%, and the average response time is 0.32 seconds.

[0109] Example 2: Fault prediction method suitable for centrifugal compressor

[0110] 2.1 Data acquisition

[0111] For centrifugal compressors, the sensor configuration is as follows:

[0112] Vibration sensor: installed at impeller, gear box, sampling frequency 20kHz, vibration velocity signal acquisition;

[0113] Temperature sensor: installed at motor winding, lubricating oil circuit, sampling frequency 2kHz;

[0114] Flow sensor: installed at inlet and outlet pipe, sampling frequency 1kHz, gas flow signal acquisition.

[0115] Collect original data containing normal operation, impeller fouling (slight fault), shaft misalignment (serious fault), a total of 300,000.

[0116] 2.2 Data preprocessing

[0117] Denoising: median filter window 5 for vibration signal, db4 wavelet 3-layer denoising (hard threshold 0.03) for temperature / flow signal;

[0118] LSH processing: principal component analysis is converted into 60-dimensional feature vectors, 12 random projection hash functions, mapped to 2048 hash buckets, forming 156 data clusters.

[0119] 2.3 LSTM model construction and training Model structure: input layer 60 dimensions, 3 LSTM hidden layers (64 units), attention layer time step T=40, output layer 3 class fault probability; Training parameters: batch size 64, iteration number 40, learning rate 0.001.

[0120] 2.4 Incremental learning and model updating

[0121] When the medium temperature rises from 30℃ to 50℃, the new data matches the #89 data cluster, and 1500 samples are selected for fine-tuning, and the verification set accuracy rate is improved from 91.5% to 93.8%.

[0122] 2.5 Fault prediction results

[0123] Test set accuracy 94.9%, slight fault recognition rate 93.5%, severe fault recognition rate 96.7%, average response time 0.28 seconds.

[0124] Example 3: Fault prediction method suitable for screw compressors

[0125] 3.1 Data acquisition

[0126] Screw compressor sensor configuration:

[0127] Vibration sensor: installed on the sun and rotor bearing seat, sampling frequency 15kHz;

[0128] Temperature sensor: installed in the compression chamber and cooling water circuit, sampling frequency 1.5kHz;

[0129] Current sensor: collect motor working current, sampling frequency 500Hz.

[0130] Collect 400,000 original data containing normal operation, rotor gap increase (slight fault), and motor overload (severe fault).

[0131] 3.2 Data preprocessing

[0132] Denoising: median filter window 5 for vibration signal, db4 wavelet 3-layer denoising (soft threshold 0.025) for temperature / current signal;

[0133] LSH processing: principal component analysis is converted into 60-dimensional feature vectors, 12 random projection hash functions, mapped to 2048 hash buckets, forming 156 data clusters.

[0134] 3.3 LSTM model construction and training Model structure: input layer 55 dimensions, 3 LSTM hidden layers (64 units), attention layer time step T=35; training parameters: batch size 48, iteration number 45, learning rate 0.001.

[0135] 3.4 Incremental learning and model updating

[0136] When the working pressure increases from 0.8 MPa to 1.0 MPa, the new data matches the #52 data cluster, and 1800 samples are selected for fine-tuning. The verification set accuracy increases from 92.1% to 94.5%.

[0137] 3.5 Fault prediction results

[0138] The test set accuracy is 95.3%, the slight fault recognition rate is 93.9%, the severe fault recognition rate is 97.0%, and the average response time is 0.30 seconds.

[0139] Comparative Example 1: Fault prediction method without LSH processing

[0140] 4.1 Method difference

[0141] Remove the LSH processing step in data preprocessing, directly input the denoised data into the LSTM model, and the remaining steps are the same as Example 1.

[0142] 4.2 Experimental results

[0143] The test set accuracy is 88.6%, which is 7.1% lower than that of Example 1;

[0144] The slight fault recognition rate is 85.3%, which decreases by 8.9%;

[0145] The model training time increases by 3 times, and the incremental learning response time increases by 2.5 times.

[0146] Comparative Example 2: LSTM fault prediction method without attention mechanism

[0147] 5.1 Method difference

[0148] Remove the attention layer in the LSTM model, directly input the last time hidden state into the output layer, and the remaining steps are the same as Example 1.

[0149] 5.2 Experimental results

[0150] The test set accuracy is 91.2%, which is 4.5% lower than that of Example 1;

[0151] The severe fault recognition rate is 92.5%, which decreases by 4.8%;

[0152] The ability to capture long-term dependence features is weakened, and the fault early warning time is shortened by 20%.

[0153] 5.3 Result Analysis

[0154] Attention mechanism can highlight the feature contribution of key time steps, and the missing post-model is difficult to focus on important state changes, resulting in decreased sensitivity and accuracy of fault prediction.

[0155] The above three embodiments and two groups of comparative examples of key parameters and results comparison table are as follows:

[0156]

[0157] The application of the hybrid fault prediction model based on LSH and LSTM to the intelligent fault prediction and maintenance system of compressor type equipment brings many beneficial effects:

[0158] Quickly adapt to changes in working conditions: by introducing LSH algorithm for data preprocessing, new data can be quickly mapped to similar data clusters, and only the LSTM model parameters corresponding to the similar data clusters are fine-tuned, realizing the quick adaptation of the model to the changes in the working condition of the equipment. Compared with the traditional method of retraining the entire model, it greatly saves time and computing resources.

[0159] Improve prediction accuracy: optimize the traditional LSTM network structure, increase the attention mechanism, and make the model pay more attention to key features related to fault prediction. Whether in normal working conditions or in working condition changes, it can more accurately extract fault features, thereby improving the accuracy of fault prediction. Experiments show that under the same data set, the fault prediction accuracy of the model of the application is improved by 10%-15% compared with the unoptimized LSTM model, effectively reducing the occurrence of false positives and false negatives, and providing more reliable protection for the stable operation of the equipment.

[0160] Efficiently process massive data: the application of LSH algorithm enables quick positioning of similar data when processing massive industrial data, avoiding repeated processing of all data. In the face of growing compressor operation data, it significantly improves data processing efficiency and reduces the computational burden of the model. For example, in the scenario of processing thousands of data per day, the processing time of the model is shortened by 30%-50% compared with the traditional method, enabling the model to complete the prediction task in a shorter time and meeting the real-time data processing needs of industrial production.

[0161] Enhance the anti-noise ability: in the data preprocessing stage, median filter and wavelet denoising method is adopted, effectively remove the noise and abnormal value in the data. After denoising processing data input to the model, reduce the noise to the model prediction result of interference, improve the robustness of the model to noise data. Even in the case of sensor occasionally malfunction caused by data deviation, the model of the application can still maintain high prediction accuracy, enhance the reliability of the model in complex industrial environment.

[0162] Reduce maintenance cost: accurate fault prediction and timely maintenance decision suggestion, make the maintenance personnel can make good maintenance preparation in advance, reasonable arrangement of maintenance resources, avoid the production interruption and unnecessary maintenance cost caused by equipment sudden failure. At the same time, through real-time monitoring equipment running state, can realize preventive maintenance, prolong the service life of equipment, further reduce the enterprise equipment maintenance cost. For example, through the system application of the application, the compressor equipment unplanned downtime can be reduced by 30%-40%, bring economic benefits for enterprises.

[0163] The above shows and describes the basic principle, main features and advantages of the present application. The skilled in the art should understand that the application is not limited to the above examples, the above examples and the description described in this application is only the preferred example, and is not intended to limit the application, without departing from the spirit and scope of the application, the application will have various changes and improvements, these changes and improvements all fall within the scope of the application claimed. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A compressor hybrid fault prediction method based on LSH and LSTM, characterized by: The method comprises the following steps: S10, data acquisition: real-time acquisition of multi-dimensional time series data in the operation process of the compressor through sensors installed on the compressor to form an original data set; S20, data preprocessing: sequentially performing denoising processing and LSH processing on the original data set to obtain preprocessed data; wherein the denoising processing adopts a method combining median filtering and wavelet denoising, and the LSH processing maps the denoised data to corresponding hash buckets through a hash function to determine similar data clusters; S30, LSTM model construction and training: constructing an LSTM network model containing an attention layer, and training the model by using a training set in the preprocessed data set, wherein the attention layer highlights key features by calculating attention scores of time step hidden states; S40, incremental learning and model updating: when the working state of the compressor changes, the new data is matched to the corresponding similar data cluster after LSH processing, and only the LSTM model parameters associated with the data cluster are fine-tuned; S50, fault prediction: using the trained and updated model to predict real-time preprocessed data, and outputting the operating state and fault type of the compressor.

2. The LSH and LSTM based compressor hybrid fault prediction method according to claim 1, characterized in that: The specific steps of the LSH processing in the data preprocessing include: S21: converting the denoised data into a fixed-dimensional high-dimensional feature vector through principal component analysis; S22: selecting a hash function family based on random projection to construct a hash function group; S23: mapping the feature vector to the hash bucket through the hash function group to form multiple similar data clusters; S24: after feature extraction and hash mapping, the new data is quickly located to the corresponding similar data cluster. 3.The LSH and LSTM based compressor hybrid fault prediction method of claim 1, wherein: The LSTM network model containing the attention layer comprises: An input layer: receiving the LSH-processed feature vector, the dimension of which is consistent with that of the feature vector; At least two LSTM hidden layers: learning the long-term dependence of the time series through input gates, forgetting gates and output gates; Attention layer: calculate the attention score of each time step hidden state, in order to measure the importance of each time step hidden state, use a scoring function to calculate the attention score α t , assuming that the LSTM hidden layer output is H=[h1,h2……,h T ], where h t is the hidden state at time t, T is the time step, the hidden state h t is mapped to a new dimensional space through a learnable linear transformation, and e t is obtained, the formula is: ； where W e is a weight matrix, b e is a bias vector; An output layer: concatenating the context vector output by the attention layer with the last time hidden state, and outputting the fault prediction result through a fully connected layer.

4. The LSH and LSTM based compressor hybrid fault prediction method of claim 1, wherein: The incremental learning strategy specifically includes: S41: selecting a preset number of new samples from the similar data cluster after the new data is matched to the similar data cluster; S42: fixing most of the parameters in the trained model, and only adjusting the weight parameters related to the features of the new data; S43: fine-tuning the selected parameters using a cross-entropy loss function and an Adam optimizer until the performance of the validation set meets the standard.

5. The LSH and LSTM based compressor hybrid fault prediction method of claim 1, wherein: The specific steps of the denoising processing are: S51, median filtering: setting the window size to 5, performing sliding window processing on one-dimensional time series data, and replacing the center value in the window with the middle value in the window; S52, wavelet denoising: using a db4 wavelet basis function for 3-layer decomposition, and reconstructing the signal after threshold processing of each sub-band coefficient.

6. A compressor hybrid fault prediction system based on LSH and LSTM, comprising the compressor hybrid fault prediction method based on LSH and LSTM of any one of claims 1-5, characterized in that, It comprises: A data acquisition module: composed of vibration, temperature and pressure sensors installed on the compressor, for real-time acquisition of multi-dimensional time series data; The data preprocessing module comprises a denoising unit and an LSH processing unit, the denoising unit adopts a median filter and a wavelet denoising method to process original data, and the LSH processing unit maps the denoised data into similar data clusters; The model training module constructs an LSTM network with an attention mechanism, trains the model using training set data, and optimizes it through a validation set; The incremental learning module locates similar data clusters according to the LSH processing result, and performs incremental update on the model associated parameters; The fault prediction module receives real-time preprocessed data, and outputs compressor operating state and fault warning information.

7. The LSH and LSTM based compressor hybrid fault prediction system of claim 6, wherein: The LSH processing unit comprises: A feature extraction subunit that converts multi-dimensional data into a fixed-dimensional feature vector through principal component analysis; A hash function subunit that adopts a random projection hash function family to construct multiple hash functions; A mapping subunit that maps the feature vector to a hash bucket through the hash function, realizing fast clustering of similar data.

8. The LSH and LSTM based compressor hybrid fault prediction system of claim 6, wherein: The LSTM network in the model training module comprises: An input layer that receives a feature vector with a dimension of 50; Three LSTM hidden layers, each containing 64 LSTM units; An attention layer that calculates time step weights through learnable parameters; An output layer that outputs normal, slight fault and severe fault probability distributions through a softmax function.

9. The LSH and LSTM based compressor hybrid fault prediction system of claim 6, wherein: The working mechanism of the incremental learning module is as follows: Receiving the similar data cluster identifier output by the LSH processing unit; S61: Extracting feature samples related to the cluster from newly collected data; S62: Freezing the parameters associated with other data clusters in the model, and only fine-tuning the input layer to the hidden layer weights corresponding to the current cluster; S63: Updating the parameters through backpropagation until the validation set accuracy is improved to a preset threshold.

10. The LSH and LSTM based compressor hybrid fault prediction system of claim 6, wherein: The denoising unit of the data preprocessing module performs the following operations: S71: Removing impulse noise from the vibration signal using a window size of 5 median filter; S72: Denoising temperature and pressure signals using db4 wavelet basis 3-layer decomposition; S73: Outputting the denoised multi-dimensional time series data to the LSH processing unit.

Images