Sludge dewatering machine health assessment and service life estimation method

Abstract

The invention discloses a health assessment and life estimation method for a sludge dewatering machine, and belongs to the field of equipment health assessment and prediction. Comprising the following steps: acquiring operating data such as temperature and vibration of each key part of equipment and working data such as cumulative processing amount from a sludge dewatering machine monitoring system, and extracting a total feature set of the operating data after preprocessing by a one-dimensional convolutional neural network and a stacked denoising auto-encoder; and then inputting a scoring model formed by two self-organizing mapping neural networks to carry out health scoring. And merging the total feature set and the working data, inputting the merged total feature set and working data into a feed-forward neural network life estimation model based on genetic algorithm optimization, and finally carrying out residual life estimation by adopting an FNN (Feedforward Neural Network) of an optimal hyper-parameter combination. The influence of multiple parameters on the health state of the sludge dewatering machine is considered, and the evaluation accuracy is improved; the intelligent level of equipment management of a current sewage treatment plant can be effectively improved, the problems of under-maintenance, over-maintenance and the like are avoided, and the purposes of reducing cost and improving efficiency are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of equipment health assessment and prediction, and more specifically relates to a health assessment and life prediction method for a sludge dewatering machine. Background Art

[0002] With the rapid progress of industrial big data and artificial intelligence technology, industrial production and manufacturing are moving towards intelligence. The various equipment or links that make up the production system are interrelated and interact with each other. Once any of the equipment or links fails, it will cause system function failure, interfere with the normal production process, and cause significant economic losses to the enterprise; it may cause personnel safety accidents and cause losses to the country and the people. Therefore, it is very necessary to use big data and artificial intelligence technology to collect the operating data of key equipment in industrial production, and conduct health assessment and life estimation of the equipment. This process involves real-time monitoring and quantitative evaluation of the health status of the equipment, as well as the prediction of the equipment life, aiming to help enterprises scientifically and rationally arrange production and maintenance plans and achieve the goal of reducing costs and increasing efficiency.

[0003] At present, equipment health management technology is gaining more and more attention. Among the key monitoring data of equipment, data such as temperature and vibration are usually closely related to the health status of the equipment, so they receive the most attention. At the same time, indicators such as cumulative working hours and cumulative processing volume may also be related to the health status and remaining life of the equipment. Although the traditional method of health diagnosis based on equipment operation data and using statistical analysis methods is easy to implement, it has great limitations, over-reliance on manual experience, and insufficient robustness and versatility. Past studies mainly rely on vibration data models to evaluate and predict equipment status. The analysis results are only related to vibration data and cannot timely reflect the impact of changes in other monitoring parameters on the health status of the equipment. In addition, existing methods usually only collect single-point data of the equipment (such as bearing data on the drive end of the equipment), and do not integrate multiple data for equipment health scoring and life prediction. Therefore, the existing equipment health assessment methods are insufficient in prediction accuracy. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a model for equipment health assessment and life prediction to improve the health score and life prediction accuracy of the current sludge dewatering machine. This method aims at the technical problem that the current health detection method is difficult to fully consider the influence of various important parameters, resulting in limited assessment accuracy. It collects operating data such as temperature and vibration of each key part of the sludge dewatering machine, as well as working data such as cumulative processing volume, and uses a neural network model to construct a model for health assessment and life prediction of the sludge dewatering machine, which realizes real-time assessment of the health of the sludge dewatering machine, helps enterprises to understand the health status of the equipment in a timely manner, and thus formulate more scientific production and maintenance plans.

[0006] A method for health assessment and life prediction of a sludge dewatering machine is characterized in that: after the operation data of the sludge dewatering machine is normalized and the key variables are determined by principal component analysis, several variables are respectively extracted with a one-dimensional convolutional neural network (1D-CNN) to extract sub-feature sets. The sub-feature sets of each variable are then fused and reduced in dimension and further extracted with a stacked denoising autoencoder (SDAE) to obtain a total feature set. The total feature set is then input into a scoring system consisting of two self-organizing map neural networks (SOM) for health assessment. The total feature set and working data are combined and input into a feedforward neural network (FNN) based on genetic algorithm (GA) optimization, and finally the FNN using the optimal hyperparameter combination is used for life prediction.

[0007] Step 1: Filter the collected sludge dewatering machine related data (including high-frequency operation data and low-frequency working data), filter key variables, and construct a health assessment data set; the operation data includes the front-end temperature of the bearing, the rear-end temperature of the bearing, the hydraulic oil temperature, vibration, speed difference, etc.; the working data includes the cumulative processing volume, the cumulative operating time, the unit processing power consumption, etc. This data set is used for the training and testing of the neural network model of the system of the present invention.

[0008] Step 2: Input various types of operation data into 1D-CNN to extract features. The data generated by each processing task is grouped together. During the data processing process, the parameters of each 1D-CNN are gradually adjusted to finally complete the processing of various types of operation data. The generated data are merged and input into SDAE for fusion and dimensionality reduction to obtain the total feature set.

[0009] Step 3: Based on the data set, a dual SOM neural network model is established to evaluate the health status of the sludge dewatering machine. The dataset label is constructed using the fusion scoring method. The first 15% of the data collected from the sludge dewatering machine after maintenance is considered to be fault-free and healthy data; the 15% of the data before maintenance is considered to be unhealthy samples. The data is normalized again, and the selected samples are used to train two SOM models respectively.

[0010] Step 4: SOM1 is trained with healthy samples. After training, the distance D1 between the sample and the healthy sample is output based on the difference between the test sample and the training sample. SOM2 is trained with unhealthy samples. After training, the distance D2 between the sample and the unhealthy sample is output based on the difference between the test sample and the training sample. Finally, D2 / (D1+D2)*100% is output as the health score of the device.

[0011] The total feature set is combined with the working data as a new sample set to estimate the life of the sludge dewatering machine. Normalization is required again before input.

[0012] Step 5: Use genetic algorithm to optimize the hyperparameters of the life prediction model based on feedforward neural network (including 5 hyperparameters, including the number of neurons in two layers, number of iterations, learning rate, and batch size). Based on the experience of repeated tests, the genetic algorithm population size is set to 40, the maximum number of iterations is set to 300, the crossover probability is set to 0.75, and the mutation probability is set to 0.05. The optimal parameter combination for different tasks may be different.

[0013] Step 6: Use the feedforward neural network model that has completed hyperparameter optimization to estimate the remaining life of the equipment. Then, you can use charts and quantitative indicators to evaluate the prediction effect of the model. Charts are used to compare the actual values ​​with the predicted values, which can relatively intuitively compare the advantages and disadvantages of the model. Commonly used quantitative indicators are MAE (mean absolute error), MAPE (mean absolute percentage error), MSE (mean square error) and RMSE (root mean square error). In most tasks, several indicators can clearly reflect the performance of the model.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: the method of the present invention realizes data-driven health assessment and life prediction of sludge dewatering machines, and considers multiple parameters of the sludge dewatering machine for fusion assessment, which is more accurate than the common single vibration parameter assessment; the dual SOM neural network extracts the healthy and unhealthy sample features of the signal at the same time, which greatly improves the scoring accuracy while ensuring the calculation speed. And there is no need to add labels manually, which greatly saves manpower; the genetic algorithm is used to optimize the hyperparameters of the life prediction model, which effectively improves the convenience of model operation and improves the accuracy. These advantages can significantly improve the level of informatization of factory management, help management to timely understand the health status of front-line equipment, reasonably arrange production and maintenance plans, avoid unnecessary waste of resources caused by over-maintenance and under-maintenance, and ultimately achieve the purpose of reducing costs and increasing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 : Structural diagram of horizontal screw dehydration centrifuge.

[0016] Figure 2 : Block diagram of sludge dewatering machine health scoring and life prediction system.

[0017] Figure 3 :1D-CNN structure diagram.

[0018] Figure 4 : DAE network diagram.

[0019] Figure 5 : SDAE network structure diagram.

[0020] Figure 6 : SOM structure diagram.

[0021] Figure 7 : Genetic algorithm optimization process.

[0022] Figure 8 : FNN structure diagram.

[0023] Fig. 9 : The overall flow chart of the method of the present invention. DETAILED DESCRIPTION

[0024] The following is a detailed description of the implementation case of the present invention. This implementation case is implemented under the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following implementation case. Step 1: Obtain the monitoring data of the sludge dewatering machine. The historical monitoring data is stored in CSV format. The tools provided by the Numpy, Scikit-learn and other program packages in Python are used to clean the data, delete zero values, replace null values, etc.; select vibration, bearing front end temperature, bearing rear end temperature, hydraulic oil temperature, speed difference, cumulative working time, cumulative processing volume and other data to form a data set.

[0025] Step 2: Use 1D-CNN to process vibration, temperature and other working data separately, and extract several eigenvalues ​​from these original sequences according to the start-stop cycle. 1D-CNN extracts features from the sequence data layer by layer through a combination of convolution layers, pooling layers and activation functions. The convolution layer captures local features by sliding the convolution kernel, the pooling layer extracts the most important information by reducing the dimension, and the activation function enhances the expression ability of the model by introducing nonlinearity. These operations together constitute the core of 1D-CNN, enabling it to effectively extract eigenvalues ​​from sequence data. The specific number of extractions depends on the specific data sample, and multiple comparisons may be required.

[0026] Formula (1) is the processing process of the convolution layer, where: K l i and b l i is the weight and bias of the i-th convolution kernel in the l-th layer; X l (j) is the jth local area in the lth layer; y (l+1) i (j) is the input of the jth neuron in the result of the i-th convolution kernel operation in the l+1th layer. The activation function selects the rectified linear unit Relu, which uses the principle of unilateral inhibition for output, which can increase the sparsity of the network and thus improve the computational efficiency of the network.

[0027] The convolutional layer is usually followed by a pooling layer, and the most commonly used one is the maximum pooling layer, which can retain the largest local features in the feature map. Its function expression is formula (2). l i (t) represents the value of the tth neuron in the ith feature vector of the lth layer, where t∈[(j-1)W+1,jW]; W represents the size of the pooling area P l+1 i (j) represents the value corresponding to the neuron in the l+1th layer after the pooling operation.

[0028] The fully connected layer is usually located after the pooling layer. The output of the last layer of the pooling layer is connected head to tail and converted into a one-dimensional vector. Then the Soft-max classifier is used to perform the fault classification task. Its operation process is shown in formula (3). In the formula: j represents a category in k categories, and the value of this category is z j express.

[0029] After all kinds of working signals are processed by 1D-CNN, SDAE performs feature fusion and dimensionality reduction to obtain the total feature set. SDAE is a stacked denoising autoencoder, which consists of two parts: encoder and decoder. The main function of the encoder is to compress the input data into a low-dimensional representation, while the decoder is responsible for reconstructing these low-dimensional representations back to the original data. In this process, the low-dimensional representation can effectively reveal the key features of the data, achieve noise removal and model simplification, and then apply the extracted features to downstream tasks. The specific operation steps are as follows: First, convert the preprocessed data samples into a format suitable for autoencoder input. Then add noise to the input samples to enhance the robustness of the model to noise. Take the most commonly used Gaussian noise as an example: generate a random number that follows a standard normal distribution (mean 0, standard deviation 1) at each data point, then multiply this random number by the specified noise level (standard deviation), and finally add it to the original data point.

[0030] Inoisy=Ioriginal+η×σ (4) Where: Inoisy—sample after adding noise; Ioriginal—original sample; η—random noise that obeys the standard normal distribution; σ—noise level (standard deviation). By adding Gaussian noise, the autoencoder can learn the processing logic of recovering useful information from noisy data without label information, thereby improving the model's robustness to noise that may be encountered in actual equipment fault diagnosis.

[0031] Where: N is the number of input data, and Xi and Xi' are the ith original input and ith reconstructed input, respectively. In unsupervised learning algorithms, the purpose of pre-training the autoencoder is to ensure that the encoder and decoder learn the low-dimensional representation of the input data and smoothly achieve the dimensionality reduction of the input data in actual operation.

[0032] Where: θ represents the model parameters, ɑ is the learning rate, Lr controls the step size of parameter update and is the gradient of the loss function Lrecon with respect to the parameter θ. In actual operation, the parameters need to be adjusted continuously in multiple iterations until the reconstruction error drops to an acceptable level or reaches a predetermined number of iterations, so that the autoencoder can learn an effective representation of the data and improve the performance of the learning model.

[0033] Feature extraction is to convert raw data into feature vectors that can represent the essential characteristics of the data. These feature vectors can be used for subsequent tasks such as classification, clustering, and regression. In the fine-tuned SDAE model, the output of the encoding layer can be used as the feature representation of the raw data. After the feature extraction is completed, the features are evaluated and selected to better understand the contribution of each feature to the task, so as to select the most representative and relevant features. Commonly used methods include feature importance analysis and correlation analysis.

[0034] The total feature set obtained is input into the dual SOM model for health scoring. SOM1 is trained with healthy samples, and SOM2 is trained with unhealthy samples. SOM is an unsupervised artificial neural network. Unlike general neural networks that are trained based on the back propagation of the loss function, it uses a competitive learning strategy, relies on competition between neurons to gradually optimize the network, and uses a neighborhood function to maintain the topological structure of the input space. This makes the two-dimensional mapping include the relative distance between data points, and adjacent samples in the input space are mapped to adjacent output neurons. Since SOM is based on unsupervised learning, this means that no manual intervention is required during the training phase (that is, no sample labels are required), and data can be clustered without knowing the category to identify features that are intrinsically related to a certain problem.

[0035] D=D2 / (D1+D2)*100% (7) Where D is the final health score, D1 is the distance between the sample and the healthy sample, and D2 is the distance between the sample and the unhealthy sample. This module compares sample features through two SOM neural networks, and the scoring accuracy is effectively improved.

[0036] The previously obtained total feature set of running data and working data (accumulated working time, accumulated processing volume, etc.) are combined into a new sample set for normalization. Subsequently, by integrating TensorFlow and DEAP libraries on Python, an automated feedforward neural network hyperparameter optimization process is implemented, the network structure and data processing process are defined, and random seeds are set to ensure repeatability. During the execution of the genetic algorithm, each generation evaluates the fitness of individuals (i.e., model performance) and continuously optimizes the population through selection, crossover, and mutation operations. Finally, when the predetermined number of iterations is reached, the algorithm outputs the hyperparameter combination that performs best on the training set and its corresponding fitness value, realizing the automated tuning of the neural network hyperparameters.

[0037] Y=g[w T f(lX+η0)+η0] This formula is the output expression of the three-layer feedforward neural network detection model, where the input vector X = [X1, X2, ..., Xm] T It is the preprocessed input data, g and f are activation functions; l is the number of output results; the output Y of the model is the final life estimation result.

[0038] The present invention improves the equipment management level of the sewage treatment plant by realizing the health assessment and life prediction algorithm of the sludge dewatering machine. Through the real-time assessment of the equipment status, unplanned downtime and excessive maintenance are avoided, which is beneficial to ensuring production safety and efficiency.

Claims

1. A method for health assessment and life prediction of a sludge dewatering machine, characterized by: After the sludge dewatering machine operation data is normalized and the principal component analysis is performed to determine the key variables, several variables are respectively extracted with a one-dimensional convolutional neural network 1D-CNN. The sub-feature sets of each variable are then fused and reduced in dimension using a stacked denoising autoencoder SDAE and further features are extracted to obtain a total feature set. The total feature set is then input into a scoring system consisting of two self-organizing map neural networks SOM for health assessment. The total feature set and working data are combined and input into a feedforward neural network FNN based on genetic algorithm GA optimization, and finally the FNN using the optimal hyperparameter combination is used to estimate the life span. The method comprises the following steps: Step 1: Screen the collected sludge dewatering machine related data, which includes high-frequency operation data and low-frequency working data; screen key variables and construct a health assessment data set. Step 2: Input various types of operating data into 1D-CNN to extract features, merge the generated data and input them into SDAE for fusion and dimensionality reduction to obtain the total feature set. Step 3: Establish a dual SOM neural network model to evaluate the health status of the sludge dewatering machine, normalize the data again, and train two SOM models with the screened samples respectively. Step 4: Use healthy samples and unhealthy samples to train two SOM neural networks respectively. Finally, use D2 / (D1+D2)*100% to output the health score of the device. Step 5: Use a genetic algorithm to optimize the hyperparameters of the life prediction model based on the feedforward neural network, wherein the hyperparameters include the number of neurons in the two layers, the number of iterations, the learning rate, and the batch size. Step 6: Use the feedforward neural network model that has completed hyperparameter optimization to estimate lifespan, and then use charts and quantitative indicators to evaluate the prediction effect of the model.

2. According to claim 1, a sludge dewatering machine health assessment and life prediction method is characterized in that: Step 2 uses a multi-parameter fusion evaluation method to construct a dataset label. The evaluation is based on a single vibration data, by considering multiple parameters such as vibration, bearing front temperature, bearing rear temperature, hydraulic oil temperature, and speed difference. All kinds of working signals are processed by 1D-CNN, and then feature fusion and dimensionality reduction are performed by SDAE to obtain the total feature set for subsequent work.

3. The method for health assessment and life prediction of a sludge dewatering machine according to claim 1 is characterized in that: Step 3: Input the obtained total feature set into the dual SOM neural network model for health scoring. The dual SOM neural network model includes SOM1 and SOM2; SOM1 is trained with healthy samples, and SOM2 is trained with unhealthy samples. The two SOM neural networks compare the healthy and unhealthy features of the samples respectively to improve the model accuracy.

4. The method for health assessment and life prediction of a sludge dewatering machine according to claim 1 is characterized in that: Step 5: The previously obtained total feature set of running data and working data are combined into a new sample set for normalization. Genetic algorithm is used to optimize the feedforward neural network to obtain the optimal hyperparameter combination.

5. The method for health assessment and life prediction of a sludge dewatering machine according to claim 1 is characterized in that: Step 6 uses the feedforward neural network model that has completed hyperparameter optimization to estimate lifespan, and then uses charts and quantitative indicators to evaluate the prediction effect of the model.

Images