Hydraulic system fault prediction method based on PSO-BP neural network algorithm

Abstract

The invention provides a hydraulic system fault prediction method based on a PSO-BP neural network algorithm, and the method comprises the steps: firstly analyzing the mechanism of a fault of a hydraulic system part, employing a BP neural network, combing the related data of the hydraulic system and the historical fault data of the part, and building a prediction model; and performing fault prediction or risk grade evaluation on the operation state of the hydraulic system by using the prediction model. In order to improve the prediction performance of the BP neural network, a particle swarm optimization algorithm is adopted to optimize the weight and threshold value of the model, and the prediction result is compared with the original prediction result to verify the effectiveness of the algorithm.

Description

Technical Field

[0001] The invention relates to the technical field of turbine speed regulation system monitoring, and in particular to a hydraulic system fault prediction method based on a PSO-BP neural network algorithm. Background Art

[0002] The turbine speed control system consists of the electrical part of the speed governor and the mechanical hydraulic part. The PLC in the speed governor control cabinet controls the operation of the entire hydraulic system, maintains the pressure, oil temperature, liquid level and other characteristic quantities of the hydraulic system within the normal working range, and provides stable and reliable operating oil pressure for the mechanical actuator of the speed governor. The reliability of the hydraulic system is crucial, but it is an independent and real-time individual, so it is difficult to realize the online evaluation and diagnosis of its component status.

[0003] Common faults in hydraulic systems include: sudden faults, wear faults (this type of fault can be predicted early through testing, that is, hydraulic system faults have a process of formation and development), oil leakage (caused by aging of seals, wear and loosening of seals, etc., divided into internal leakage and external leakage, manifested as liquid loss and liquid foam), misuse faults (load, pressure, flow exceeding the rated value: for example, load use causes a failure that burns the motor), inherent weakness faults, and other common faults (oil contamination, hydraulic system overheating, insufficient hydraulic system pressure, noise and abnormal vibration, slow movement or no movement).

[0004] Common fault diagnosis methods include: pressure test, sensory diagnosis (using hearing, vision, touch and other sensory means to determine whether there is any abnormality in the hydraulic system), actuator speed test (by testing the speed of the actuator to reflect the changes of certain components in the hydraulic system); Wavelet packet support vector machine combination: Through wavelet packet analysis of hydraulic systems (such as related signals of hydraulic pumps), a fault prediction model of hydraulic systems (such as hydraulic pumps) combined with wavelet packet support vector machines is established. This method can capture weak fault characteristics in hydraulic systems and improve the accuracy of fault prediction.

[0005] However, the existing prediction methods have the following disadvantages: (1) High computational complexity: Support vector machines (SVM) have high computational complexity when processing large-scale data sets and high-dimensional data. Especially after mapping to high-dimensional space, the computational time and computational space will increase significantly. Although wavelet packet analysis can refine signal features, it also requires processing a large number of data points. This may result in a long training time for the fault prediction model and high requirements for hardware resources.

[0006] (2) Parameter sensitivity: The performance of SVM is highly dependent on the choice of kernel function and the setting of regularization parameters (such as C parameter and gamma parameter). The selection of these parameters has a great impact on the classification results and requires repeated experiments and adjustments. If the parameters are not selected properly, the accuracy of fault prediction may decrease. In addition, the parameter adjustment process is relatively cumbersome, which increases the difficulty of model establishment.

[0007] (3) High requirements for data preprocessing: SVM is sensitive to data scaling. If the data is not normalized, it may lead to deviations in classification results. At the same time, wavelet packet analysis also requires proper preprocessing of the original signal to extract effective feature information. This requires strict data preprocessing steps before using wavelet packet support vector machine for fault prediction, which increases the complexity and workload of data processing.

[0008] Therefore, accurate and rapid prediction of hydraulic system device failures is of great significance to the safe and stable operation of power plants, and also helps to improve device utilization and reduce cost losses. Summary of the invention

[0009] The present invention provides a hydraulic system fault prediction method based on a PSO-BP neural network algorithm. The method utilizes a BP neural network and adopts a particle swarm optimization algorithm to optimize the model weights and thresholds. Then, the hydraulic system-related data and the historical fault data of the components are sorted out and fully mined to establish a prediction model. The prediction model is used to predict the operating status of the hydraulic system components. Compared with other prediction methods, the method of the present invention has higher prediction accuracy in hydraulic system fault prediction.

[0010] In order to realize the above technical features, the object of the present invention is achieved as follows: a hydraulic system fault prediction method based on PSO-BP neural network algorithm, comprising: Raw data collection: Collect the original data of the operation of related components from the hydraulic system; Sample data acquisition: By preprocessing the original data, the original data related to the hydraulic system failure is identified and constituted into sample data; Model training: The fault diagnosis model is established using BP neural network, and the sample data is used as input and the actual fault type is used as output for training. In this process, the weights and thresholds of the BP neural network are optimized through the PSO algorithm to improve the performance and generalization ability of the network. Failure prediction: When the hydraulic system is running, the hydraulic system operation data is collected in real time and preprocessed and feature extracted. Then, these data are input into the trained PSO-BP neural network model. The model will give the possible fault type and severity according to the input data, so as to realize the fault prediction. Verification and optimization: The prediction results of the model are verified by actual fault data, and the model is further optimized based on the verification results to improve the accuracy and efficiency of the prediction.

[0011] The specific process of obtaining the sample data is as follows: The total operating time of the hydraulic system components is divided into multiple equal unit time periods. Each time period of the components corresponds to a sample of the hydraulic system fault prediction model. The fault condition of the components measured by the monitoring system in each time period is used as the output of the sample corresponding to the previous time period. If the component fails in the next time period, the sample corresponding to the previous time period of this time period is recorded as the fault sample. Similarly, if the system operates normally in the next time period, the sample corresponding to the previous time period of this time period is recorded as the normal sample.

[0012] For example, if you want to predict mechanical failures of hydraulic system valve groups, you can use the material quality coefficient of hydraulic system components, the number of actions, and the pressure difference, oil temperature, oil quality, and speed regulator working condition information obtained by the monitoring system as multi-dimensional feature quantities of the sample.

[0013] The fault diagnosis model based on BP neural network includes input layer, hidden layer, output layer and weight; A single hidden layer BP neural network is used, in which the input layer consists of 5 neurons, each neuron represents the 4-dimensional input feature of a research sample, and the remaining neuron is used to adjust the threshold of the hidden layer activation function, and its value is set to -1, resulting in a x i =[ x i1 , x i2 … x i(p-1) ,-1] T Input samples of Then, the number of neurons in the hidden layer is set to 6, of which 5 neurons are connected to the input layer through weights, and the weights are w jk , indicating that the j neurons and k The neurons of the first hidden layer are connected, and the remaining neuron is adjusted according to the threshold effect of the activation function of the output layer, its size is set to -1 and remains unchanged. In this hidden layer, the firstk The net input to each neuron T k Defined as: (1) In the formula, x ij For the i Samples and j The output of a neuron; i For the i input samples; p is the number of neurons in the input layer; Use the sigmoid function as the hidden layer activation function. k The net output of the neurons m k Defined as: (2) In the formula, q is the number of hidden layer neurons.

[0014] The fault prediction of the fault diagnosis model based on BP neural network adopts a binary classification method, which only requires one neuron as output to achieve binary classification; In addition, the sigmoid function is used as the activation function, where each neuron in the hidden layer is connected by a weight w jk Connected to the output neuron, the network output, individual network output, is expressed by the following formula: (3) (4) In the formula, T Output for this network; w k is the weight of the connection between this neuron and the hidden layer neuron; y i For and individual i The output of the neural network.

[0015] The PSO algorithm is used to optimize the initial weights of the BP neural network, including: According to the neural network structure, the initial population is set to several pq + q dimensional row vector, each particle contains all the network's value information, and the mean square error s of the network is used to represent the fitness value of the particle: (5) Where: n represents the number of samples, y i For and individual iThe output of the neural network is y i ' Represents an individual i The actual output of , in each iteration, will track the individual i Best historical position P i,k and the historical best position of the group P q,k , in order to update the speed and position of the particle, so as to obtain the optimal fitness in the population, that is, the particle corresponding to the minimum mean square error. The formulas for these updates are: (6) (7) In the formula, V i,k Indicates i The particle in k The movement speed in the iterations, and c 1 and c 2 is the acceleration factor, which is a non-negative constant; r 1 and r 2 is a random number between 0 and 1; X i,k The i The particle in k The position in the iteration, and w is a non-negative variable inertia weight, and its value decreases linearly with the increase of the number of iterations; (8) In the formula, w max , w min are the maximum and minimum inertia weights, k max is the maximum number of iterations; k is the number of iterations; at the beginning of the iteration, w The larger it is, the more it can exert its global search capabilities, and w The smaller it is, the more it can focus on local search.

[0016] The verification of the prediction results specifically includes: Through the binary classification method, the faults are divided into positive and negative categories. If the fault can be accurately predicted, it is marked as T P , i.e., the true class; if the fault is predicted incorrectly, it is marked as F N , which is a false negative class; if it runs normally but is predicted incorrectly, it is marked as FP , which is the false positive class; if normal operation is correctly predicted, it is marked as T N , that is, the true negative class; the prediction effect is measured by evaluating the hit rate, positive rate, negative rate, and false positive rate of the fault, where R TP and R FP , R ACC are the most important indicators, namely hit rate, false alarm rate and accuracy rate; (9) (10) (11)

[0017] Beneficial effects of the present invention: 1. The present invention provides a hydraulic system fault prediction method using a BP neural network optimized by a particle swarm optimization algorithm. Compared with other prediction methods, this method has higher prediction accuracy in hydraulic system fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0019] Figure 1 This is the structural diagram of the BP neural network of the present invention.

[0020] Figure 2 This is a flow chart of the hydraulic system fault prediction method based on the PSO-BP neural network of the present invention.

[0021] Figure 3 This is a comparison of the prediction effects of the prediction model of the present invention. DETAILED DESCRIPTION

[0022] The embodiments of the present invention are further described below in conjunction with the accompanying drawings.

[0023] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0024] Embodiment 1: See also Figure 1-3 The present invention is a hydraulic system fault prediction method based on PSO-BP neural network, which is described in detail as follows: 1. Failure mechanism of hydraulic system: Common faults of hydraulic systems include: (1) Sudden failures: hydraulic cylinder body explosion, pump shaft breakage: This type of failure usually cannot be predicted early, and once it occurs, the consequences are more serious.

[0025] (2) Wear failure: such as wear on the outer diameter of the plunger of a plunger pump. This type of failure can be predicted early through testing, that is, the failure has a process of formation and development.

[0026] (3) Oil leakage: It is caused by aging of seals, wear and looseness of seals, etc. It is divided into internal leakage and external leakage, which manifests as liquid loss, liquid foam, etc. Leakage will not only lead to a reduction in hydraulic oil, but may also cause safety hazards such as fire.

[0027] (4) Failure due to misuse: The load, pressure, or flow rate exceeds the rated value. For example, the load may burn out the motor.

[0028] (5) Inherent weakness failure: If the design is unreasonable, although the load, pressure, flow, etc. during use do not exceed the specified values ​​at the time of design, these values ​​are not applicable to the actual situation, resulting in failure.

[0029] (6) Other common faults: oil contamination, hydraulic system overheating, insufficient hydraulic system pressure, noise and abnormal vibration, slow movement or no movement, etc.

[0030] Common fault diagnosis methods include: pressure testing, sensory diagnosis (using auditory, visual, tactile and other sensory means to determine whether there is an abnormality in the hydraulic system), actuator speed testing (by testing the speed of the actuator to reflect the changes in certain components in the hydraulic system).

[0031] The types of hydraulic system faults are complex. Therefore, the fault prediction of the hydraulic system selects wear faults as the main research direction for the design of the weak link identification method of the distribution network. The fault prediction method mainly obtains the characteristic quantities of each dimension of the target components of the hydraulic system through the monitoring system, substitutes them into the corresponding prediction model, and performs fault prediction on them.

[0032] 2. Generate samples: The present invention divides the total operation time of hydraulic system components into multiple equal unit time periods. Each time period of the component corresponds to a sample of the hydraulic system fault prediction model, and the fault condition of the component measured by the monitoring system in each time period is used as the output of the sample corresponding to the previous time period. For example, if the mechanical fault of the hydraulic system valve group is to be predicted, the material quality coefficient of the hydraulic system component, the number of actions (duration of use), and the pressure difference, oil temperature, oil quality, and working condition of the speed regulator obtained by the monitoring system are used as multi-dimensional feature quantities of the sample. If a component fails in the next time period, the sample corresponding to the previous time period of this time period is recorded as a fault sample; similarly, if the system operates normally in the next time period, the sample corresponding to the previous time period of this time period is recorded as a normal sample.

[0033] 3. BP neural network model: To deeply understand the BP neural network, we must master four basic concepts: input layer, hidden layer, output layer and weight. Each layer can be regarded as a network composed of three basic layers and the weights between them, but the number of layers and nodes in the hidden layer may vary, so it is necessary to flexibly handle it according to the actual situation to achieve the best effect. The structure of the BP neural network can be seen in Figure 1 .

[0034] (1) Input layer: The input end of information. In the figure, the input layer has only one node. In the actual network, it depends on the specific situation.

[0035] (2) Hidden layer: the information processing end, used to simulate a calculation process. There is only one hidden layer in the figure, with 5 nodes.

[0036] (3) Output layer: The output end of information, which is the result we want. In the above figure, R is the only node in the output layer. In fact, there may be many output nodes.

[0037] (4) Weight: The parameter that connects each layer of information. In the above figure, it is only reflected in the form of product.

[0038] The BP neural network is a complex feedforward mechanism consisting of three layers: input layer, hidden layer, and output layer. This mechanism is able to capture and process complex information and transform it into a model that can be used for decision-making. In these three layers, each layer will determine the behavior of other layers based on its own capabilities and performance, and these decisions will have an important impact on the performance of the entire machine. If the output result is far from the expected result, the algorithm will adopt a backpropagation method to adjust the weights and thresholds of the network according to the deviation of the predicted result.

[0039] After a forward or reverse propagation, the parameters of the network can be adjusted to build a model, that is, a training network. The purpose of this model is to make the performance of the model close to the situation of real data through continuous iteration. Although BP neural networks have a higher level and number of nodes in theory, which can effectively explore the connection between various research data, in practical applications, due to the increase in computational complexity, the data they can process is still quite limited.

[0040] This embodiment uses a single hidden layer BP neural network, in which the input layer consists of 5 neurons, each neuron represents a 4-dimensional input feature of a research sample, and the remaining neuron is used to adjust the threshold of the hidden layer activation function, and its value is set to -1, resulting in a x i =[ xi1 , x i2 … x i(p-1) ,-1] T Input samples of Then, the number of neurons in the hidden layer is set to 6, of which 5 neurons are connected to the input layer through weights, and the weights are w jk , indicating that the j neurons and k The neurons of the first hidden layer are connected, and the remaining neuron is adjusted according to the threshold effect of the activation function of the output layer, its size is set to -1 and remains unchanged. In this hidden layer, the first k The net input to each neuron T k Defined as: (1) In the formula, x ij For the i Samples and j The output of a neuron; i For the i input samples; p is the number of neurons in the input layer; Use the sigmoid function as the hidden layer activation function. k The net output of the neurons m k Defined as: (2) In the formula, q is the number of hidden layer neurons.

[0041] In the present invention, fault prediction adopts a binary classification method. Therefore, only one neuron is needed as output to achieve the purpose of binary classification. In addition, this method also adopts a special activation function, namely the sigmoid function, in which each neuron in the hidden layer is weighted. w jk Connected to the output neuron. The network output and individual network output can be expressed by the following formula: (3) (4) In the formula, T Output for this network; w k is the weight of the connection between this neuron and the hidden layer neuron; y i For and individual iThe output of the neural network.

[0042] 4. PSO algorithm: The PSO algorithm is a global search algorithm based on the foraging behavior of bird flocks. It can effectively improve the problem of traditional BP neural network training that is prone to local extreme values. It can achieve local optimization while maintaining globality, thereby achieving ideal training results. Therefore, this paper uses the PSO algorithm to optimize the initial weights.

[0043] It symbolically regards the solution process as a flock of birds hunting, and each bird is a particle, which is also an optimization problem. The bird can sense the food around it, so the optimal solution for hunting is to find the bird closest to the food, remember its position, and then judge the location of the food based on experience.

[0044] Therefore, each bird is characterized by two quantities: speed and position. Speed ​​refers to how fast the bird moves, while position refers to the direction in which the bird moves. When the program runs, each bird will fly in the feasible domain and search for the optimal solution individually. The value found is called the current individual extreme value. This value is recorded and then iterated and compared with the individual extreme values ​​found by other birds. Finally, the best one among all individual extreme values ​​is recorded and called the global extreme value. In each iteration, the speed and position will change. The specific update formula is as follows:

[0045]

[0046] in, is the inertia weight; , is the learning factor; , is a random number greater than or equal to 0 and less than or equal to 1; is the individual extreme value, which is also the value remembered by the particle in the last optimization process; is the group extreme value, that is, the entire global optimal solution; , is the speed before and after the update; , The positions before and after the update.

[0047] The update formula shows that the speed of the particle after the update is mainly determined by three parts: memory, self-cognition and group cognition. The so-called memory term is the motion state of the particle at the moment, also known as the influence of the last speed on the current speed after the update; the so-called self-cognition term is derived from the particle's own experience, and adjusts its own search direction according to memory; the so-called group cognition term means that the particle not only relies on itself, but also combines the group communication with all particles to change the search direction.

[0048] It can be seen that the standard PSO is easy to calculate due to its fewer preset parameters and simple iterative update method. In addition, the goal is clear throughout the solution process. Once the direction of solving the problem is determined, it will persist in approaching this direction, avoiding the waste of time caused by repetition, and the solution efficiency is high. Not only that, in most cases, the results it solves are almost the same as the actual value, that is, as long as a sufficient number of iterations and a large enough population size are selected, the accuracy of the results is very high.

[0049] 5. Optimization of BP neural network algorithm by PSO algorithm: The PSO algorithm is a global search algorithm based on the foraging behavior of bird flocks. It can effectively improve the problem of traditional BP neural network training process that is prone to fall into local extreme values. It can achieve local optimization while maintaining globality, thereby achieving ideal training results. Therefore, the present invention uses the PSO algorithm to optimize the initial weights.

[0050] According to the neural network structure, the initial population is set to several pq + q dimensional row vector, each particle contains all the network's value information, and the mean square error s of the network is used to represent the fitness value of the particle: (5) Where: n represents the number of samples, y i For and individual i The output of the neural network is y i ' Represents an individual i The actual output of , in each iteration, will track the individual i Best historical position P i,k and the historical best position of the group P q,k , in order to update the speed and position of the particle, so as to obtain the optimal fitness in the population, that is, the particle corresponding to the minimum mean square error. The formulas for these updates are: (6) (7) In the formula, V i,k Indicates i The particle in k The movement speed in the iterations, and c 1 and c 2 is the acceleration factor, which is a non-negative constant; r 1 andr 2 is a random number between 0 and 1; X i,k The i The particle in k The position in the iteration, and w is a non-negative variable inertia weight, and its value decreases linearly with the increase of the number of iterations; (8) In the formula, w max , w min are the maximum and minimum inertia weights, k max is the maximum number of iterations; k is the number of iterations; at the beginning of the iteration, w The larger it is, the more it can exert its global search capabilities, and w The smaller it is, the more it can focus on local search.

[0051] 6. Fault prediction method of hydraulic system components: like Figure 2 As shown in the figure, the specific situation of using the established model to realize the fault prediction of hydraulic system components is as follows: (1) Allow the computer to read the material quality coefficient of the hydraulic system components, the number of actions (duration of use) detected by the monitoring system, and the historical data such as the pressure difference of the components, the oil temperature, the oil quality coefficient, the working condition of the speed regulator, etc. obtained by the monitoring system.

[0052] (2) The total operating time of the system is divided into multiple equal unit time periods. Each time period of the component corresponds to a sample of the hydraulic system fault prediction model. The fault condition of the component measured by the monitoring system in each time period is used as the output of the sample corresponding to the previous time period.

[0053] If the system operates normally in the next time period, the sample corresponding to the previous time period of this time period is recorded as a normal sample, and the output value of the sample is recorded as 1; conversely, if the component fails in the next time period, the sample corresponding to the previous time period of this time period is recorded as a faulty sample, and the output value of the sample is set to 2.

[0054] The data required for the hydraulic system fault prediction model used in the present invention comes from some monitoring data of a power station and the corresponding hydraulic system fault information. The total time of system operation is divided into multiple unit time periods of unit time length, and the corresponding information in each unit time period is counted, including the number of hydraulic system component actions (use time), and the historical data such as the pressure difference, oil temperature, oil quality coefficient, and working condition of the governor obtained by the monitoring system. A random sampling algorithm is used to make the fault samples and normal samples the same to avoid interference from the data itself to the neural network. From a total of 2,000 samples, we selected 80% of the samples for training and the remaining 20% ​​for testing, so that 800 fault samples and 800 normal samples can be obtained in the training samples, and 200 fault samples and 200 normal samples can be obtained in the test samples, thereby achieving effective simulation of the system.

[0055] VII. Forecasting effect evaluation indicators: Through the binary classification method, the faults are divided into positive and negative categories. If the fault can be accurately predicted, it is marked as T P , i.e., the true class; if the fault is predicted incorrectly, it is marked as F N , which is a false negative class; if it runs normally but is predicted incorrectly, it is marked as F P , which is the false positive class; if normal operation is correctly predicted, it is marked as T N , that is, the true negative class; the prediction effect is measured by evaluating the hit rate, positive rate, negative rate, and false positive rate of the fault, where R TP and R FP , R ACC are the most important indicators, namely hit rate, false alarm rate and accuracy rate; (9) (10) (11).

[0056] 8. Use the model for verification: In order to measure the prediction effect of each prediction model and evaluate the hit rate, false alarm rate and accuracy of fault prediction, RTP, RFP and RACC are the most important indicators. The comparison of the early warning accuracy of hydraulic system fault prediction samples based on BP neural network model, PSO-BP neural network model and SVM model is shown in Table 1.

[0057] From the results, we can see that the PSO-BP neural network model has a high hit rate, a low false alarm rate, and the accuracy is the best among the three prediction models. Figure 3 As shown in the figure, the proposed method has better performance in predicting wear failures of hydraulic system components.

[0058] Table 1 Comparison of prediction model results

Claims

1. A hydraulic system fault prediction method based on PSO-BP neural network algorithm, characterized in that: include: Raw data collection: Collect the original data of the operation of related components from the hydraulic system; Sample data acquisition: By preprocessing the original data, the original data related to the hydraulic system failure is identified and constituted into sample data; Model training: The fault diagnosis model is established using BP neural network, and the sample data is used as input and the actual fault type is used as output for training. In this process, the weights and thresholds of the BP neural network are optimized through the PSO algorithm to improve the performance and generalization ability of the network. Failure prediction: When the hydraulic system is running, the hydraulic system operation data is collected in real time and preprocessed and feature extracted. Then, these data are input into the trained PSO-BP neural network model. The model will give the possible fault type and severity according to the input data, so as to realize the fault prediction. Verification and optimization: The prediction results of the model are verified by actual fault data, and the model is further optimized based on the verification results to improve the accuracy and efficiency of the prediction.

2. The hydraulic system fault prediction method based on PSO-BP neural network algorithm according to claim 1 is characterized in that: The specific process of obtaining the sample data is as follows: The total operating time of the hydraulic system components is divided into multiple equal unit time periods. Each time period of the components corresponds to a sample of the hydraulic system fault prediction model. The fault condition of the components measured by the monitoring system in each time period is used as the output of the sample corresponding to the previous time period. If the component fails in the next time period, the sample corresponding to the previous time period of this time period is recorded as the fault sample. Similarly, if the system operates normally in the next time period, the sample corresponding to the previous time period of this time period is recorded as the normal sample.

3. The hydraulic system fault prediction method based on PSO-BP neural network algorithm according to claim 2 is characterized in that: For example, if you want to predict mechanical failures of hydraulic system valve groups, you can use the material quality coefficient of hydraulic system components, the number of actions, and the pressure difference, oil temperature, oil quality, and speed regulator working condition information obtained by the monitoring system as multi-dimensional feature quantities of the sample.

4. The hydraulic system fault prediction method based on PSO-BP neural network algorithm according to claim 2 is characterized in that: The fault diagnosis model based on BP neural network includes input layer, hidden layer, output layer and weight; A single hidden layer BP neural network is used, in which the input layer consists of 5 neurons, each neuron represents the 4-dimensional input feature of a research sample, and the remaining neuron is used to adjust the threshold of the hidden layer activation function, and its value is set to -1, resulting in a x i =[ x i1 , x i2 … x i(p-1) ,-1] T Input samples of Then, the number of neurons in the hidden layer is set to 6, of which 5 neurons are connected to the input layer through weights, and the weights are w jk , indicating that the j neurons and k The neurons of the first hidden layer are connected, and the remaining neuron is adjusted according to the threshold effect of the activation function of the output layer, its size is set to -1 and remains unchanged. In this hidden layer, the first k The net input to each neuron T k Defined as: （1） In the formula, x ij For the i The samples and j The output of a neuron; i For the i input samples; p is the number of neurons in the input layer; Use the sigmoid function as the hidden layer activation function. k The net output of the neurons m k Defined as: （2） In the formula, q is the number of hidden layer neurons.

5. The hydraulic system fault prediction method based on PSO-BP neural network algorithm according to claim 4 is characterized in that: The fault prediction of the fault diagnosis model based on BP neural network adopts a binary classification method, which only requires one neuron as output to achieve binary classification; In addition, the sigmoid function is used as the activation function, where each neuron in the hidden layer is connected by a weight w jk Connected to the output neuron, the network output, individual network output, is expressed by the following formula: （3） （4） In the formula, T Output for this network; w k is the weight of the connection between this neuron and the hidden layer neuron; y i For and individual i The output of the neural network.

6. The hydraulic system fault prediction method based on PSO-BP neural network algorithm according to claim 5 is characterized in that: The PSO algorithm is used to optimize the initial weights of the BP neural network, including: According to the neural network structure, the initial population is set to several pq + q dimensional row vector, each particle contains all the network's value information, and the mean square error s of the network is used to represent the fitness value of the particle: （5） Where: n represents the number of samples, y i For and individual i The output of the neural network is y i ' Represents an individual i The actual output of , in each iteration, will track the individual i Best historical position P i,k and the historical best position of the group P q,k , in order to update the speed and position of the particle, so as to obtain the optimal fitness in the population, that is, the particle corresponding to the minimum mean square error. The formulas for these updates are: （6） （7） In the formula, V i,k Indicates i The particle in k The movement speed in the iterations, and c 1 and c 2 is the acceleration factor, which is a non-negative constant; r 1 and r 2 is a random number between 0 and 1; X i,k The i The particle in k The position in the iteration, and w is a non-negative variable inertia weight, and its value decreases linearly with the increase of the number of iterations; （8） In the formula, w max , w min are the maximum and minimum inertia weights, k max is the maximum number of iterations; k is the number of iterations; at the beginning of the iteration, w The larger it is, the more it can exert its global search capabilities, and w The smaller it is, the more it can focus on local search.

7. The hydraulic system fault prediction method based on PSO-BP neural network algorithm according to claim 6 is characterized in that: The verification of the prediction results specifically includes: Through the binary classification method, the faults are divided into positive and negative categories. If the fault can be accurately predicted, it is marked as T P , i.e., the true class; if the fault is predicted incorrectly, it is marked as F N , which is a false negative class; if it runs normally but is predicted incorrectly, it is marked as F P , which is the false positive class; if normal operation is correctly predicted, it is marked as T N , that is, the true negative class; the prediction effect is measured by evaluating the hit rate, positive rate, negative rate, and false positive rate of the fault, where R TP and R FP , R ACC are the most important indicators, namely hit rate, false alarm rate and accuracy rate; （9） （10） （11）。

Images