Customer relationship management prediction maintenance method based on deep learning

Abstract

The invention discloses a customer relationship management prediction maintenance method based on deep learning. The method comprises the following steps: collecting customer data and preprocessing the customer data; feature engineering extraction is carried out; establishing a customer loss prediction model by combining elastic weight integration and an enhanced dynamic generative adversarial network; adjusting a model structure and parameters by using a heterogeneous graph neural network and deep algebraic geometry learning; capturing a time dynamic state through a recurrent neural network to perform real-time data flow analysis; generating a personalized recommendation scheme by applying collaborative filtering analysis; and predicting faults and taking maintenance measures by combining long and short-term memory network monitoring equipment data. Through the steps, the prediction accuracy, the personalized service level and the resource configuration efficiency are improved, and the customer loss and the maintenance cost are reduced.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a customer relationship management predictive maintenance method based on deep learning. Background Art

[0002] With the rapid development of artificial intelligence and big data technology, customer relationship management systems play an increasingly important role in modern enterprise management. Customer relationship management systems help companies optimize customer service processes, improve customer satisfaction and loyalty, and thus enhance their market competitiveness by collecting and analyzing customer behavior data, transaction data, and real-time data. However, existing customer relationship management systems still face many challenges in data processing, analysis, and prediction. In particular, when faced with massive data and complex customer relationships, traditional methods have low prediction accuracy and real-time performance.

[0003] First, the existing customer relationship management system has obvious deficiencies in data processing efficiency. Traditional customer relationship management systems usually rely on manual and semi-automatic means to collect and pre-process customer data, and data cleaning, missing value filling and standardization processing are inefficient. This not only increases the time cost of data processing, but also makes it difficult to adapt to the needs of real-time data processing, resulting in low data quality and affecting the accuracy of subsequent analysis and prediction.

[0004] Secondly, existing technologies have obvious defects in feature engineering. Feature engineering is a very important part of data mining and machine learning, which directly affects the prediction effect of the model. When processing customer data, traditional feature engineering methods are often unable to fully extract and select features that can reflect customer behavior and preferences, resulting in low prediction accuracy of the model. In addition, traditional methods usually rely on manual experience when extracting and selecting features, lack automation and intelligent means, and it is difficult to find potential useful information in massive data.

[0005] Third, existing customer churn prediction models are deficient in terms of accuracy and stability. Traditional customer churn prediction methods mainly rely on simple statistical analysis and machine learning methods, such as logistic regression and decision trees, which are difficult to effectively capture the deep patterns and changes in customer behavior when dealing with large-scale and complex customer data. Especially when customer behavior and market environment change, the prediction effect of traditional models is often greatly reduced. In addition, traditional models are prone to overfitting during the training process, resulting in poor performance of the model on new data.

[0006] In summary, the existing customer relationship management systems have many shortcomings in data processing, feature engineering and customer churn prediction, and it is difficult to meet the needs of modern enterprises for efficient, accurate and real-time customer management. Therefore, how to provide a customer relationship management prediction and maintenance method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] One purpose of the present invention is to propose a customer relationship management predictive maintenance method based on deep learning, which makes full use of technologies such as elastic weight integration, enhanced dynamic generative adversarial networks, heterogeneous graph neural networks and deep algebraic geometry learning, and describes in detail a method for intelligently realizing customer churn prediction and maintenance, which has the advantages of high prediction accuracy, high personalized service level and high resource allocation efficiency.

[0008] A customer relationship management predictive maintenance method based on deep learning according to an embodiment of the present invention comprises the following steps:

[0009] S1. Collect customer data, including historical behavior data, transaction data and real-time data;

[0010] S2. Preprocess the collected customer data, including data cleaning, missing value filling and standardization;

[0011] S3. Perform feature engineering on the preprocessed customer data to extract and select features that can reflect customer behaviors and preferences;

[0012] S4. A customer churn prediction model is established by combining elastic weight integration and enhanced dynamic generative adversarial network. The elastic weight integration is used to maintain parameter stability, and the enhanced dynamic generative adversarial network is used to generate and adjust customer churn samples.

[0013] S5. Use heterogeneous graph neural networks to process multi-type customer relationship data, and adjust model structure and parameters in combination with deep algebraic geometry learning;

[0014] S6. Capture the temporal dynamics of customer behavior through recurrent neural networks and combine them with optimized models for real-time data stream analysis;

[0015] S7. Based on the prediction results of the optimization model, collaborative filtering is used to analyze customer historical behavior data and generate personalized product and service recommendation solutions;

[0016] S8. Combine long short-term memory networks to monitor and analyze equipment operation data, predict potential failures, and take maintenance measures in advance.

[0017] Optionally, S3 includes the following steps:

[0018] S31, extract features from the preprocessed customer data set D′ to generate an initial feature set F, where each feature f i Can reflect specific customer behaviors or preferences;

[0019] S32, using the principal component analysis method to perform dimensionality reduction processing on the initial feature set F, projecting the high-dimensional feature data into a low-dimensional space, retaining the main components, and generating a feature matrix F′ after dimensionality reduction;

[0020] S33. Use the mutual information method to evaluate each feature f i The correlation between the customer churn label y and the mutual information value I(f i ,y);

[0021]

[0022] Among them, p(x,y) represents the joint probability distribution, p(x) and p(y) represent the marginal probability distribution respectively;

[0023] S34, select features based on the mutual information value, set the threshold θ, and retain the mutual information value I(f i ,y)≥θ features, forming the filtered feature set F”;

[0024] S35, construct features for the filtered feature set F" to generate a new combined feature f c ;

[0025] S36, standardize the constructed feature set to generate the final feature set F f ;

[0026] S37, the final feature set F f Combined with the customer data label y, a feature label pair (F f ,y), as the data input for subsequent model training.

[0027] Optionally, S4 includes the following steps:

[0028] S41, initialize the parameters of the customer churn prediction model, define the neural network structure, including the input layer, multiple hidden layers and the output layer, the input layer receives the feature set F f , the output layer generates the probability of customer churn

[0029] S42. Use elastic weight integration technology to calculate the weight importance of each parameter Ω i , the specific steps include:

[0030] After each task training, the Fisher information matrix is ​​calculated

[0031]

[0032] Among them, p(x|θ i ) is based on the parameter θ i Likelihood function, x is the observed data, E is the expected value of the data distribution;

[0033] The specific form of the likelihood function is:

[0034]

[0035] Among them, x j represents the jth observed data point, and n is the total number of observed data points;

[0036] S43, accumulate the Fisher information matrix and obtain the parameter importance:

[0037]

[0038] S44. During the training process, adjust the loss function to include parameter importance terms and update the neural network parameters:

[0039]

[0040] Among them, L original represents the original loss function, λ represents the regularization coefficient, θ i and Represent the parameters of the current and previous tasks respectively;

[0041] S45. Introduce the enhanced dynamic generative adversarial network, including the construction and training of the generator G and the discriminator D:

[0042] The generator G is based on the noise vector z and the parameter θ g Generate customer churn samples

[0043]

[0044] where z is sampled from a standard normal distribution, θ g Represents the parameters of the generator;

[0045] The discriminator D evaluates the difference between the generated samples and the real samples and calculates the loss:

[0046]

[0047] Among them, D(x) and They represent the samples generated by the discriminator for the real sample x The discriminant output of

[0048] S46, reinforcement learning method optimizes between the generator and the discriminator, and updates the parameters of the generator through the policy gradient method:

[0049]

[0050] Among them, J(θ g ) is the objective function of the generator, Represents the generator parameter θ g The gradient of

[0051] S47, the optimization process of the discriminator D includes using the real customer data x and the generated customer churn samples Update the parameters of the discriminator and minimize the loss function L D ;

[0052] S48, the optimization process of the generator G comprises the following steps:

[0053] Apply the policy gradient method to update the parameters of the generator so that the generated customer churn samples are closer to the real data distribution;

[0054] According to the feedback from the discriminator, the parameters of the generator are gradually adjusted:

[0055]

[0056] Among them, α is the learning rate, θ g Represents the parameters of the generator;

[0057] S49. Combining the advantages of generative adversarial networks and reinforcement learning, the reward mechanism R of reinforcement learning is used to dynamically adjust the objective functions of the generator and discriminator:

[0058] R = E[r t |s t ,a t ];

[0059] Among them, r t is the reward value at time step t, s t is the current state, a t for the actions taken;

[0060] The outputs of the generator and discriminator are calculated using the multi-layer perceptron structure in the generative adversarial network:

[0061]

[0062] Among them, σ is the activation function, W is the weight matrix, b is the bias term, x is the input feature, is the output;

[0063] S410, introduce the time series embedding method to encode the time information into the generator input to achieve time-aware generation:

[0064]

[0065] Where t represents the time step, represents the customer churn sample generated at time t, z t is the noise vector of time step t;

[0066] S411. Use the optimized generator and discriminator to generate more high-quality customer churn samples, and perform mixed training with real customer data;

[0067] S412: Apply the optimized customer churn prediction model to the comprehensive customer data set D' to calculate the churn probability of each customer The results are then used for subsequent analysis and maintenance decisions.

[0068] Optionally, S5 includes the following steps:

[0069] S51. Construct a heterogeneous graph neural network and define a heterogeneous graph Q=(V,E) containing multiple types of nodes and edges, where V represents a node set, E represents an edge set, and different types of nodes and edges represent different types of customer data and relationships, respectively;

[0070] S52, for each node v i ∈V initializes the feature vector These feature vectors come from the preprocessed feature set F f ;

[0071] S53, performing feature aggregation on the heterogeneous graph through a message passing mechanism, specifically comprising the following steps:

[0072] For each node v on each layer l i , from its neighbor node N(v i ) receives messages and aggregates them:

[0073]

[0074] Among them, σ is the activation function, W h and W e are the weight matrices of node features and edge features, e ij Represents node v i and node v j The edge features between

[0075] For each node v i , update its eigenvector:

[0076]

[0077] in, and They are the weight matrices of the node features and message features of the current layer respectively;

[0078] S54, through multi-layer feature aggregation operations, finally generate the embedding representation h of each node i ;

[0079] S55. Apply deep algebraic geometry learning to optimize node embedding, specifically including the following steps:

[0080] Use algebraic geometry methods to extract high-order relationships between node embeddings, construct algebraic geometry optimization objective functions, and define two algebraic geometry mapping functions f and g as:

[0081] f(h i )=A i h i ;

[0082] g(h j )=B j h j ;

[0083] Among them, A i and B j is the mapping matrix;

[0084] The algebraic geometry optimization objective function is:

[0085]

[0086] Topology optimization is introduced in algebraic geometry space, and the following loss function is constructed through topology structure preservation technology:

[0087]

[0088] Among them, w ij Represents node v i and v j The weights between them are calculated through the topological relationships embedded in the nodes;

[0089] Combining algebraic geometry optimization and topology preservation loss, a comprehensive optimization goal is formed:

[0090] L total =L DAGL +βL TSP ;

[0091] Among them, β is the trade-off coefficient;

[0092] S56. Optimize the comprehensive loss function L by gradient descent method total,update the parameters of the heterogeneous graph neural network so that the node embeddings better reflect the high-order structure and topological features in the customer relationship data;

[0093] S57, embed the optimized node into h i Used for subsequent customer churn prediction model training;

[0094] S58. Apply the adjusted model structure and parameters to the comprehensive customer data set D′ to calculate the churn probability of each customer The results are then used for subsequent analysis and maintenance decisions.

[0095] Optionally, the S6 includes the following steps:

[0096] S61. Build a recursive neural network model, define the input layer, hidden layer and output layer, and the input layer receives time series data x t , the output layer generates the predicted value for the corresponding time step

[0097] S62, for the input time series data x t Preprocess and standardize the data features of each time step to meet the input requirements of the recurrent neural network model:

[0098] S63. Define the hidden layer state h of the recurrent neural network t The update formula is as follows:

[0099] h t =σ(W h x t +U h h t-1 +b h );

[0100] Among them, σ is the activation function, W h is the input weight matrix, U h is the hidden state weight matrix, b h is the bias vector, x t is the input of the current time step, h t-1 is the hidden state of the previous time step;

[0101] S64. Combined with the variational autoencoder to further capture long-term dependencies, the update formula of VAE is defined as follows:

[0102] For the input time series data x t , first generate the latent variable z through the encoder q(z|x) t , the formula is:

[0103] z t =μ+σ⊙∈;

[0104] Where μ and σ represent the mean and standard deviation of the latent variable, respectively, and ∈ is the noise sampled from the standard normal distribution;

[0105] Then the decoder p(x|z) generates the reconstructed time series data The formula is:

[0106]

[0107] Among them, W z is the weight matrix of the decoder, b z is the bias vector of the decoder;

[0108] S65. Use the optimized node embedding to represent h i and time series data x t Combine and dynamically adjust through the following formula:

[0109]

[0110] Among them, W y is the output weight matrix, b y is the output bias vector, is the predicted output at time step t;

[0111] S66. Combine the optimized heterogeneous graph neural network and deep algebraic geometry learning model to combine the prediction results of time series data with node embedding representation to generate comprehensive customer churn prediction results

[0112] S67. Analyze real-time data streams and continuously train and optimize models based on new data;

[0113] S68. Apply the optimized prediction model to the comprehensive customer data set D' to calculate the churn probability of each customer at different time steps The results are then used for subsequent analysis and maintenance decisions.

[0114] Optionally, the S7 includes the following steps:

[0115] S71. Collect the customer's historical behavior data B to form a behavior matrix R, where R ij represents the rating or behavior value of customer i on product j;

[0116] S72. Standardize the behavior matrix R and normalize the score or behavior value to the interval [0,1]:

[0117] S73, applying a collaborative filtering method to perform similarity calculation on the standardized behavior matrix R, specifically comprising the following steps:

[0118] Calculate the similarity between customers using the cosine similarity formula:

[0119]

[0120] Among them, sim(u,v) represents the similarity between customer u and customer v, R uj and R vj They represent the ratings of customer u and customer v on product j respectively;

[0121] S74. Generate a recommendation list for the target customer based on customer similarity. The specific steps are as follows:

[0122] Based on the top k similar customers with the highest similarity, calculate the target customer's predicted rating for the unrated product:

[0123]

[0124] in, represents the predicted rating of target customer i for product j, N k (i) represents the top k customers with the highest similarity to target customer i;

[0125] S75. Combined prediction results To optimize the recommendation list, the specific steps are as follows:

[0126] The probability of customer churn With recommendation rating Combine them to generate an optimized recommendation score:

[0127]

[0128] in, represents the optimized recommendation score, α is the weight coefficient, which represents the balance coefficient between the recommendation score and the churn probability;

[0129] S76. Sort the optimized recommendation scores to generate a personalized recommendation list, and apply the recommendation results to the customer relationship management system to provide personalized product and service recommendation solutions for each customer;

[0130] S77. Continuously optimize and update the recommendation model based on actual feedback data.

[0131] Optionally, the S8 includes the following steps:

[0132] S81. Collect equipment operation data, including sensor data, log data, and historical maintenance records, to form equipment operation data set D e ;

[0133] S82. Run data set D on the devicee Perform preprocessing, including data cleaning, missing value filling and standardization;

[0134] S83. Build a long short-term memory network LSTM model, define the input layer, LSTM unit layer and output layer, and the input layer receives time series data x t , the output layer generates the predicted value for the corresponding time step

[0135] S84. Define the update formula of the LSTM unit as follows:

[0136] i t =σ(W i x t +U i h t-1 +b i );

[0137] f t =σ(W f x t +U f h t-1 +b f );

[0138] o t =σ(W o x t +U o h t-1 +b o );

[0139] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c );

[0140] h t =o t ⊙tanh(c t );

[0141] Among them, i t 、f t , o t They represent the input gate, forget gate and output gate respectively, and c t is the unit state, ⊙ represents the element-by-element product operation, W i , W f , W o , W c is the input weight matrix, U i , Uf , U o , U c is the hidden state weight matrix, b i , b f , b o , b c is the bias vector;

[0142] S85, the pre-processed equipment operation data x t Input into the LSTM model and obtain the hidden state h of each time step through recursive calculation t and the cell state c t ;

[0143] S86. Combining latent Dirichlet allocation (LDA) to perform topic modeling on the equipment operation data and extract potential failure modes, specifically including the following steps:

[0144] For each document d, its topic distribution θ d Generated by Dirichlet distribution, its formula is:

[0145]

[0146] Among them, α=(α1,α2,…,α K ) is the hyperparameter of the Dirichlet distribution, θ di represents the proportion of topic i in document d, Γ(·) is the Gamma function;

[0147] For each topic k, its word distribution Generated by Dirichlet distribution, its formula is:

[0148]

[0149] Where β=(β1,β2,…,β V ) is the hyperparameter of the Dirichlet distribution, represents the proportion of word j in topic k.

[0150] S87. Use Gibbs sampling method to estimate the parameters of LDA model and obtain the topic distribution θ of each document through sampling. d And the word distribution of each topic

[0151] S88. Prediction results combined with LSTM model and the failure mode θ of the LDA model d and Generate a comprehensive maintenance plan to proactively implement maintenance actions, including the following steps:

[0152] Calculate the failure probability P(f t |x t ):

[0153] P(f t |x t )=σ(W f h t +b f );

[0154] Among them, σ is the activation function, W f is the output weight matrix, b f is the output bias vector;

[0155] According to the failure probability P(f t |x t ) and failure modes of the LDA model Generate maintenance plans, prioritizing high-risk equipment;

[0156] S88. Apply comprehensive maintenance plans to equipment management systems for real-time monitoring and dynamic adjustment, and continuously optimize maintenance strategies by combining new equipment operation data and model prediction results;

[0157] S89. Apply the optimized maintenance plan and equipment failure prediction model to the comprehensive equipment data set D e , calculate the failure probability of each device at different time steps The results are used for subsequent analysis and maintenance decisions. The beneficial effects of the present invention are:

[0158] (1) The present invention proposes a customer relationship management predictive maintenance method based on deep learning. By collecting and preprocessing customer data, including historical behavior data, transaction data and real-time data, the problem of low data processing efficiency in the prior art is solved, and the speed and quality of data processing are significantly improved.

[0159] (2) The present invention establishes a customer churn prediction model by combining elastic weight integration and enhanced dynamic generative adversarial network, maintains the stability of model parameters, and generates and adjusts customer churn samples at the same time, thereby improving the accuracy and stability of customer churn prediction and solving the problem of insufficient prediction accuracy in the prior art.

[0160] (3) The present invention uses heterogeneous graph neural networks to process multi-type customer relationship data, and combines deep algebraic geometry learning to optimize model structure and parameters, thereby improving the modeling and analysis capabilities of complex customer relationships and solving the problem of weak ability to process complex customer relationships in the prior art.

[0161] (4) By capturing the temporal dynamics of customer behavior through a recursive neural network and combining it with the aforementioned optimized model for real-time data stream analysis, dynamic monitoring and response to customer behavior can be achieved, solving the problem of the lack of real-time analysis capabilities in existing technologies and significantly improving the accuracy and timeliness of personalized recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0162] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0163] Figure 1 A flowchart of a customer relationship management predictive maintenance method based on deep learning proposed by the present invention;

[0164] Figure 2 A flowchart of establishing a customer churn prediction model by using elastic weight integration and enhanced dynamic generative adversarial network in the present invention;

[0165] Figure 3 Flowchart of using heterogeneous graph neural network and deep algebraic geometry learning to process and optimize customer relationship data in the present invention. DETAILED DESCRIPTION

[0166] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0167] refer to Figure 1-3 , a customer relationship management predictive maintenance method based on deep learning, comprising the following steps:

[0168] S1. Collect customer data, including historical behavior data, transaction data and real-time data;

[0169] S2. Preprocess the collected customer data, including data cleaning, missing value filling and standardization;

[0170] S3. Perform feature engineering on the preprocessed customer data to extract and select features that can reflect customer behaviors and preferences;

[0171] S4. A customer churn prediction model is established by combining elastic weight integration and enhanced dynamic generative adversarial network. The elastic weight integration is used to maintain parameter stability, and the enhanced dynamic generative adversarial network is used to generate and adjust customer churn samples.

[0172] S5. Use heterogeneous graph neural networks to process multi-type customer relationship data, and adjust model structure and parameters in combination with deep algebraic geometry learning;

[0173] S6. Capture the temporal dynamics of customer behavior through recurrent neural networks and combine them with optimized models for real-time data stream analysis;

[0174] S7. Based on the prediction results of the optimization model, collaborative filtering is used to analyze customer historical behavior data and generate personalized product and service recommendation solutions;

[0175] S8. Combine long short-term memory networks to monitor and analyze equipment operation data, predict potential failures, and take maintenance measures in advance.

[0176] In this implementation, S3 includes the following steps:

[0177] S31, extract features from the preprocessed customer data set D′ to generate an initial feature set F, where each feature f i Can reflect specific customer behaviors or preferences;

[0178] S32, using the principal component analysis method to perform dimensionality reduction processing on the initial feature set F, projecting the high-dimensional feature data into a low-dimensional space, retaining the main components, and generating a feature matrix F′ after dimensionality reduction;

[0179] S33. Use the mutual information method to evaluate each feature f i The correlation between the customer churn label y and the mutual information value I(f i ,y);

[0180]

[0181] Among them, p(x,y) represents the joint probability distribution, p(x) and p(y) represent the marginal probability distribution respectively;

[0182] S34, select features based on the mutual information value, set the threshold θ, and retain the mutual information value I(f i ,y)≥θ features, forming the filtered feature set F”;

[0183] S35, construct features for the filtered feature set F" to generate a new combined feature f c ;

[0184] S36, standardize the constructed feature set to generate the final feature set F f ;

[0185] S37, the final feature set F f Combined with the customer data label y, a feature label pair (F f ,y), as the data input for subsequent model training.

[0186] In this implementation, S4 includes the following steps:

[0187] S41, initialize the parameters of the customer churn prediction model, define the neural network structure, including the input layer, multiple hidden layers and the output layer, the input layer receives the feature set F f , the output layer generates the probability of customer churn

[0188] S42. Use elastic weight integration technology to calculate the weight importance of each parameter Ω i , the specific steps include:

[0189] After each task training, the Fisher information matrix is ​​calculated

[0190]

[0191] Among them, p(x|θ i ) is based on the parameter θ i Likelihood function, x is the observed data, E is the expected value of the data distribution;

[0192] The specific form of the likelihood function is:

[0193]

[0194] Among them, x j represents the jth observed data point, and n is the total number of observed data points;

[0195] S43, accumulate the Fisher information matrix and obtain the parameter importance:

[0196]

[0197] S44. During the training process, adjust the loss function to include parameter importance terms and update the neural network parameters:

[0198]

[0199] Among them, L original represents the original loss function, λ represents the regularization coefficient, θ i and Represent the parameters of the current and previous tasks respectively;

[0200] S45. Introduce the enhanced dynamic generative adversarial network, including the construction and training of the generator G and the discriminator D:

[0201] The generator G is based on the noise vector z and the parameter θ g Generate customer churn samples

[0202]

[0203] where z is sampled from a standard normal distribution, θ g Represents the parameters of the generator;

[0204] The discriminator D evaluates the difference between the generated samples and the real samples and calculates the loss:

[0205]

[0206] in, and They represent the samples generated by the discriminator for the real sample x The discriminant output of

[0207] S46, reinforcement learning method optimizes between the generator and the discriminator, and updates the parameters of the generator through the policy gradient method:

[0208]

[0209] Among them, J(θ g ) is the objective function of the generator, Represents the generator parameter θ g The gradient of

[0210] S47, the optimization process of the discriminator D includes using the real customer data x and the generated customer churn samples Update the parameters of the discriminator and minimize the loss function L D ;

[0211] S48, the optimization process of the generator G comprises the following steps:

[0212] Apply the policy gradient method to update the parameters of the generator so that the generated customer churn samples are closer to the real data distribution;

[0213] According to the feedback from the discriminator, the parameters of the generator are gradually adjusted:

[0214]

[0215] Among them, α is the learning rate, θ g Represents the parameters of the generator;

[0216] S49. Combining the advantages of generative adversarial networks and reinforcement learning, the reward mechanism R of reinforcement learning is used to dynamically adjust the objective functions of the generator and discriminator:

[0217] R = E[r t |s t ,a t ];

[0218] Among them, r t is the reward value at time step t, s tis the current state, a t for the actions taken;

[0219] The outputs of the generator and discriminator are calculated using the multi-layer perceptron structure in the generative adversarial network:

[0220]

[0221] Among them, σ is the activation function, W is the weight matrix, b is the bias term, x is the input feature, is the output;

[0222] S410, introduce the time series embedding method to encode the time information into the generator input to achieve time-aware generation:

[0223]

[0224] Where t represents the time step, represents the customer churn sample generated at time t, z t is the noise vector of time step t;

[0225] S411. Use the optimized generator and discriminator to generate more high-quality customer churn samples, and perform mixed training with real customer data;

[0226] S412: Apply the optimized customer churn prediction model to the comprehensive customer data set D' to calculate the churn probability of each customer The results are then used for subsequent analysis and maintenance decisions.

[0227] In this implementation, S5 includes the following steps:

[0228] S51. Construct a heterogeneous graph neural network and define a heterogeneous graph Q=(V,E) containing multiple types of nodes and edges, where V represents a node set, E represents an edge set, and different types of nodes and edges represent different types of customer data and relationships, respectively;

[0229] S52, for each node v i ∈V initializes the feature vector These feature vectors come from the preprocessed feature set F f ;

[0230] S53, performing feature aggregation on the heterogeneous graph through a message passing mechanism, specifically comprising the following steps:

[0231] For each node v on each layer l i , from its neighbor node N(v i ) receives messages and aggregates them:

[0232]

[0233] Among them, σ is the activation function, W h and W e are the weight matrices of node features and edge features, e ij Represents node v i and node v j The edge features between

[0234] For each node v i , update its eigenvector:

[0235]

[0236] in, and They are the weight matrices of the node features and message features of the current layer respectively;

[0237] S54, through multi-layer feature aggregation operations, finally generate the embedding representation h of each node i ;

[0238] S55. Apply deep algebraic geometry learning to optimize node embedding, specifically including the following steps:

[0239] Use algebraic geometry methods to extract high-order relationships between node embeddings, construct algebraic geometry optimization objective functions, and define two algebraic geometry mapping functions f and g as:

[0240] f(h i )=A i h i ;

[0241] g(h j )=B j h j ;

[0242] Among them, A i and B j is the mapping matrix;

[0243] The algebraic geometry optimization objective function is:

[0244]

[0245] Topology optimization is introduced in algebraic geometry space, and the following loss function is constructed through topology structure preservation technology:

[0246]

[0247] Among them, w ij Represents node v i and v j The weights between them are calculated through the topological relationships embedded in the nodes;

[0248] Combining algebraic geometry optimization and topology preservation loss, a comprehensive optimization goal is formed:

[0249] L total =L DAGL +βL TSP ;

[0250] Among them, β is the trade-off coefficient;

[0251] S56. Optimize the comprehensive loss function L by gradient descent method total ,update the parameters of the heterogeneous graph neural network so that the node embeddings better reflect the high-order structure and topological features in the customer relationship data;

[0252] S57, embed the optimized node into h i Used for subsequent customer churn prediction model training;

[0253] S58. Apply the adjusted model structure and parameters to the comprehensive customer data set D′ to calculate the churn probability of each customer The results are then used for subsequent analysis and maintenance decisions.

[0254] In this implementation, S6 includes the following steps:

[0255] S61. Build a recursive neural network model, define the input layer, hidden layer and output layer, and the input layer receives time series data x t , the output layer generates the predicted value for the corresponding time step

[0256] S62, for the input time series data x t Preprocess and standardize the data features of each time step to meet the input requirements of the recurrent neural network model:

[0257] S63. Define the hidden layer state h of the recurrent neural network t The update formula is as follows:

[0258] h t =σ(W h x t +U h h t-1 +b h );

[0259] Among them, σ is the activation function, W h is the input weight matrix, U h is the hidden state weight matrix, b h is the bias vector, x t is the input of the current time step, h t-1is the hidden state of the previous time step;

[0260] S64. Combined with the variational autoencoder to further capture long-term dependencies, the update formula of VAE is defined as follows:

[0261] For the input time series data x t , first generate the latent variable z through the encoder q(z|x) t , the formula is:

[0262] z t =μ+σ⊙∈;

[0263] Where μ and σ represent the mean and standard deviation of the latent variable, respectively, and ∈ is the noise sampled from the standard normal distribution;

[0264] Then the decoder p(x|z) generates the reconstructed time series data The formula is:

[0265]

[0266] Among them, W z is the weight matrix of the decoder, b z is the bias vector of the decoder;

[0267] S65. Use the optimized node embedding to represent h i and time series data x t Combine and dynamically adjust through the following formula:

[0268]

[0269] Among them, W y is the output weight matrix, b y is the output bias vector, is the predicted output at time step t;

[0270] S66. Combine the optimized heterogeneous graph neural network and deep algebraic geometry learning model to combine the prediction results of time series data with node embedding representation to generate comprehensive customer churn prediction results

[0271] S67. Analyze real-time data streams and continuously train and optimize models based on new data;

[0272] S68. Apply the optimized prediction model to the comprehensive customer data set D' to calculate the churn probability of each customer at different time steps The results are then used for subsequent analysis and maintenance decisions.

[0273] In this implementation, S7 includes the following steps:

[0274] S71. Collect the customer's historical behavior data B to form a behavior matrix R, where R ij represents the rating or behavior value of customer i on product j;

[0275] S72. Standardize the behavior matrix R and normalize the score or behavior value to the interval [0,1]:

[0276] S73, applying a collaborative filtering method to perform similarity calculation on the standardized behavior matrix R, specifically comprising the following steps:

[0277] Calculate the similarity between customers using the cosine similarity formula:

[0278]

[0279] Among them, sim(u,v) represents the similarity between customer u and customer v, R uj and R vj They represent the ratings of customer u and customer v on product j respectively;

[0280] S74. Generate a recommendation list for the target customer based on customer similarity. The specific steps are as follows:

[0281] Based on the top k similar customers with the highest similarity, calculate the target customer's predicted rating for the unrated product:

[0282]

[0283] in, represents the predicted rating of target customer i for product j, N k (i) represents the top k customers with the highest similarity to target customer i;

[0284] S75. Combined prediction results To optimize the recommendation list, the specific steps are as follows:

[0285] The probability of customer churn With recommendation rating Combine them to generate an optimized recommendation score:

[0286]

[0287] in, represents the optimized recommendation score, α is the weight coefficient, which represents the balance coefficient between the recommendation score and the churn probability;

[0288] S76. Sort the optimized recommendation scores to generate a personalized recommendation list, and apply the recommendation results to the customer relationship management system to provide personalized product and service recommendation solutions for each customer;

[0289] S77. Continuously optimize and update the recommendation model based on actual feedback data.

[0290] In this implementation, S8 includes the following steps:

[0291] S81. Collect equipment operation data, including sensor data, log data, and historical maintenance records, to form equipment operation data set D e ;

[0292] S82. Run data set D on the device e Perform preprocessing, including data cleaning, missing value filling and standardization;

[0293] S83. Build a long short-term memory network LSTM model, define the input layer, LSTM unit layer and output layer, and the input layer receives time series data x t , the output layer generates the predicted value for the corresponding time step

[0294] S84. Define the update formula of the LSTM unit as follows:

[0295] i t =σ(W i x t +U i h t-1 +b i );

[0296] f t =σ(W f x t +U f h t-1 +b f );

[0297] o t =σ(W o x t +U o h t-1 +b o );

[0298] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c);

[0299] h t =o t ⊙tanh(c t );

[0300] Among them, i t 、f t , o t They represent the input gate, forget gate and output gate respectively, and c t is the unit state, ⊙ represents the element-by-element product operation, W i , W f , W o , W c is the input weight matrix, U i , U f , U o , U c is the hidden state weight matrix, b i , b f , b o , b c is the bias vector;

[0301] S85, the pre-processed equipment operation data x t Input into the LSTM model and obtain the hidden state h of each time step through recursive calculation t and the cell state c t ;

[0302] S86. Combining latent Dirichlet allocation (LDA) to perform topic modeling on the equipment operation data and extract potential failure modes, specifically including the following steps:

[0303] For each document d, its topic distribution θ d Generated by Dirichlet distribution, its formula is:

[0304]

[0305] Among them, α=(α1,α2,…,α K ) is the hyperparameter of the Dirichlet distribution, θ di represents the proportion of topic i in document d, Γ(·) is the Gamma function;

[0306] For each topic k, its word distribution Generated by Dirichlet distribution, its formula is:

[0307]

[0308] Where β=(β1,β2,…,β V ) is the hyperparameter of the Dirichlet distribution, represents the proportion of word j in topic k.

[0309] S87. Use Gibbs sampling method to estimate the parameters of LDA model and obtain the topic distribution θ of each document through sampling. d And the word distribution of each topic

[0310] S88. Prediction results combined with LSTM model and the failure mode θ of the LDA model d and Generate a comprehensive maintenance plan to proactively implement maintenance actions, including the following steps:

[0311] Calculate the failure probability P(f t |x t ):

[0312] P(f t |x t )=σ(W f h t +b f );

[0313] Among them, σ is the activation function, W f is the output weight matrix, b f is the output bias vector;

[0314] According to the failure probability P(f t |x t ) and failure modes of the LDA model Generate maintenance plans, prioritizing high-risk equipment;

[0315] S88. Apply comprehensive maintenance plans to equipment management systems for real-time monitoring and dynamic adjustment, and continuously optimize maintenance strategies by combining new equipment operation data and model prediction results;

[0316] S89. Apply the optimized maintenance plan and equipment failure prediction model to the comprehensive equipment data set D e , calculate the failure probability of each device at different time steps The results are then used for subsequent analysis and maintenance decisions.

[0317] Embodiment 1:

[0318] A large retail company, Company A, is facing problems such as high customer churn rate, low customer satisfaction, and high equipment maintenance costs. The company has a large customer base and complex customer relationship data, including historical purchase records, browsing behavior, feedback information, etc. In order to improve customer retention, optimize personalized recommendations, and reduce maintenance costs, the company decided to introduce a customer relationship management predictive maintenance method based on deep learning.

[0319] Company A first collected a large amount of customer data, including historical behavior data, transaction data, and real-time data. Through data preprocessing technology, these data were cleaned, missing values ​​were filled, and standardized, laying a solid foundation for subsequent feature engineering and model training.

[0320] In the feature engineering stage, Company A conducted in-depth feature extraction and selection on the pre-processed customer data to generate a feature set that can reflect customer behavior and preferences. Through principal component analysis and mutual information methods, features highly correlated with customer churn were screened out, and these features were constructed and standardized to form the final feature set.

[0321] Subsequently, Company A established a customer churn prediction model by combining elastic weight integration with enhanced dynamic generative adversarial networks. Through the elastic weight integration technology, the model maintained parameter stability when facing new tasks, while the enhanced dynamic generative adversarial network generated and adjusted customer churn samples, greatly improving the accuracy and stability of churn prediction.

[0322] In terms of processing complex customer relationship data, Company A adopted heterogeneous graph neural networks and combined them with deep algebraic geometry learning to optimize the model structure and parameters. This step effectively improved the modeling capabilities of multi-type customer data, enabling the model to more accurately reflect the complex relationships between customers.

[0323] In order to capture the dynamic changes of customer behavior in real time, Company A introduced a recursive neural network. Through time series analysis, the network can dynamically monitor changes in customer behavior and combine the aforementioned optimization model to perform real-time data flow analysis, ensuring the timeliness and accuracy of personalized recommendations.

[0324] Based on the results of the above prediction model, Company A combined collaborative filtering methods to conduct an in-depth analysis of customer historical behavior data and generated personalized product and service recommendations, which significantly improved customer satisfaction and loyalty.

[0325] In terms of equipment management, Company A uses long short-term memory networks to monitor and analyze equipment operation data, predict potential failures, and take maintenance measures in advance. This step not only reduces maintenance costs, but also improves equipment operation efficiency and reliability.

[0326] Concrete data proves our beneficial effects

[0327] By applying the above methods, Company A achieved significant results in a short period of time. The following are specific data proofs:

[0328] Table 1 Data table

[0329]

[0330] It can be seen from the above embodiments that the present invention effectively solves the problems of customer churn, low customer satisfaction and high equipment maintenance costs, and significantly improves the operational efficiency and customer satisfaction of the enterprise.

[0331] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A customer relationship management predictive maintenance method based on deep learning, characterized in that: The following steps are involved: S1. Collect customer data, including historical behavior data, transaction data and real-time data; S2. Preprocess the collected customer data, including data cleaning, missing value filling and standardization; S3. Perform feature engineering on the preprocessed customer data to extract and select features that can reflect customer behaviors and preferences; S4. A customer churn prediction model is established by combining elastic weight integration and enhanced dynamic generative adversarial network. The elastic weight integration is used to maintain parameter stability, and the enhanced dynamic generative adversarial network is used to generate and adjust customer churn samples. S5. Use heterogeneous graph neural networks to process multi-type customer relationship data, and adjust model structure and parameters in combination with deep algebraic geometry learning; S6. Capture the temporal dynamics of customer behavior through recurrent neural networks and combine them with optimized models for real-time data stream analysis; S7. Based on the prediction results of the optimization model, collaborative filtering is used to analyze customer historical behavior data and generate personalized product and service recommendation solutions; S8. Combine long short-term memory networks to monitor and analyze equipment operation data, predict potential failures, and take maintenance measures in advance.

2. The customer relationship management prediction and maintenance method based on deep learning according to claim 1, characterized in that: The S3 comprises the following steps: S31, extract features from the preprocessed customer data set D′ to generate an initial feature set F, where each feature f i Can reflect specific customer behaviors or preferences; S32, using the principal component analysis method to perform dimensionality reduction processing on the initial feature set F, projecting the high-dimensional feature data into a low-dimensional space, retaining the main components, and generating a feature matrix F′ after dimensionality reduction; S33. Use the mutual information method to evaluate each feature f i The correlation between the customer churn label y and the mutual information value I(f i ,y); Among them, p(x,y) represents the joint probability distribution, p(x) and p(y) represent the marginal probability distribution respectively; S34, select features based on the mutual information value, set the threshold θ, and retain the mutual information value I(f i ,y)≥θ features, forming the filtered feature set F”; S35, construct features for the filtered feature set F" to generate a new combined feature f c ; S36, standardize the constructed feature set to generate the final feature set F f ; S37, the final feature set F f Combined with the customer data label y, a feature label pair (F f ,y), as the data input for subsequent model training.

3. The customer relationship management predictive maintenance method based on deep learning according to claim 1, characterized in that: The S4 comprises the following steps: S41, initialize the parameters of the customer churn prediction model, define the neural network structure, including the input layer, multiple hidden layers and the output layer, the input layer receives the feature set F f , the output layer generates the probability of customer churn S42. Use elastic weight integration technology to calculate the weight importance of each parameter Ω i , the specific steps include: After each task training, the Fisher information matrix is ​​calculated Among them, p(x|θ i ) is based on the parameter θ i Likelihood function, x is the observed data, E is the expected value of the data distribution; The specific form of the likelihood function is: Among them, x j represents the jth observed data point, and n is the total number of observed data points; S43, accumulate the Fisher information matrix and obtain the parameter importance: S44. During the training process, adjust the loss function to include parameter importance terms and update the neural network parameters: Among them, L original represents the original loss function, λ represents the regularization coefficient, θ i and Represent the parameters of the current and previous tasks respectively; S45. Introduce the enhanced dynamic generative adversarial network, including the construction and training of the generator G and the discriminator D: The generator G is based on the noise vector z and the parameter θ g Generate customer churn samples where z is sampled from a standard normal distribution, θ g Represents the parameters of the generator; The discriminator D evaluates the difference between the generated samples and the real samples and calculates the loss: Among them, D(x) and They represent the samples generated by the discriminator for the real sample x The discriminant output of S46, reinforcement learning method optimizes between the generator and the discriminator, and updates the parameters of the generator through the policy gradient method: Among them, J(θ g ) is the objective function of the generator, Represents the generator parameter θ g The gradient of S47, the optimization process of the discriminator D includes using the real customer data x and the generated customer churn samples Update the parameters of the discriminator and minimize the loss function L D ; S48, the optimization process of the generator G comprises the following steps: Apply the policy gradient method to update the parameters of the generator so that the generated customer churn samples are closer to the real data distribution; According to the feedback from the discriminator, the parameters of the generator are gradually adjusted: Among them, α is the learning rate, θ g Represents the parameters of the generator; S49. Combining the advantages of generative adversarial networks and reinforcement learning, the reward mechanism R of reinforcement learning is used to dynamically adjust the objective functions of the generator and discriminator: R=E[r t |s t ,a t ]; Among them, r t is the reward value at time step t, s t is the current state, a t for the actions taken; The outputs of the generator and discriminator are calculated using the multi-layer perceptron structure in the generative adversarial network: Among them, σ is the activation function, W is the weight matrix, b is the bias term, x is the input feature, is the output; S410, introduce the time series embedding method to encode the time information into the generator input to achieve time-aware generation: Where t represents the time step, represents the customer churn sample generated at time t, z t is the noise vector of time step t; S411. Use the optimized generator and discriminator to generate more high-quality customer churn samples, and perform mixed training with real customer data; S412: Apply the optimized customer churn prediction model to the comprehensive customer data set D' to calculate the churn probability of each customer The results are then used for subsequent analysis and maintenance decisions.

4. The customer relationship management predictive maintenance method based on deep learning according to claim 1, characterized in that: The S5 comprises the following steps: S51. Construct a heterogeneous graph neural network and define a heterogeneous graph Q=(V,E) containing multiple types of nodes and edges, where V represents a node set, E represents an edge set, and different types of nodes and edges represent different types of customer data and relationships, respectively. S52, for each node v i ∈V initializes the feature vector These feature vectors come from the preprocessed feature set F f ; S53, performing feature aggregation on the heterogeneous graph through a message passing mechanism, specifically comprising the following steps: For each node v on each layer l i , from its neighbor node N(v i ) receives messages and aggregates them: Among them, σ is the activation function, W h and W e are the weight matrices of node features and edge features, e ij Represents node v i and node v j The edge features between For each node v i , update its feature vector: in, and They are the weight matrices of the node features and message features of the current layer respectively; S54, through multi-layer feature aggregation operations, finally generate the embedding representation h of each node i ; S55. Apply deep algebraic geometry learning to optimize node embedding, specifically including the following steps: Use algebraic geometry methods to extract high-order relationships between node embeddings, construct algebraic geometry optimization objective functions, and define two algebraic geometry mapping functions f and g as: f(h i )=A i h i ; g(h j )=B j h j ; Among them, A i and B j is the mapping matrix; The algebraic geometry optimization objective function is: Topology optimization is introduced in algebraic geometry space, and the following loss function is constructed through topology structure preservation technology: Among them, w ij Represents node v i and v j The weights between them are calculated through the topological relationships embedded in the nodes; Combining algebraic geometry optimization and topology preservation loss, a comprehensive optimization goal is formed: L total =L DAGL +βL TSP ; Among them, β is the trade-off coefficient; S56. Optimize the comprehensive loss function L by gradient descent method total ,update the parameters of the heterogeneous graph neural network so that the node embeddings better reflect the high-order structure and topological features in the customer relationship data; S57, embed the optimized node into h i Used for subsequent customer churn prediction model training; S58. Apply the adjusted model structure and parameters to the comprehensive customer data set D′ to calculate the churn probability of each customer The results are then used for subsequent analysis and maintenance decisions.

5. The customer relationship management predictive maintenance method based on deep learning according to claim 1, characterized in that: The S6 comprises the following steps: S61. Build a recursive neural network model, define the input layer, hidden layer and output layer, and the input layer receives time series data x t , the output layer generates the predicted value for the corresponding time step S62, for the input time series data x t Preprocess and standardize the data features of each time step to meet the input requirements of the recurrent neural network model: S63. Define the hidden layer state h of the recurrent neural network t The update formula is as follows: h t =σ(W h x t +U h h t-1 +b h ); Among them, σ is the activation function, W h is the input weight matrix, U h is the hidden state weight matrix, b h is the bias vector, x t is the input of the current time step, h t-1 is the hidden state of the previous time step; S64. Combined with the variational autoencoder to further capture long-term dependencies, the update formula of VAE is defined as follows: For the input time series data x t , first generate the latent variable z through the encoder q(z|x) t , the formula is: With t =μ+σ⊙∈; Where μ and σ represent the mean and standard deviation of the latent variable, respectively, and ∈ is the noise sampled from the standard normal distribution; Then the decoder p(x|z) generates the reconstructed time series data The formula is: Among them, W z is the weight matrix of the decoder, b z is the bias vector of the decoder; S65. Use the optimized node embedding to represent h i and time series data x t Combine and dynamically adjust through the following formula: Among them, W y is the output weight matrix, b y is the output bias vector, is the predicted output at time step t; S66. Combine the optimized heterogeneous graph neural network and deep algebraic geometry learning model to combine the prediction results of time series data with node embedding representation to generate comprehensive customer churn prediction results S67. Analyze real-time data streams and continuously train and optimize models based on new data; S68. Apply the optimized prediction model to the comprehensive customer data set D' to calculate the churn probability of each customer at different time steps The results are then used for subsequent analysis and maintenance decisions.

6. The customer relationship management predictive maintenance method based on deep learning according to claim 1, characterized in that: The S7 comprises the following steps: S71. Collect the customer's historical behavior data B to form a behavior matrix R, where R ij represents the rating or behavior value of customer i on product j; S72. Standardize the behavior matrix R and normalize the score or behavior value to the interval [0,1]: S73, applying a collaborative filtering method to perform similarity calculation on the standardized behavior matrix R, specifically comprising the following steps: Calculate the similarity between customers using the cosine similarity formula: Among them, sim(u,v) represents the similarity between customer u and customer v, R uj and R vj They represent the ratings of customer u and customer v on product j respectively; S74. Generate a recommendation list for the target customer based on customer similarity. The specific steps are as follows: Based on the top k similar customers with the highest similarity, calculate the target customer's predicted rating for the unrated product: in, represents the predicted rating of target customer i for product j, N k (i) represents the top k customers with the highest similarity to target customer i; S75. Combined prediction results To optimize the recommendation list, the specific steps are as follows: The probability of customer churn With recommendation rating Combine them to generate an optimized recommendation score: in, represents the optimized recommendation score, α is the weight coefficient, which represents the balance coefficient between the recommendation score and the churn probability; S76. Sort the optimized recommendation scores to generate a personalized recommendation list, and apply the recommendation results to the customer relationship management system to provide personalized product and service recommendation solutions for each customer; S77. Continuously optimize and update the recommendation model based on actual feedback data.

7. The customer relationship management prediction and maintenance method based on deep learning according to claim 1, characterized in that: The S8 comprises the following steps: S81. Collect equipment operation data, including sensor data, log data, and historical maintenance records, to form equipment operation data set D e ; S82. Run data set D on the device e Perform preprocessing, including data cleaning, missing value filling and standardization; S83. Build a long short-term memory network LSTM model, define the input layer, LSTM unit layer and output layer, and the input layer receives time series data x t , the output layer generates the predicted value for the corresponding time step S84. Define the update formula of the LSTM unit as follows: i t =σ(W i x t +U i h t-1 +b i ); f t =σ(W f x t +U f h t-1 +b f ); the t =σ(W o x t +U o h t-1 +b o ); c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c ); h t =o t ⊙tanh(c t ); Among them, i t 、f t , o t They represent the input gate, forget gate and output gate respectively, and c t is the unit state, ⊙ represents the element-by-element product operation, W i , W f , W o , W c is the input weight matrix, U i , U f , U o , U c is the hidden state weight matrix, b i , b f , b o , b c is the bias vector; S85, the pre-processed equipment operation data x t Input into the LSTM model and obtain the hidden state h of each time step through recursive calculation t and the cell state c t ; S86. Combining latent Dirichlet allocation (LDA) to perform topic modeling on the equipment operation data and extract potential failure modes, specifically including the following steps: For each document d, its topic distribution θ d Generated by Dirichlet distribution, its formula is: Among them, α=(α1,α2,…,α K ) is the hyperparameter of the Dirichlet distribution, θ di represents the proportion of topic i in document d, Γ(·) is the Gamma function; For each topic k, its word distribution Generated by Dirichlet distribution, its formula is: Where β=(β1,β2,…,β V ) is the hyperparameter of the Dirichlet distribution, represents the proportion of word j in topic k. S87. Use Gibbs sampling method to estimate the parameters of LDA model and obtain the topic distribution θ of each document through sampling. d And the word distribution of each topic S88. Prediction results combined with LSTM model and the failure mode θ of the LDA model d and Generate a comprehensive maintenance plan to proactively implement maintenance actions, including the following steps: Calculate the failure probability P(f t |x t ): P(f t |x t )=σ(W f h t +b f ); Among them, σ is the activation function, W f is the output weight matrix, b f is the output bias vector; According to the failure probability P(f t |x t ) and failure modes of the LDA model Generate maintenance plans, prioritizing high-risk equipment; S88. Apply comprehensive maintenance plans to equipment management systems for real-time monitoring and dynamic adjustment, and continuously optimize maintenance strategies by combining new equipment operation data and model prediction results; S89. Apply the optimized maintenance plan and equipment failure prediction model to the comprehensive equipment data set D e , calculate the failure probability of each device at different time steps The results are then used for subsequent analysis and maintenance decisions.

Images