Power material warehousing behavior identification method, system, device and storage medium

Abstract

The application relates to the technical field of warehouse behavior recognition, and discloses a power material warehouse behavior recognition method, system, device and storage medium, which comprises the following steps: obtaining radio frequency data and work order data of to-be-recognized materials in a power material warehouse, and performing pretreatment to obtain radio frequency features and intervention actions; a time sequence hidden state coding module constructed based on a causal time sequence autoencoder is used to perform hidden state coding on the radio frequency features to obtain observation hidden states; the observation hidden states, the intervention actions and a double-constraint boundary are input into a causal dynamic world model to obtain a future multi-step hidden state sequence; and the future multi-step hidden state sequence is input into a double-task parallel decoding module to obtain a position prediction result and a behavior prediction result. Through the world model of the double-constraint endogenous embedding causal time sequence architecture, high-precision material position prediction, behavior recognition and business compliance judgment can be realized, so that the intelligent management and control level and the safety production capacity of the power material warehouse are improved.

Description

Technical Field

[0001] This invention relates to the field of warehouse behavior recognition technology, and in particular to a method, system, device and storage medium for recognizing warehouse behavior of power materials. Background Technology

[0002] With the advancement of digital transformation in the power industry, RFID (Radio Frequency Identification) technology has been widely applied to the location and lifecycle management of power materials in warehouses. Power materials are mostly made of metal and are large and heavy-duty. Warehouses often have dense shelving and significant metal obstructions, making RFID signals prone to multipath effects, packet loss due to obstruction, and random jumps, resulting in insufficient location accuracy and discontinuous tracking. Simultaneously, power material warehousing management places extremely high demands on the compliance of material movement and the timeliness of early warnings of abnormal behavior, which traditional location and control methods can no longer meet.

[0003] Most existing positioning methods employ models such as Long Short-Term Memory (LSTM) networks and graph convolutional networks, which only model the correlation of RFID time-series data. They struggle to capture physical semantics or causal mechanisms, and their generalization ability is poor, especially when the signal is affected by multipath propagation, obstruction, or interference. Furthermore, existing solutions either rely on manual post-processing filtering of rules after prediction or require pre-defined fixed business process rules, which can easily lead to invalid predictions that do not conform to business specifications. At the same time, manually preset rules lack flexibility, requiring simultaneous modification when warehouse business processes are adjusted, resulting in high maintenance costs and an inability to adapt to the personalized business needs of different warehouses. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method, system, device, and storage medium for identifying the warehousing behavior of power materials, enabling accurate prediction of material location and compliant identification of movement behavior, thereby achieving the technical effect of high reliability, strong compliance, and intelligent management and control requirements for power material warehousing.

[0005] In a first aspect, the present invention provides a method for identifying power material storage behavior, the method comprising: The radio frequency data and work order data of the materials to be identified in the power materials warehouse are acquired and preprocessed to obtain radio frequency characteristics and intervention actions. The radio frequency features are encoded using a time-series hidden state coding module based on a causal time-series autoencoder to obtain the observed hidden state. The observed hidden state, the intervention action, and the preset dual-constraint boundary are input into a preset causal dynamic world model. Multiple rounds of hidden state prediction are performed through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on a constrained causal Transformer. The state transition function has embedded endogenous constraints, including physical space constraints and business behavior compliance constraints. The future multi-step hidden state sequence is input into a dual-task parallel decoding module to obtain position prediction results and behavior prediction results. The dual-task parallel decoding module includes a position prediction decoding head and a behavior prediction decoding head.

[0006] Furthermore, the step of acquiring the radio frequency data and work order data of the materials to be identified in the power material warehouse, and performing preprocessing to obtain radio frequency timing characteristics and intervention actions includes: Acquire radio frequency data and work order data of materials to be identified in the power material warehouse, wherein the materials to be identified include materials associated with work orders and materials not associated with work orders; The radio frequency data is standardized and preprocessed to obtain radio frequency characteristics; An intervention action is generated based on the work order status of the work order data.

[0007] Furthermore, the temporal hidden state coding module consists of multiple layers of first causal Transformer encoders, and each layer of first causal Transformer encoder is equipped with a self-attention layer with multiple heads and lower triangular causal masks. The position encoding of the temporal hidden state encoding module adopts a fusion position encoding of time relative position encoding and reader spatial coordinate embedding, wherein the reader spatial coordinate embedding is obtained based on the real physical position mapping of the reader.

[0008] Furthermore, the state transition function of the causal dynamic world model is composed of multiple layers of second causal Transformer encoders. Each layer of the second causal Transformer encoder block is composed of cascaded cross-attention units, dual-constraint gating units, feedforward neural networks, layer normalization and residual connection units. The cross-attention unit is used to calculate and output a context vector based on the hidden state and the intervention action through cross-attention. The dual-constraint gating unit includes physical constraint gating and business constraint gating. The physical constraint gating is used to filter illegal physical paths of the context vector according to the physical space constraints, and the business constraint gating is used to filter illegal business paths of the filtered context vector according to the business behavior compliance constraints. The physical space constraint is a constraint matrix based on the accessibility boundary within the warehouse, and the business behavior compliance constraint is the similarity between the time-series trajectory of the most recent time step and each compliance behavior prototype. The compliance behavior prototype is obtained by clustering analysis of historical compliance trajectory features.

[0009] Furthermore, the step of inputting the observed hidden state, the intervention action, and the preset dual-constraint boundary into a preset causal dynamic world model, and performing multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence includes: Using the posterior hidden state of the previous moment, the intervention action of the current moment, and the double-constraint boundary as a causal dynamic world model, the prior hidden state of the current moment is obtained. Determine whether the radio frequency data is valid data. If not, use the prior hidden state at the current moment as the subsequent hidden state at the current moment. If so, then calculate the Kalman gain based on the observed hidden state and the prior hidden state at the current time, and fuse the observed hidden state and the prior hidden state at the current time based on the Kalman gain to obtain the posterior hidden state at the current time. The posterior hidden state at the current moment, the intervention action at the next moment, and the double-constraint boundary are input into the causal dynamic world model. Autoregressive multi-round prediction is performed according to the preset step size to obtain the future multi-step hidden state sequence.

[0010] Furthermore, the location prediction decoding head adopts a 3-layer fully connected network, and the output location prediction results include the current material location, the material location prediction sequence for future times, and the location prediction confidence sequence. The behavior prediction decoding head consists of a cascaded two-layer fully connected network and a Softmax activation function. The output behavior prediction results include the current behavior classification result, the future behavior prediction sequence, and the risk level.

[0011] Furthermore, the loss function of the causal dynamic world model includes multi-step dynamic prediction loss, causal consistency loss, constraint compliance loss, and KL divergence constraint loss; The multi-step dynamic prediction loss is constructed based on the mean squared error loss of future multiple steps, the causal consistency loss is constructed based on the contrastive learning loss, and the constraint compliance loss includes physical constraint loss and business compliance loss.

[0012] Secondly, the present invention provides a power material warehousing behavior identification system, the system comprising: The data processing module is used to acquire radio frequency data and work order data of materials to be identified in the power material warehouse, and to preprocess them to obtain radio frequency characteristics and intervention actions. The data encoding module is used to encode the radio frequency features using a time-series hidden state encoding module based on a causal time-series autoencoder to obtain the observed hidden state. The state prediction module is used to input the observed hidden state, the intervention action, and the preset dual-constraint boundary into a preset causal dynamic world model, and perform multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on the constrained causal Transformer. The state transition function has embedded endogenous constraints, including physical space constraints and business behavior compliance constraints. The behavior recognition module is used to input the future multi-step hidden state sequence into the dual-task parallel decoding module to obtain the position prediction result and the behavior prediction result. The dual-task parallel decoding module includes a position prediction decoding head and a behavior prediction decoding head.

[0013] Thirdly, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0015] This invention provides a method, system, device, and storage medium for identifying the behavior of power material storage. By intrinsically embedding physical space constraints and business compliance features into the world model kernel of a causal temporal architecture, this invention improves the adaptability of the model. Through a causal modeling mechanism with decoupled intervention states and multi-step joint loss training, it can achieve high-precision prediction of material location in extreme scenarios, enabling accurate behavior identification and business compliance judgment. This invention can be quickly adapted to multiple types of power warehouses and replicated on a large scale, thereby improving the intelligent management and control level and safe production capability of power material storage. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method for identifying power material storage behavior in an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of the power material warehousing behavior recognition system in an embodiment of the present invention; Figure 3 This is an internal structural diagram of the computer device in an embodiment of the present invention.

[0017] Figure label: 10. Data processing module; 20. Data encoding module; 30. State prediction module; 40. Behavior recognition module. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The first embodiment of the present invention proposes a method for identifying the storage behavior of power materials, including steps S10 to S40: Step S10: Obtain radio frequency data and work order data of the materials to be identified in the power materials warehouse, and perform preprocessing to obtain radio frequency characteristics and intervention actions; Step S20: The radio frequency features are encoded using a time-series hidden state coding module based on a causal time-series autoencoder to obtain the observed hidden state. Step S30: Input the observed hidden state, the intervention action, and the preset dual-constraint boundary into the preset causal dynamic world model, and perform multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on the constrained causal Transformer. The state transition function has embedded endogenous constraints, including physical space constraints and business behavior compliance constraints. Step S40: Input the future multi-step hidden state sequence into the dual-task parallel decoding module to obtain the position prediction result and the behavior prediction result. The dual-task parallel decoding module includes a position prediction decoding head and a behavior prediction decoding head.

[0020] This embodiment employs Radio Frequency Identification (RFID) technology. Multiple RFID readers deployed in the power material warehouse and RFID tags bound to each item enable the location tracking and anomaly detection of power materials' movement. In this embodiment, the power material to be identified is any item in the warehouse that bears an RFID tag. This item can be associated with a work order or not. Taking any item in the warehouse as an example, the method provided in this embodiment will be described in detail.

[0021] First, RFID technology is used to acquire radio frequency (RF) data of the materials to be identified. This RF data is time-series data, including the unique code of the material tag, the reader ID, RSSI signal strength, phase value, and millisecond-level timestamps. Standardized preprocessing, including noise removal and normalization, is then performed to obtain a fixed-dimensional time-series feature sequence, i.e., RF features. Taking data length L as an example, the RF features at the current time t include RF data from time step t-L+1 to time step t.

[0022] Simultaneously, the work order data for the material at the current time t is retrieved from the warehouse management system. Based on the real-time work order status, the work order actions are categorized to obtain intervention actions. The intervention action is the sole causal variable driving the material's state transition. According to the work order status, actions can be categorized as inbound operations, outbound operations, transfer operations, inventory operations, requisition and return, inactive state, and abnormal movement. When there is no work order, the default is inactive state. Intervention actions are represented in the form of embedded vectors.

[0023] The processed radio frequency features are encoded using a temporal hidden state encoding module built on a causal temporal autoencoder to obtain the observed hidden state. In this embodiment, the temporal hidden state encoder is constructed using a 3-layer causal Transformer encoder. Each layer is configured with an 8-head self-attention layer with a lower triangular causal mask. The role of the lower triangular causal mask in the attention mechanism is to force autoregressive generation, ensuring that each position can only focus on itself and previous positions, preventing "peeking" at future information. Its main diagonal and lower elements are "allowed to pay attention" (usually represented as 0 or True), and the elements above the main diagonal are "masked of the future" (usually represented as -∞ or False), ensuring that the feature encoding at each time step can only use current and historical temporal information, and preventing the leakage of future information.

[0024] In a preferred embodiment, the position encoding of the temporal hidden state encoding module employs fused position encoding. This fused position encoding is obtained by weighted fusion of temporal relative position encoding and reader spatial coordinate embedding, replacing traditional sinusoidal position encoding and achieving a joint representation of temporal order and spatial location. Specifically, temporal relative position encoding is used to represent the relative distance between any two moments in the temporal sequence. The relative position encoding scheme proposed by Transformer-XL is used for temporal relative position encoding. The specific encoding steps can be referenced from conventional relative position encoding schemes and will not be elaborated upon here.

[0025] Reader spatial coordinate embedding is used to map the real physical location of RFID readers to the feature space of the model, realizing endogenous encoding of spatial location information. Specifically, the real physical coordinates of all fixed RFID readers in the warehouse are pre-measured and normalized. Then, a learnable spatial coordinate embedding matrix is ​​multiplied with the normalized two-dimensional coordinates to map the normalized two-dimensional coordinates into a vector with the same dimension as the temporal relative position encoding. The spatial coordinate embedding matrix can be optimized during model training. For each time step of the input sequence, i.e., the input RFID features, the corresponding spatial coordinate embedding vector is obtained according to the reader ID read at each time step, forming a spatial location embedding sequence. Finally, the temporal relative position encoding and the spatial coordinate embedding are weighted and fused to obtain the final fused location encoding. Preferably, the fusion weight of the temporal relative position encoding is preset to 0.6, which can be flexibly adjusted according to the temporal / spatial importance of different warehouse scenarios in practical applications.

[0026] The observed hidden states are obtained by encoding radio frequency features using a pre-trained temporal hidden state encoding module. The observed hidden states, intervention actions, and pre-set double-constraint boundaries are then input into a causal dynamic world model. Through multiple rounds of hidden state prediction, a sequence of future hidden states is obtained.

[0027] In the field of artificial intelligence, a world model refers to a modeling mechanism by which an intelligent agent models the internal dynamics of its environment. The goal is to enable the agent not only to perceive the current state but also to predict future states. A world model can be represented as a function mapping system. This system takes the environmental state and actions at the current time step as input and predicts the environmental state at the next time step through a state transition function f. By learning the functional form of f, the agent can internally simulate environmental evolution. The theoretical foundation of world models is closely related to Markov Decision Processes (MDPs). Within the MDP framework, the environment can be described as a quadruple: state space, action space, state transition probabilities, and immediate reward function. The core objective of world models is to achieve effective prediction and planning of environmental dynamics by learning approximate forms of state transition probabilities and state representations.

[0028] The causal dynamic world model in this embodiment is a world model with dual-constraint endogenous embedding, and its state transition function is defined as: in, Let be the hidden state at time step t. For the intervention action at time step t, It is a double-constraint boundary variable, which is formed by concatenating the physical space constraint vector and the business behavior compliance constraint vector; The constrained causal state transition function is defined by learnable parameters. Composition. It should be noted here that the radio frequency data corresponding to time step t in the causal dynamic world model is the data at time t. For ease of description, in the causal dynamic world model, the hidden state corresponding to the radio frequency characteristics at time t is described as the hidden state at time step t.

[0029] In this embodiment, the hidden state dimension is fixed at 256 dimensions. It is obtained by updating the hidden state output by the encoder in conjunction with the intervention action and the dual-constraint variables. It can be understood as completely containing all information about the spatial location, movement dynamics, and operational status of the materials. A causal modeling mechanism with intervention-state decoupling is adopted to decouple the intervention action... Completely decoupled from hidden states, defined It is the only causal variable driving the transition of material status.

[0030] This embodiment designs a dual-constraint endogenous embedding mechanism, directly embedding physical space constraints and business behavior compliance constraints into the state transition function, rather than post-processing correction, and dual-constraint boundary variables. This is a fixed-dimensional vector, composed of a physical space constraint vector and a business behavior compliance constraint vector. Specifically, the physical space constraints characterize the accessibility boundaries within the warehouse and are expressed in a rasterized matrix form. First, the warehouse plane is rasterized according to a preset resolution (e.g., 0.2m × 0.2m), and then the physical constraint matrix is ​​constructed. Each element in the matrix corresponds to the accessibility score of the corresponding grid cell. For accessible areas (such as passageways and work areas), the matrix element at that location is 1; for impassable areas (such as shelves, walls, and obstacles), the matrix element at that location is 0; and for restricted access areas (such as the boundary of a hazardous materials isolation zone), the matrix element at that location can be set to a value between 0 and 1, such as 0.2. Finally, the physical constraint matrix is ​​flattened into a one-dimensional vector as the physical space constraint variable. This constraint can be updated in real time during model execution.

[0031] Business behavior compliance constraints are constructed based on a historical compliant business time-series trajectory feature library for the warehouse. First, all compliant business trajectory data from the past two years are retrieved from the warehouse management system. Trajectory data refers to continuous trajectory data composed of radio frequency time-series data of the same material linked by its unique label code. The length of each trajectory is extracted. Temporal feature vectors (32 is the feature dimension of a single time step, then) (Corresponding to 2-second historical data), thus obtaining all historical compliant trajectory features. Then, the K-Means clustering algorithm is used to cluster all historical compliant trajectory features, and the number of clusters is... The number of business types is determined accordingly; taking the aforementioned 7 types of intervention actions as an example, K=7. The center vector of each cluster is calculated as the prototype of the compliant behavior for that business type. Finally, all compliant behavior prototypes are concatenated into a business compliance feature library. After building the business compliance feature library, for the current time step t, the trajectory features (i.e., the radio frequency features of the L most recent time steps) are taken, and their cosine similarity with each compliant behavior prototype is calculated. The vector composed of these K similarities is used as the business behavior compliance constraint. Finally, the physical space constraint vector and the business behavior compliance constraint vector are concatenated to obtain the final double-constraint boundary variables: This variable serves as the input to the state transition function and is updated in real time during each state transition step.

[0032] The state transition function in this embodiment This is achieved through a four-layer constrained causal Transformer encoder. Each layer consists of a cross-attention unit, a dual-constraint gating unit, a feedforward neural network, and a layer normalization and residual connection unit. The cross-attention unit takes as input data the hidden state and the intervention action, using the hidden state as the query and the intervention action as the key and value. A context vector is obtained through the cross-attention mechanism, which models the causal relationship between the intervention action and the state change, thus effectively eliminating false associations caused by RFID signal fluctuations.

[0033] The dual-constraint gating unit is the core unit of constraint-inherent embedding. During state transitions, it directly filters out illegal paths that do not conform to physical / business constraints. The dual-constraint gating unit consists of two parts: physical constraint gating and business constraint gating. The physical constraint gating filters out context vector components where the predicted location falls into an impassable region based on physical spatial constraints. Specifically, it decodes the predicted location probability distribution map from the context vector. In the formula, This is a probability map of location distribution. For context vectors, It is a fully connected layer. It is a normalized exponential function.

[0034] Then calculate the physical constraint gating weight vector: in, This represents element-wise dot product. It is the Sigmoid activation function. As a learnable linear transformation layer, it outputs a gated weight vector with the same dimension as the context vector. .

[0035] Finally, physical constraint filtering is performed on the context vector: in, This is the context vector after filtering by physical constraints.

[0036] Preferably, when When the average value is less than 0.3, it is determined that the predicted location falls entirely into an impassable area, and the location is directly... Set to 0 to block this illegal transfer path.

[0037] Business constraint gating, based on business compliance constraints, also uses vector-based gating weights to selectively filter the context vector, filtering out feature components that deviate significantly from historical compliance behavior patterns. The specific implementation steps are as follows: Context vector filtered from physical constraints Decode the current behavior feature vector: in, This is the current behavior feature vector. A fully connected layer with learnable parameters; Then, calculate the similarity vector between the current behavior feature vector and each prototype in the business compliance feature library: in, The k-th compliant behavior prototype was obtained through historical data clustering. The similarity between the current behavior feature vector and the kth compliant behavior prototype, where K is the total number of compliant behavior prototypes; Then calculate the business constraint gating weight vector: in, As a learnable linear transformation layer, it outputs a gated weight vector with the same dimension as the context vector. .

[0038] Finally, the context vector is filtered according to business constraints: when When τ is the similarity threshold, such as 0.5, the prediction behavior is judged to be non-compliant and is directly... Set to 0 to block this illegal transfer path.

[0039] The vector filtered by double-constraint gating is input into a feedforward neural network. After layer normalization and residual connections, the final result of the constrained causal Transformer encoder is output. For a 4-layer constrained causal Transformer, the intervention action is a key and value, independently queried by the cross-attention of each layer. The double-constraint variables serve as gating conditions, independently computed in the gating unit of each layer. The output of the previous encoder layer becomes the input of the next encoder layer, and the output of the last encoder layer is the hidden state. Finally, the hidden state is probabilistically modeled as a Gaussian distribution with diagonal covariance.

[0040] Based on the above-described causal dynamic world model architecture, this embodiment uses the autoregressive approach of the causal dynamic world model to predict hidden states in multiple rounds, obtaining the future multi-step hidden state sequence through the following steps: Using the posterior hidden state of the previous moment, the intervention action of the current moment, and the double-constraint boundary as a causal dynamic world model, the prior hidden state of the current moment is obtained. Based on the probability distribution of the observed hidden state, determine whether the radio frequency data is valid. If not, take the prior hidden state at the current moment as the posterior hidden state at the current moment. Conversely, based on the observed hidden state and the prior hidden state at the current time, the Kalman gain is calculated, and based on the Kalman gain, the observed hidden state and the prior hidden state at the current time are fused to obtain the posterior hidden state at the current time. The posterior hidden state at the current moment, the intervention action at the next moment, and the double-constraint boundary are input into the causal dynamic world model. Autoregressive multi-round prediction is performed according to the preset step size to obtain the future multi-step hidden state sequence.

[0041] In this embodiment, the future multi-step hidden state sequence is obtained by the world model through multiple rounds of prediction using an autoregressive approach. Each round of prediction outputs a hidden state. During a certain round of iterative prediction, the posterior hidden state of the previous time step is used as the basis for prediction. and intervention actions at the current moment and double-constrained boundary variables As the initial input to the causal dynamic world model, if it is the first round, the posterior hidden state is a preset value. The prior hidden state at the current moment is predicted by the state transition function of the causal dynamic world model. ,in, The mean of the prior hidden states. Let I be the variance of the prior hidden state, I be the identity matrix, and N(*) denote the Gaussian distribution.

[0042] Then, the corresponding hidden state update method is selected based on whether the RFID signal is valid data. When the RFID signal is normal (the current time step is valid RFID data), the hidden state is updated using a weighted fusion method. When the RFID signal is blocked or lost (there is no valid RFID data in the current time step), the hidden state is updated using a rolling prediction method.

[0043] Specifically, the first step is to determine whether the radio frequency (RF) signal at the current time step is valid. As can be seen from the above embodiment, the data at time step t input to the timing hidden state encoding module is a sliding window timing data, which contains the RF data of the most recent L time steps. When determining whether the RF information at the current time step is valid, it can be determined by the RSSI signal strength of the RF signal. For example, if the signal strength is weak or there is no response, the signal is considered invalid. Only when the signal strength is greater than the threshold is the signal considered valid. To simplify the calculation, this embodiment only uses the signal strength of the RF signal at the current moment in the sliding window to determine whether the signal is valid. Of course, the signal validity can also be determined by comparing the packet loss rate within the sliding window with the packet loss rate threshold.

[0044] If the radio frequency data is invalid, then the model outputs the prior hidden state at the current time. As the posterior hidden state at the current moment Conversely, based on the observed hidden state and the prior hidden state at the current time, the Kalman gain is calculated. Specifically, the observed hidden state is obtained by encoding the RFID observation data (i.e., radio frequency features) at the current time t using a causal temporal autoencoder. ,in, The mean of the observed hidden states, The variance of the observed hidden states.

[0045] Then, the Kalman gain is calculated based on the variance of the observed hidden state and the variance of the prior hidden state: In the formula, Let be the Kalman gain at time t. Let V be the variance of the prior hidden state. The variance of the observed hidden states.

[0046] Based on the Kalman gain, the prior hidden state and the observed hidden state are fused to obtain the Gaussian distribution of the posterior hidden state at the current time step: In the formula, Let be the mean of the posterior hidden states at time t. Let be the variance of the posterior hidden state at time t. The mean of the prior hidden states. This represents the mean of the observed hidden states.

[0047] Finally, the posterior hidden state at the current time is obtained by sampling from the distribution of the fused posterior hidden states. .

[0048] The above steps yield the posterior hidden state at the current moment. Based on this posterior hidden state, multiple rounds of autoregressive prediction using a causal dynamic world model are performed to obtain a sequence of hidden states for the next multiple steps. Each round of prediction by the causal dynamic world model outputs a hidden state. The specific process is as follows: Using the posterior hidden state at the current moment As the initial input to the causal dynamic world model, for the prediction step size arrive (for example (corresponding to the next 60 seconds), input the intervention action corresponding to the step size k. and double-constrained boundary variables In this process, the intervention action is predicted by the work order status in the warehouse management system during normal operation, based on the work order type and preset work order process. When there is no work order, the system remains static with no action. The double-constraint boundary variables are extrapolated from the current constraint variables, and then the hidden state with step size k is predicted through the state transition function. The hidden state obtained through K rounds of prediction This forms a sequence of hidden states for future multiple steps. In this embodiment, the final output sequence of hidden states for future multiple steps from the causal dynamic world model includes the hidden states for the next K steps. In addition, it also includes the hidden state at the current time t. That is, the complete sequence is There are a total of K+1 time steps. This is crucial for the subsequent decoding head to output the result at the current moment.

[0049] Among them, the double-constraint boundary corresponding to the step size k (k=1,2,...,K) in the autoregressive prediction process. The following extrapolation method was used to obtain: 1. Steps for extrapolating physical constraint boundaries: Step 1: Obtain the physical constraint matrix at the current time t ; Step 2: Based on the current location of the supplies and movement speed Predict the approximate location range of materials when the step size is k. The moving speed is obtained by dividing the coordinate distance of the current position of the material obtained from sampling at adjacent time points by the time interval; Step 3: Within the location range Internally, based on the assumption of spatial continuity, physical accessibility does not undergo abrupt changes, therefore exist Internal retention and same; Step 4: The global physical constraints of the warehouse (fixed obstacles such as shelves and walls) remain unchanged in the prediction time domain and can be directly reused; Step 5: Flattening yields the physical constraint vector .

[0050] 2. Steps for extrapolating business constraint boundaries: Step 1: Obtain the business similarity vector at the current time t: Step 2: Perform time-series smoothing extrapolation using exponential moving average (EMA) to obtain the business similarity vector with step size k: in, =0.9 is the smoothing coefficient; Step 3: Ensure that the extrapolated similarity value remains within the range of [0,1]; Finally, the physical constraint vector and the business constraint vector are concatenated to obtain the double-constraint boundary for step size k: The future multi-step hidden state sequence (including the current moment) output by the causal dynamic world model is input into the dual-task parallel decoding module to achieve joint prediction of location and behavior. In this embodiment, the dual-task parallel decoding module consists of two parallel decoding heads: a location prediction decoding head and a behavior prediction decoding head. The input to both decoding heads is the future multi-step hidden state sequence. The location prediction decoding head uses a 3-layer fully connected network. Its output includes the current material location, a predicted material location sequence for future moments, and a location prediction confidence sequence. Specifically, when RFID signal loss or obstruction prevents direct acquisition of effective location information at the current moment, the model uses historical trajectory information, the current intervention action, and dual constraint variables to complete the precise location of the material at the current moment through encoding and decoding processes. This completion process only uses current and historical information and does not involve future information, conforming to causal logic. The specific implementation mechanism of the current location completion is as follows: the input sequence received by the decoding head... The first element is the hidden state at the current time t. ; The historical trajectory information has been fully encoded (all historical radio frequency features from t-L+1 to t are aggregated through the encoder's causal attention mechanism), and the current intervention action has been included. (Through cross-attention integration) and double-constrained variables (Integrated through gating mechanism); Decoding head extracted via index. The current location coordinates of the materials are obtained directly through a 3-layer fully connected network.

[0051] The future material location prediction sequence is the three-dimensional coordinates of the material at each prediction step time, and the location prediction confidence sequence is the prediction confidence of the three-dimensional coordinates of the material at each prediction step time. The confidence can be calculated based on the latent state variance or by using other conventional confidence calculation methods, which will not be elaborated here.

[0052] The behavior prediction decoder employs a cascaded architecture of a two-layer fully connected network and a Softmax activation function. Its input is a sequence of future multi-step hidden states, and its output includes the current behavior classification result, the predicted future behavior sequence, and the risk level. The steps for obtaining the current behavior classification result are as follows: the decoder extracts the first element of the input sequence using an index. (Current hidden state); The system has already encoded the complete state information for the current moment, including location, movement trajectory, and business status. After passing through a two-layer fully connected network and Softmax activation, it directly outputs the probability distribution of each behavior type at the current moment, i.e., the current behavior classification result. The behavior classification result corresponds to a pre-defined behavior type, including normal behavior types and abnormal behavior types. Normal behavior types include inbound operations, outbound operations, transfer operations, inventory counts, requisition and return, inactivity, and abnormal movement. Abnormal behavior types include outbound movement, movement outside of operating hours, unauthorized operation, misplacement, and unauthorized outbound movement. It should be noted that the classification of normal and abnormal behavior in this embodiment is not only based on the presence or absence of a work order, but also on whether the movement trajectory conforms to the physical constraints of the warehouse and the business compliance prototype. Normal behavior refers to the material movement path, activity area, and movement duration completely matching the warehouse access rules and historical compliant trajectories. The movement trajectory includes not only compliant movement with a work order, but also compliant movement without a work order, i.e., abnormal movement. Abnormal movement can be understood as only lacking formal work order documents, but the entire movement conforms to physical space constraints and matches the historical fine-tuning compliant trajectory within the warehouse. There are no safety, inventory, or process risks, so it is classified into the normal behavior category, and is only used as a sub-category for differentiated risk rating.

[0053] The current behavior classification result is the probability vector of each behavior type at the current moment. The future behavior prediction sequence refers to the probability vector of the behavior type corresponding to the prediction step size. The risk level is determined by the business compliance score and the current behavior type. The business compliance score does not require additional input because... The trajectory feature information has already been encoded in the data, therefore from... The trajectory feature vector is decoded, and its maximum similarity with each prototype in the business compliance feature library is calculated: The score This is the business compliance score. The risk level is determined by both the compliance score and the behavior type. For example, low risk is a compliance score ≥ 0.7 and normal behavior; medium risk is 0.4 ≤ compliance score < 0.7 or abnormal movement; high risk is a compliance score < 0.4 or abnormal behavior. The risk level judgment logic is encoded in the last layer of the fully connected network of the decoder and is learned through end-to-end training.

[0054] In a preferred embodiment, a multi-objective joint loss function is designed to complete the phased training and closed-loop optimization of the model. Specifically, the total loss function of the causal dynamic world model consists of multi-step dynamic prediction loss, causal consistency loss, constraint compliance loss, and KL divergence constraint loss, and its expression is as follows: In the formula, Let be the total loss function of the causal dynamic world model. To predict losses dynamically in multiple steps, For loss of causal consistency, To constrain compliance losses, The loss is the KL divergence constraint loss. , , and All are loss weights.

[0055] The multi-step dynamic prediction loss is the MSE loss that simultaneously optimizes the prediction of hidden states in multiple future steps. Its weight is preferably set to 0.45, and the longer the prediction step size K, the higher the weight. This inherently solves the problem of long-term error accumulation. Its expression is: In the formula, K represents the maximum prediction step size. Indicates the first The weight of each step, where N represents the batch size. Indicates the first One sample in The mean of the predicted hidden states of the step. Indicates the first One sample in The mean of the true hidden states of each step.

[0056] The causal consistency loss employs contrastive learning loss to ensure that, under the same initial state, the same intervention action corresponds to similar state transition outcomes, while different intervention actions correspond to significantly different transition outcomes. Its expression is: In the formula, N represents the batch size. This indicates an indicator function, which is 1 if the condition is true and 0 otherwise. These are the intervention action labels for the i-th and j-th samples, respectively; These are the predicted hidden states of the i-th and j-th samples, respectively; This represents the cosine similarity function.

[0057] The loss due to compliance constraints is divided into two parts: loss due to physical constraints and loss due to business compliance. Among them, physical constraint loss The expression used to penalize out-of-bounds predictions is: In the formula, For the first The average value of the physical constraint gating weight vector of each sample.

[0058] Business compliance losses The expression used to penalize predictions that do not conform to compliance patterns is: In the formula, For the first The average value of the business constraint gating weight vector for each sample.

[0059] KL divergence constraint loss To constrain the difference between the Gaussian distribution and the standard normal distribution of the latent states and prevent the latent space from collapsing, the expression is: In the formula, D represents the total dimension of the hidden states. For the first The hidden state of the nth sample The mean of the dimension, For the first The hidden state of the nth sample The variance of the dimension.

[0060] In addition to the total loss of the causal dynamic world model mentioned above, the loss of the dual-task parallel decoding module also needs to be calculated. The total loss of the dual-task parallel decoding module consists of location prediction loss and behavior prediction loss. The location prediction loss... The mean squared error loss is expressed as follows: In the formula, For the first One sample in The actual position coordinates of the step, For the first One sample in The predicted position coordinates of the step. The boundary constraint loss is used to penalize predictions that exceed the warehouse boundary. The boundary loss weights are used.

[0061] The specific formula for calculating the boundary constraint loss is as follows: In the formula, For the first One sample in The normalized score of the warehouse grid corresponding to the predicted coordinates of the step is determined. When the predicted location falls within the legally accessible area of ​​the warehouse, the score is ≤1 and the loss is 0. When the predicted location goes out of bounds or falls into illegal areas such as shelves / walls, the score is >1, resulting in a positive penalty loss and forcing the model to converge to the legal spatial range.

[0062] Behavioral prediction loss The expression is: In the formula, For cross-entropy loss, For compliance reasons, the loss weights are compared. Compare the business compliance losses with those of the causal dynamic world model. Unlike other methods, this loss directly affects the behavior classification result, expressed as: In the formula, The behavior of the i-th sample is classified, with 0 representing normal behavior and 1 representing abnormal behavior. Indicates the first The maximum similarity between the trajectory features of each sample and the compliant prototype.

[0063] Based on the aforementioned loss function, this embodiment employs a four-stage training strategy. The first stage involves pre-training the temporal hidden state encoding module, using unlabeled data to learn temporal feature representations. The second stage involves pre-training the causal dynamic world model kernel, using simulation datasets and historical compliant business datasets to learn causal state transition patterns and freeze encoder weights. The third stage involves end-to-end joint fine-tuning, using real-world scenario datasets to unfreeze the top-level encoder and world model weights, jointly optimizing the decoder head's position and behavioral tasks. The fourth stage involves incremental closed-loop self-learning, automatically identifying difficult examples with prediction errors exceeding thresholds and newly added compliant business samples during model inference, performing incremental fine-tuning weekly to continuously optimize model performance and automatically adapt to adjustments in business processes. It should be noted that the loss function in the above embodiment is a preferred setting; other loss functions can also be used for model training, and this is not a limitation.

[0064] The trained model is deployed using a lightweight approach. By reducing the model size and inference latency, a three-tiered distributed architecture of "cloud, edge, and device" is adopted, which is the standard for the power industry. The device is responsible for raw data collection, while the edge deploys a lightweight inference engine to achieve real-time prediction and early warning. The cloud is responsible for centralized model training, version management, and centralized monitoring of multiple warehouses. At the same time, it is deeply integrated with the warehouse management system, video surveillance system, access control and security system, and ERP system to achieve closed-loop management of the entire life cycle of materials.

[0065] The identification process of this embodiment will be described in detail below using an indoor power supply warehouse as an example. Assume that the warehouse area is 2000㎡, divided into an inbound area, an acceptance area, a shelf area, a sorting area, an outbound area, and a hazardous chemical isolation area. 16 fixed UHF RFID readers conforming to the GB / T 29768-2013 standard are deployed, along with 8 UWB positioning anchors (positioning accuracy ≤10cm). All devices are synchronized through an NTP network time server, with a synchronization error ≤1ms, and the data sampling frequency is fixed at 10Hz.

[0066] Based on the above-mentioned material inventory leveling, the specific implementation steps of this embodiment are as follows: First, multi-source data collection and standardized preprocessing are performed. The data collection content includes: a) RFID core time-series data: collect the unique code of the anti-metal tag of the material, read the reader ID, RSSI signal strength, phase value, and millisecond-level timestamp, with a sampling frequency of 10Hz; b) UWB ground truth data: Collect the real-time coordinates (x, y, z) of the UWB tag that is attached to the RFID tag, with a sampling frequency of 50Hz, as the ground truth position for model training; c) Business data: Connect to the warehouse management system to collect historical compliant work order data, real-time work order status, work process, material type, operator, and work period data for the past two years, extract time-series trajectory features of different business types, and build a compliant business time-series feature library; d) Prior data: Complete a 1:1 on-site survey of the warehouse, draw a rasterized digital map with a resolution of 0.2m×0.2m, and mark the accessible / inaccessible areas, shelf locations, area boundaries, and the matching relationship between material types and corresponding storage areas.

[0067] The standardized preprocessing workflow includes: a) Time synchronization and resampling: Unify the data from RFID, UWB and warehouse management system to the same time axis, and use linear interpolation to resample all data to a fixed frequency of 10Hz to ensure that the features and true values ​​of each time step are strictly aligned. b) Noise Reduction and Anomaly Removal: Kalman filtering is used to smooth the RSSI signal with a processing window size of 5; the 3σ criterion is used to remove outliers of RSSI exceeding [-90dBm, -30dBm], UWB coordinate outliers exceeding the warehouse boundary, and filter EPC misalignment data caused by tag collisions. c) Trajectory association and segmentation: The time-series data of the same material are associated as a continuous trajectory according to the unique code of the anti-metal tag of the material. The trajectory is segmented according to the start and end nodes of the work order in the warehouse management system. Each trajectory segment is labeled with the UWB location truth value, business action label and behavior type label. d) Feature normalization and encoding: Min-max normalization is performed on continuous features such as RSSI, phase, and coordinates, scaling them to the [0,1] interval; 8-dimensional embedding encoding is performed on discrete features such as reader ID, region code, and material type, and finally a 32-dimensional fixed feature vector is output at a single time step. e) Data augmentation: Perform random packet loss simulation (up to 50% packet loss rate), add Gaussian noise, scale the time axis, and slightly shift the path to improve the model's anti-interference ability.

[0068] Then, the dataset is constructed, including: a) Real-world scenario dataset: 12,000 complete business trajectories were collected, with each trajectory having a length of ≥100 time steps, covering 52 types of power materials, 7 types of normal business scenarios, and 5 types of abnormal scenarios, with a total time series data volume of 12 million frames. b) Simulation dataset: A 1:1 digital twin model of the warehouse was built based on Unity to simulate RFID signal propagation, multipath effect, occlusion and packet loss, generating 1.2 million complete trajectories, with a single trajectory length of ≥200 time steps, and a total time series data volume of 240 million frames. c) Dataset splitting: Split according to business scenarios, with 70% for training set, 15% for validation set, and 15% for test set. The test set includes material types and operation scenarios not present in the training set to verify the model's generalization ability.

[0069] Then, the temporal hidden state encoding module is constructed. The encoder architecture adopts a 3-layer causal Transformer encoder, with each layer configured with 8-head causal self-attention and a strict lower triangular causal mask to prevent future information leakage. Each layer is configured with a 2-layer feedforward neural network with a hidden layer dimension of 512, GELU activation function, dropout rate of 0.1, and layer normalization and residual connections are added. The position encoding adopts temporal relative position encoding + reader spatial coordinate embedding, directly integrating the reader's normalized physical coordinates (x,y) into the position encoding, replacing the spatial feature extraction network.

[0070] Then, a causal dynamic world model is constructed, where the hidden state dimension is fixed at 256, intervention actions are divided into 7 categories, and dual-constraint variables are synchronized in real time with the warehouse grid map and the compliance business time-series feature library. The constrained causal Transformer block adopts a 4-layer structure, with each layer consisting of a cross-attention unit, a dual-constraint gating unit, a feedforward neural network (hidden layer dimension 512), layer normalization, and residual connections. In the dual-constraint gating unit, the physical constraint gating calculates the accessibility of the location grid corresponding to the predicted state in real time, and the business constraint gating calculates the similarity between the time-series features of the predicted behavior and the compliance business feature library to achieve dual-constraint filtering. The location prediction decoding head of the dual-task parallel decoding module adopts a 3-layer fully connected network, and the behavior prediction decoding head adopts a 2-layer fully connected network with softmax activation at the end.

[0071] Finally, model training is performed. In the first stage of encoder pre-training, the encoding module uses temporal reconstruction self-supervised pre-training, with a batch size of 64, a learning rate of 1e-4, an optimizer of AdamW (weight decay of 1e-4), and 50 epochs, with the learning rate decaying by 0.5 every 10 epochs. After pre-training, the weights of the bottom two encoder layers are frozen. In the second stage of causal dynamic world model pre-training, the encoder weights are frozen, with a batch size of 32, a learning rate of 5e-5, an optimizer of AdamW, and 100 epochs. Course learning is used, first training 5 steps of short temporal prediction, gradually increasing to the maximum prediction step size, with the learning rate decaying by 0.5 every 20 epochs. In the third stage of end-to-end joint fine-tuning, the weights of the top one layer of the encoder are unfrozen, and the batch size is set to 64, the learning rate to 1e-4, the optimizer to AdamW (weight decay of 1e-4), and the learning rate to 100. With size=16, learning rate=1e-5, epoch=50, and total loss consisting of the total loss of the causal dynamic world model, location prediction loss, and behavior prediction loss, the learning rate decays by 0.5 every 10 epochs. In the fourth stage of incremental closed-loop self-learning, a prediction error threshold of 1m is set, and difficult examples with errors exceeding the threshold and newly added compliant business samples are automatically collected. Incremental fine-tuning is performed once a week with a learning rate of 1e-6 and epoch=10, and only the weights of the dock and the top two layers of the world model are fine-tuned.

[0072] Finally, deployment and integration were implemented. The trained model underwent INT8 quantization, and TensorRT was used for operator fusion and inference optimization. Knowledge distillation was used to compress the model size, resulting in a final model memory footprint of ≤6GB and an end-to-end inference latency of ≤80ms per trajectory. The deployment architecture adopted a three-tiered "cloud-edge-device" architecture. At the edge, an NVIDIA Jetson Orin 32GB edge server was used to deploy the inference engine, enabling local real-time prediction and anomaly warning. The cloud deployment was on a private power cloud, responsible for model training, version management, and multi-warehouse monitoring. Business integration included interface with the warehouse WMS system to achieve dynamic inventory and automatic verification of incoming and outgoing materials; interface with access control and video surveillance systems to automatically trigger light and sound alarms, access control locking, and video verification for abnormal behavior; and interface with the ERP system to achieve full lifecycle traceability management of materials.

[0073] This embodiment provides a method for identifying the behavior of power material storage. By intrinsically embedding physical space constraints and business compliance features into the world model kernel of a causal temporal architecture, the method improves the adaptability of the model when warehouse layout is adjusted, equipment failures occur, and business processes change. Through the intervention state decoupling causal modeling mechanism and multi-step joint loss training, it effectively removes false associations caused by RFID signal fluctuations and solves the problem of long-term prediction error accumulation. It can achieve high-precision material location prediction under extreme scenarios such as metal obstruction and high packet loss rate, and realize accurate behavior recognition and business compliance judgment. This embodiment can be quickly adapted to multiple types of power warehouses and can be replicated on a large scale, and can comprehensively realize the intelligent management and control level and safety production capability upgrade of power material storage.

[0074] Please see Figure 2 Based on the same inventive concept, the second embodiment of the present invention proposes a power material warehousing behavior identification system, comprising: The data processing module 10 is used to acquire radio frequency data and work order data of materials to be identified in the power material warehouse, and to preprocess them to obtain radio frequency characteristics and intervention actions. The data encoding module 20 is used to encode the radio frequency features using a time-series hidden state encoding module based on a causal time-series autoencoder to obtain the observation hidden state. The state prediction module 30 is used to input the observed hidden state, the intervention action and the preset double-constraint boundary into the preset causal dynamic world model, and perform multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on the constrained causal Transformer. The state transition function has embedded endogenous constraints, including physical space constraints and business behavior compliance constraints. The behavior recognition module 40 is used to input the future multi-step hidden state sequence into the dual-task parallel decoding module to obtain the position prediction result and the behavior prediction result. The dual-task parallel decoding module includes a position prediction decoding head and a behavior prediction decoding head.

[0075] The technical features and effects of the power material warehousing behavior identification system proposed in this embodiment of the invention are the same as those of the method proposed in this embodiment of the invention, and will not be repeated here. Each module in the above-mentioned power material warehousing behavior identification system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0076] Furthermore, embodiments of the present invention also propose a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0077] Please see Figure 3 The diagram illustrates the internal structure of a computer device in one embodiment. This computer device can specifically be a terminal or a server. The computer device includes a processor, memory, network interface, display, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for identifying the behavior of power material storage. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input devices of the computer device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse, etc.

[0078] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0079] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0080] In summary, the present invention proposes a method, system, device, and storage medium for identifying power material storage behavior. The method acquires radio frequency (RF) data and work order data of the materials to be identified in the power material warehouse, performs preprocessing to obtain RF features and intervention actions; employs a temporal hidden state encoding module based on a causal temporal autoencoder to perform hidden state encoding on the RF features, obtaining observed hidden states; inputs the observed hidden states, the intervention actions, and preset dual-constraint boundaries into a preset causal dynamic world model, and performs multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on a constrained causal Transformer, and the state transition function embeds endogenous constraints, including physical space constraints and business behavior compliance constraints; inputs the future multi-step hidden state sequence into a dual-task parallel decoding module to obtain location prediction results and behavior prediction results. The dual-task parallel decoding module includes a location prediction decoding head and a behavior prediction decoding head. This invention improves the adaptability of the model by intrinsically embedding physical space constraints and business compliance features into the world model kernel of the causal temporal architecture. Through the causal modeling mechanism of intervention state decoupling and multi-step joint loss training, it can achieve high-precision material location prediction in extreme scenarios, and realize accurate behavior recognition and business compliance judgment. This invention can be quickly adapted to multiple types of power warehouses and realized on a large scale, thereby improving the intelligent management and control level and safe production capability of power material storage.

[0081] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0082] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for identifying the storage behavior of electrical materials, characterized in that, include: The radio frequency data and work order data of the materials to be identified in the power materials warehouse are acquired and preprocessed to obtain radio frequency characteristics and intervention actions. The radio frequency features are encoded using a time-series hidden state coding module based on a causal time-series autoencoder to obtain the observed hidden state. The observed hidden state, the intervention action, and the preset dual-constraint boundary are input into a preset causal dynamic world model. Multiple rounds of hidden state prediction are performed through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on a constrained causal Transformer. The state transition function has embedded endogenous constraints, including physical space constraints and business behavior compliance constraints. The future multi-step hidden state sequence is input into a dual-task parallel decoding module to obtain position prediction results and behavior prediction results. The dual-task parallel decoding module includes a position prediction decoding head and a behavior prediction decoding head.

2. The method for identifying power material storage behavior according to claim 1, characterized in that, The steps of acquiring radio frequency data and work order data of materials to be identified in the power material warehouse, and preprocessing them to obtain radio frequency timing characteristics and intervention actions include: Acquire radio frequency data and work order data of materials to be identified in the power material warehouse, wherein the materials to be identified include materials associated with work orders and materials not associated with work orders; The radio frequency data is standardized and preprocessed to obtain radio frequency characteristics; An intervention action is generated based on the work order status of the work order data.

3. The method for identifying power material storage behavior according to claim 1, characterized in that, The temporal hidden state coding module consists of multiple layers of first causal Transformer encoders, and each layer of first causal Transformer encoder is equipped with a self-attention layer with a multi-head lower triangular causal mask. The position encoding of the temporal hidden state encoding module adopts a fusion position encoding of time relative position encoding and reader spatial coordinate embedding, wherein the reader spatial coordinate embedding is obtained based on the real physical position mapping of the reader.

4. The method for identifying power material storage behavior according to claim 1, characterized in that, The state transition function of the causal dynamic world model consists of a multi-layer second causal Transformer encoder. Each layer of the second causal Transformer encoder block consists of a cascaded cross-attention unit, a double-constraint gating unit, a feedforward neural network, and a layer normalization and residual connection unit. The cross-attention unit is used to calculate and output a context vector based on the hidden state and the intervention action through cross-attention. The dual-constraint gating unit includes physical constraint gating and business constraint gating. The physical constraint gating is used to filter illegal physical paths of the context vector according to the physical space constraints, and the business constraint gating is used to filter illegal business paths of the filtered context vector according to the business behavior compliance constraints. The physical space constraint is a constraint matrix based on the accessibility boundary within the warehouse, and the business behavior compliance constraint is the similarity between the time-series trajectory of the most recent time step and each compliance behavior prototype. The compliance behavior prototype is obtained by clustering analysis of historical compliance trajectory features.

5. The method for identifying power material storage behavior according to claim 1, characterized in that, The steps of inputting the observed hidden state, the intervention action, and the preset dual-constraint boundary into a preset causal dynamic world model, and performing multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence include: Using the posterior hidden state of the previous moment, the intervention action of the current moment, and the double-constraint boundary as a causal dynamic world model, the prior hidden state of the current moment is obtained. Determine whether the radio frequency data is valid data. If not, use the prior hidden state at the current moment as the subsequent hidden state at the current moment. If so, then calculate the Kalman gain based on the observed hidden state and the prior hidden state at the current time, and fuse the observed hidden state and the prior hidden state at the current time based on the Kalman gain to obtain the posterior hidden state at the current time. The posterior hidden state at the current moment, the intervention action at the next moment, and the double-constraint boundary are input into the causal dynamic world model. Autoregressive multi-round prediction is performed according to the preset step size to obtain the future multi-step hidden state sequence.

6. The method for identifying power material storage behavior according to claim 1, characterized in that, The location prediction decoding head adopts a 3-layer fully connected network, and the output location prediction results include the current material location, the material location prediction sequence for future times, and the location prediction confidence sequence. The behavior prediction decoding head consists of a cascaded two-layer fully connected network and a Softmax activation function. The output behavior prediction results include the current behavior classification result, the future behavior prediction sequence, and the risk level.

7. The method for identifying power material storage behavior according to claim 1, characterized in that, The loss function of the causal dynamic world model includes multi-step dynamic prediction loss, causal consistency loss, constraint compliance loss, and KL divergence constraint loss. The multi-step dynamic prediction loss is constructed based on the mean squared error loss of future multiple steps, the causal consistency loss is constructed based on the contrastive learning loss, and the constraint compliance loss includes physical constraint loss and business compliance loss.

8. A power material warehousing behavior identification system, characterized in that, include The data processing module is used to acquire radio frequency data and work order data of materials to be identified in the power material warehouse, and to preprocess them to obtain radio frequency characteristics and intervention actions. The data encoding module is used to encode the radio frequency features using a time-series hidden state encoding module based on a causal time-series autoencoder to obtain the observed hidden state. The state prediction module is used to input the observed hidden state, the intervention action, and the preset dual-constraint boundary into a preset causal dynamic world model, and perform multi-round hidden state prediction through autoregression to obtain a future multi-step hidden state sequence. The state transition function of the causal dynamic world model is constructed based on the constrained causal Transformer. The state transition function has embedded endogenous constraints, including physical space constraints and business behavior compliance constraints. The behavior recognition module is used to input the future multi-step hidden state sequence into the dual-task parallel decoding module to obtain the position prediction result and the behavior prediction result. The dual-task parallel decoding module includes a position prediction decoding head and a behavior prediction decoding head.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images