In recent years, rapid developments in technology have improved the collection of massive data from different industrial processes. These new developments have also brought benefits for many machines such as ball mills. It is now possible to develop soft sensors for the measurement of load parameters. Vibration and acoustical signals produced by ball mill shells, during the grinding process, contain useful information which is used to determine load parameters inside the equipment. The shell vibration signals and acoustical signals are called secondary variables. They are easy to acquire. The desired variables, the load parameters, are called the primary variables. For the aforementioned, in the present work, vibration and acoustical signals from the ball mill shells are used for modeling a soft sensor. The literature mentions three parameters for the load in ball mills:Material Ball Volume Ratio (MBVR), Pulp Density (PD) and Charge Volume Ratio (CVR). The soft sensor should be able to measure these three variables. Most of the approaches use three models of soft sensors, thus one for each variable (MBVR, PD and CVR). Nonetheless, this thesis proposes a multi-task model capable of calculating the three different variables at the same time as well as making the approach efficient and simple to model. When modeling of a soft sensor, first, it is necessary to sample the signals. Secondly, the data must be pre-processed, and then performs the extraction of characteristics by a neural network (NN). Finally, to find the intrinsic relationship between the secondary variables and the primary variable, a multi-task Recurrent Neural Network (RNN) based on Long Short Term Memory (LSTM) is modeled.　　This thesis focuses on the implementation of a multi-task soft sensor to measure the load of ball mills, highlighting the multi-task feature, calculating all load parameters (MBVR, PD and CVR) in a single iteration, solving the tedious problem of modeling three systems, making a simple model, and maintaining the precision. The procedure is conceivable since the parameters share a similar statistical distribution. In other words, the parameters share a common structure because all the load parameters belong to the same parameters in a universe, which is the Mill Load (ML) inside the ball mill. The model is based on a RNN-LSTM. Due to the ability of RNN-LSTM to work with temporal sequence, like in the load parameters for ball mills, this kind of network is chosen for modeling a soft sensor.　　RNN-LSTM work are dependent on time, making recursive the input "n"times, thereby learning intrinsic structures within the given time-interval. Their connections between units form a cycle, creating an internal state in the network, which allows it to represent temporal dynamic behavior. As a result of this behavior, the RNN-LSTM take into account the time variable, which makes them more efficient in data prediction, as in this case, the load in ball mills. The RNN-LSTM also possess great flexibility and modularity. As most of the NNs, they are able to have one or more signals to the input and one or more outputs,hence it is a multi-task approach.　　In order to validate the accuracy of the multi-task soft sensor model, based on RNN-LSTM, experiments on a lab-scale ball mill machine are carried out. The experiments simulate industrial conditions, recording the sample data, labeling each load parameter for each vibration and acoustic signal sample.　　The main research content in the present work is:　　1. Analyzing the audio and vibration signals in the time-frequency domain.　　2. Finding the temporal relationship that exists between samples with RNN-LSTM.　　3. Demonstrating that a multi-task neuronal network can substitute individual models yet achieve the same accuracy.　　The experimental results show that multi-task soft sensor model based on an RNN-LSTM is accurate and stable. Comparisons are also made with single-task models to verify the results of the RNN-LSTM.